CoreWeave owns about 16,525 Nvidia Blackwell GPUs per $1B of market cap, 77× Microsoft’s 214. Most of that gap is debt: CoreWeave’s enterprise value is 2.4× its market cap, and per $1B of enterprise value IREN leads at 11,275. Amazon has the largest disclosed claim on what comes next, at least 3M Nvidia GPUs through 2028; per dollar, the neoclouds’ unbuilt contracted power is bigger still, if they can finance it.

• 16,525 — Blackwell GPUs per $1B of market cap, CoreWeave. Microsoft 214, Alphabet 86. Nvidia GPUs only, estimates
• 11,275 — Per $1B of enterprise value, IREN. CoreWeave falls to 6,882: its enterprise value is 2.4× its market cap
• 846K — Largest Blackwell owner, Microsoft. B200 and B300 shipped through Q1 2026, per Epoch AI. Meta 519K, Amazon 401K
• 3M+ — Largest disclosed next allocation, AWS. Nvidia GPUs: more than 1M from 2026, then 2M more through 2027–28

Blackwell GPUs per $1B of market cap

Why the neoclouds top the list

It’s a leverage screen. GPUs per dollar of market cap rise as the equity shrinks against the fleet. CoreWeave’s roughly 543K Blackwell GPUs sit under $78.9B of enterprise value but $32.9B of market cap; the rest is debt and leases, $31.4B of it recourse debt at June 30. IREN and Nebius carry less (1.16× and 1.05× their market caps), so per dollar of enterprise value IREN moves to first and CoreWeave’s figure falls by more than half, to 6,882.

And the market has marked the equity down. CoreWeave closed at $59.57 on October 27, against a 52-week high of $141.79; its credit default swaps peaked at 855bp in the run-up to its second-quarter results. IREN closed at $27.44, against $69.58 the night it announced its Nvidia partnership in May. Part of the per-dollar lead is the market discounting the debt behind the fleets.

Owning the GPU isn’t owning the upside. Most neocloud capacity is contracted out at fixed prices: Microsoft at IREN and Nebius, Meta at Nebius and at CoreWeave, where it signed $21B through 2032 on top of $14B. Shareholders earn what’s left after interest and depreciation, and the customer keeps the margin on what runs on the chips. CoreWeave’s adjusted operating margin was 5% in the second quarter.

Blackwell GPUs owned

Blackwell shipped by quarter

Who owns the most

Microsoft, by Epoch AI’s estimate: about 846K B200 and B300 through the first quarter, ahead of Meta (519K), Amazon (401K) and Alphabet (332K). Epoch’s estimate for Microsoft’s Fairwater Atlanta site alone is about 304K.

The biggest single site is xAI’s. Epoch puts Colossus 2 at 1.112 million H100-equivalents, about 440K Blackwell GPUs, owned by SpaceX since it bought xAI in February. Abilene, the site Oracle runs for OpenAI, is about 201K.

The counts lag the shipments. Epoch’s owner series ends in the first quarter and Nvidia has shipped two more quarters since, so every hyperscaler owns more than its base figure today. Two more quarters at the fourth quarter’s pace lift the hyperscalers’ ratios by about half, which doesn’t change the order: the lowest neocloud is still 12× the highest hyperscaler.

Who is lined up for the next allocation, per $1B of market cap

What each forward figure rests on

• Nebius NBIS — GPUs: 1.33M · Per $1B: 31,997 · Announced: Aug 2026 · Target of 5 GW contracted by year-end, less the 0.8–1 GW it expects connected by then. Low: the 3 GW it targeted in February, less 1 GW · In Nvidia’s first Rubin cloud cohort; deployments from late 2026 or early 2027
• CoreWeave CRWV — GPUs: 900K · Per $1B: 27,397 · Announced: Aug 2026 · 2.7 GW contracted but not yet active: 4.2 GW contracted as of August 11 against 1.5 GW active · First production Vera Rubin NVL72 customer, Cognition, on September 30
• IREN IREN — GPUs: 167K · Per $1B: 15,015 · Announced: May & Aug 2026 · The 500 MW of IT load it plans to add in 2027. Its Nvidia partnership covers up to 5 GW, a framework rather than orders · No Rubin orders disclosed
• Oracle ORCL — GPUs: ≥667K · Per $1B: ≥1,745 · Announced: Oct 27, 2026 · “Multi-gigawatt” Vera Rubin expansion on OCI, read at its floor of 2 GW · Vera Rubin NVL72 generally available on OCI now
• Meta META — GPUs: ≥2M · Per $1B: ≥1,242 · Announced: Feb 18, 2026 · Multiyear Nvidia deal for “millions” of Blackwell and Rubin GPUs, read at its floor of 2 million. No unit count or value disclosed · Vera CPUs from 2027; part of the capacity comes through CoreWeave and Crusoe
• Amazon (AWS) AMZN — GPUs: ≥3M · Per $1B: ≥1,115 · Announced: Mar 16 & Aug 26, 2026 · More than 1 million Nvidia GPUs from 2026, then 2 million more through 2027–28, across Blackwell Ultra, Rubin and Rubin Ultra. The largest disclosed · In Nvidia’s first Rubin cloud cohort
• SpaceX (xAI) SPCX — GPUs: 530K · Per $1B: 265 · Announced: 2026 · Colossus build-out toward 1 million GPUs across its sites, less the roughly 470,000 Blackwell counted above · No Rubin volumes disclosed
• Microsoft MSFT — GPUs: ≥400K · Per $1B: ≥101 · Announced: Jun 2026 · Fairwater sites scaling to “hundreds of thousands” of Vera Rubin superchips, two GPUs each, read at the floor. Rubin it rents from neoclouds, such as 30,000 from Nscale, belongs to them and isn’t counted · In Nvidia’s first Rubin cloud cohort
• Alphabet GOOGL · No committed Nvidia volume disclosed. Google Cloud’s network design links up to 960,000 Rubin GPUs, a capability rather than an order · In Nvidia’s first Rubin cloud cohort

Who gets Rubin first

Nvidia named eight clouds first in line for Vera Rubin: AWS, Google Cloud, Microsoft Azure, Oracle Cloud, CoreWeave, Lambda, Nebius and Nscale. Production shipments began this fall. CoreWeave has had the first customer in production, Cognition, since September 30, and Oracle made Vera Rubin generally available on its cloud on October 27, with a multi-gigawatt expansion to come. Nvidia’s 2026 Rubin output is estimated at only 200K–300K GPUs, and it has historically sent 60%–70% of a new generation’s first year to the hyperscalers, so most of these allocations land in 2027 and 2028.

In GPUs, Amazon and Meta. AWS committed to more than 1 million Nvidia GPUs from 2026 and 2 million more through 2027–28, the largest disclosed. Meta’s February deal covers “millions” of Blackwell and Rubin GPUs, with no number given. Microsoft is building Fairwater for hundreds of thousands of Rubin superchips and also rents Rubin from neoclouds, 30,000 of them from Nscale in Norway.

Per dollar, the neoclouds, on paper. CoreWeave has 2.7 GW contracted that isn’t built, Nebius is aiming for 5 GW contracted by year-end, and IREN plans 500 MW more in 2027. That is about 31,997 next-generation GPUs per $1B of Nebius’s market cap and 27,397 per $1B of CoreWeave’s. But capex guided for this year is 1.13× CoreWeave’s market cap and 0.54× Nebius’s, and IREN’s for its fiscal 2027 is 2.48× its own. The next allocation will be paid for by lenders and by customers’ prepayments, not by today’s shareholders.

The ranking, with leverage

Sorted by GPUs per $1B of market cap. Enterprise value and guided capex for the four names where debt changes the picture.

• CoreWeave CRWV · Neocloud — Blackwell GPUs: 543K 397K–600K · As of: Jun 2026 · Market cap: $32.9B · Per $1B mkt cap: 16,525 · EV: $78.9B · Per $1B EV: 6,882 · EV ÷ mkt cap: 2.40× · Capex ÷ mkt cap: 1.13×
• IREN IREN · Neocloud — Blackwell GPUs: 145K 36K–150K · As of: Year-end 2026 (secured) · Market cap: $11.1B · Per $1B mkt cap: 13,063 · EV: $12.9B · Per $1B EV: 11,275 · EV ÷ mkt cap: 1.16× · Capex ÷ mkt cap: 2.48×
• Nebius NBIS · Neocloud — Blackwell GPUs: 161K 100K–200K · As of: Jun 2026 · Market cap: $41.7B · Per $1B mkt cap: 3,863 · EV: $43.7B · Per $1B EV: 3,684 · EV ÷ mkt cap: 1.05× · Capex ÷ mkt cap: 0.54×
• Oracle ORCL · Leveraged builder — Blackwell GPUs: 201K 100K–450K · As of: Oct 2026 · Market cap: $382B · Per $1B mkt cap: 527 · EV: $470B · Per $1B EV: 428 · EV ÷ mkt cap: 1.23× · Capex ÷ mkt cap: 0.24×
• Meta META · Hyperscaler — Blackwell GPUs: 519K 477K–818K · As of: Q1 2026 · Market cap: $1.61T · Per $1B mkt cap: 322
• SpaceX (xAI) SPCX · Leveraged builder — Blackwell GPUs: 470K 140K–585K · As of: Oct 2026 · Market cap: $2T · Per $1B mkt cap: 235
• Microsoft MSFT · Hyperscaler — Blackwell GPUs: 846K 780K–1.32M · As of: Q1 2026 · Market cap: $3.95T · Per $1B mkt cap: 214
• Amazon AMZN · Hyperscaler — Blackwell GPUs: 401K 373K–607K · As of: Q1 2026 · Market cap: $2.69T · Per $1B mkt cap: 149
• Alphabet GOOGL · Hyperscaler — Blackwell GPUs: 332K 308K–509K · As of: Q1 2026 · Market cap: $3.87T · Per $1B mkt cap: 86

What a Blackwell allocation is worth

Rubin is a big step: CoreWeave measured about 10× the tokens per megawatt of a GB200 NVL72 on DeepSeek R1 at matched interactivity. A Blackwell fleet contracted at today’s prices earns them until its contracts end. What it re-contracts for once Rubin is in volume is the main unknown in what it’s worth, and it matters most to owners who borrowed to buy it. The inference-margin pack has what a Blackwell GPU earns serving open models now, and the neocloud pack what the neoclouds still have to raise.

So owning the allocation through neocloud equity is a levered position on fleet value and refinancing more than on AI margins. Per dollar of enterprise value, with the contracts and their counterparties in view, is the fairer comparison.

How each count was built

IREN’s count is the company’s own target. The rest are Epoch AI’s model, Epoch’s site estimates or our conversion of disclosed power, so treat them as estimates.

• CoreWeave CRWV — Implied from power: 1.5 GW active on June 30, less the 360 MW its ~250,000 mostly-Hopper GPUs used at its IPO filing, at about 2.1 kW per Blackwell GPU. Range: 1.9–2.3 kW, and 20% of new power not Blackwell at the low end
• IREN IREN — Secured fleet: its 150,000-GPU target for year-end 2026 (over 50,000 B300 ordered in March plus Microsoft’s GB300), less a few thousand Hoppers. About a quarter is live: $1B of run-rate revenue against $4B contracted. The low end counts only that quarter
• Nebius NBIS — Implied from power, low confidence: 170 MW active at the end of 2025, scaled by run-rate revenue ($1.25B to $3B) to about 408 MW in June, less about 70 MW of Hopper, at about 2.1 kW per GPU
• Oracle ORCL — Abilene alone: Epoch AI’s 509,000 H100-equivalents for the site, about 201,000 Blackwell. Oracle says the Abilene supercluster has more than 100,000; Larry Ellison’s target for the site is more than 450,000. Other OCI regions aren’t counted
• Meta META — Epoch AI’s chip-owners model: B200 and B300 shipped through Q1 2026 (incomplete). Low: through Q4 2025. High: two more quarters at Q4 2025’s pace
• SpaceX (xAI) SPCX — Colossus 2: Epoch AI’s 1.112M H100-equivalents, about 440,000 Blackwell, plus 30,000 GB200 in Colossus 1. Low: Epoch’s owner data (140,000 added in Q3 2025). High: Colossus 2 at the reported 555,000
• Microsoft MSFT — Epoch AI’s chip-owners model, as for Meta. Its site estimate for Fairwater Atlanta alone is about 304,000
• Amazon AMZN — Epoch AI’s chip-owners model, as for Meta. Nvidia only: Trainium isn’t counted
• Alphabet GOOGL — Epoch AI’s chip-owners model, as for Meta. Nvidia only: Google’s TPUs, most of its compute, aren’t counted

What this is anchored on

Market caps at the October 26–27, 2026 closes.

• Hyperscaler Blackwell counts — B200 and B300 units shipped by owner and quarter, Q4 2024 to Q1 2026 (Q1 2026 incomplete) · Epoch AI, AI Chip Owners (updated May 12, 2026)
• Site estimates — Colossus 2 (SpaceX) 1.112M, Fairwater Atlanta (Microsoft) 769,000, Abilene (Oracle) 509,000 H100-equivalents · Epoch AI, AI data centers (updated Oct 26, 2026)
• CoreWeave — 1.5 GW active and 3.7 GW contracted on June 30, 4.2 GW by August 11; backlog $104B; Meta $21B through 2032 on top of $14B; recourse debt $31.4B; CDS peaked at 855bp · TechTimes, Aug 12, 2026
• CoreWeave — Adjusted operating margin 5% in Q2; FY2026 capex guide $35–39B · Futurum, Aug 2026
• CoreWeave baseline — About 250,000 GPUs, mostly Hopper, on 360 MW at its IPO filing · The Next Platform, Mar 5, 2025
• CoreWeave, Rubin — First production Vera Rubin NVL72 customer (Cognition); about 10× the tokens per megawatt of GB200 NVL72 on DeepSeek R1 · CoreWeave, Sep 30, 2026
• Nebius — About 170 MW active at the end of 2025; run-rate revenue $1.25B; Microsoft and Meta capacity delivered · Nebius Q4 2025 shareholder letter
• Nebius — Run-rate revenue $3B at June 30; 5 GW contracted-power target for year-end; FY2026 capex $20–25B; Rubin from late 2026 or early 2027 · Nebius Q2 2026 call
• IREN — 150,000-GPU target for year-end 2026; over 50,000 B300 ordered; Nvidia partnership up to 5 GW; $69.58 after hours on May 7 · Blockspace via Yahoo Finance, May 7, 2026
• IREN — $4B contracted run-rate revenue, $1B live after Microsoft accepted Horizon 1; 500 MW more IT load in 2027; FY2027 capex $25–30B · Yahoo Finance, Aug 27, 2026
• Oracle — Vera Rubin NVL72 generally available on OCI; multi-gigawatt Vera Rubin expansion planned; more than 100,000 Blackwell in the Abilene supercluster · Oracle, Oct 27, 2026
• Oracle — Abilene to house more than 450,000 GB200 eventually · DCD
• Oracle — Net debt $88.3B at August 31, 2026; FY2027 capex guide $90–95B · Cestrian Capital Research
• AWS — More than 1 million Nvidia GPUs over the next 12 months · DCD, Mar 16, 2026
• AWS — 2 million additional Nvidia GPUs through 2027–28: Blackwell Ultra, Rubin, Rubin Ultra · The Energy Mag, Aug 26, 2026
• Meta — Multiyear deal for millions of Blackwell and Rubin GPUs · PureAI, Feb 18, 2026
• Rubin — First cohort: AWS, Google Cloud, Azure, OCI, CoreWeave, Lambda, Nebius, Nscale; 2026 output estimated at 200,000–300,000, 60–70% to hyperscalers; Fairwater to scale to hundreds of thousands of Vera Rubin superchips; Google network up to 960,000 Rubin GPUs · TechTimes, Jun 27, 2026
• Microsoft — Rents 30,000 Vera Rubin from Nscale in Norway · Let’s Data Science, Apr 14, 2026
• SpaceX — Listed June 12, 2026; owns xAI since February; Colossus across sites toward 1 million GPUs · Wikipedia
• SpaceX — Colossus 2 reported at 555,000 GPUs · Introl, Jan 2026
• Market caps — Closes on October 26: Microsoft, Alphabet, Amazon, Meta, Oracle, SpaceX · StockTitan
• CoreWeave market — $59.57 close Oct 27 (52-week range $51.24–$141.79); EV · Motley Fool; StockAnalysis
• CoreWeave market — Market cap and enterprise value, Oct 23 (re-marked to Oct 27’s close) · StockAnalysis
• Nebius market — $41.66B market cap at the Oct 26 close; debt $10.06B, cash $8.04B · CompaniesMarketCap (Nasdaq); TipRanks
• IREN market — $27.44 close Oct 27, $11.10B market cap; debt $7.84B, cash $6.08B · CompaniesMarketCap; TipRanks

Method

Blackwell GPUs are counted one per B200, B300, GB200 or GB300 GPU package, not per die. For Microsoft, Meta, Amazon and Alphabet they are Epoch AI’s estimates of B200 and B300 shipped through Q1 2026, which Epoch flags as incomplete. Site estimates in H100-equivalents are converted at Epoch’s own ratio, 2.527 per Blackwell GPU.

Power to GPUs. CoreWeave and Nebius don’t publish GPU counts, so we convert active power at about 2.1 kW of IT load per Blackwell GPU (1.9–2.3): a GB200 or GB300 NVL72 rack draws about 120–140 kW for 72 GPUs, plus networking and storage. CoreWeave’s base removes the 360 MW its mostly-Hopper fleet used at its IPO filing. Nebius’s removes about 70 MW of Hopper from the roughly 408 MW we infer for June from its run-rate revenue. Forward power converts at 3 kW per Rubin-generation GPU (2.6–3.4).

Market cap and enterprise value use closing prices on October 26 (October 27 for CoreWeave and IREN) and all share classes. Enterprise value adds net debt from the latest balance sheet: CoreWeave’s from StockAnalysis, re-marked to the October 27 price; Oracle’s $88.3B at August 31; Nebius’s and IREN’s from TipRanks. Capex is the latest full-year guide at its midpoint (fiscal 2027 for Oracle and IREN).

Not counted: GPUs a company rents rather than owns (they count for the owner, so Microsoft’s capacity at Nebius, IREN and Nscale counts for them), Hopper and older GPUs, and other chips: TPUs, Trainium and AMD. Private owners, among them OpenAI, Anthropic, Lambda, Nscale and Crusoe, have no market cap to divide by.

Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at 06.ai/trackers.

Not investment advice. Estimates are labeled as estimates.

Serving DeepSeek’s V4.1-Flash on a Blackwell rack earns about $39 an hour per GPU at DeepSeek’s off-peak prices, so a host renting the rack covers its rent at 37% utilization. On an 8-GPU B200 box running NVIDIA’s standard recipe it takes 60%, and an H200 can’t cover its rent at those prices at all. The margin sits with rack-scale serving, frontier-size models and owned GPUs; on the most popular model, 33 hosts compete on price.

• $39 — Revenue per GPU-hour, Blackwell rack. V4.1-Flash at DeepSeek’s off-peak prices on a GB300 NVL72, per vLLM. The rack rents for $14.59 a GPU-hour
• 37% — Utilization a renter needs to break even. Same rack. 60% on an 8-GPU B200 box with NVIDIA’s recipe; an H200 would need 120%, so never
• 33 — Hosts selling one model for V4.1-Flash on OpenRouter. 15 list exactly DeepSeek’s peak rate; 8 discount by 30–84%
• 24.6% — Zhipu’s gross margin on API sales. H1 2026, from −0.4% a year earlier, after raising API prices about 101% this year

What a GPU earns, and the utilization a renter needs

An owner needs far less. At SemiAnalysis’s owner costs ($1.95 an hour for a B200, $2.65 for a GPU in a GB300 rack) the B200 box covers its costs at 15% utilization and the GB300 rack at 7%. Kimi K3 on H200s doesn’t cover even the cost of owning them.

How the money comes in

Routing is price-first. OpenRouter’s default spreads requests across hosts that have been stable, weighted by the inverse square of price, so a host at half the price draws four times the traffic. It adds no markup to host prices and earns a 5.5% fee when buyers load credits, so a host keeps the price it posts.

The traffic is agents, and mostly input. In the week to October 14, V4.1-Flash processed 36.6 trillion prompt tokens and 807 billion completion tokens, about 45 to 1, and its biggest senders are coding and agent harnesses. In the week to October 26 it carried 52.8T tokens, 28% of everything in OpenRouter’s daily top 50. Chinese models carry 60% of OpenRouter’s tokens over the last four weeks but 26% of the dollars (tracker).

So it’s a cache business priced per input token. On NVIDIA’s trace, 70% of a V4.1-Flash request’s revenue comes from the 10% of prompt tokens that miss the cache; output is 17%. At OpenRouter’s 45-to-1 mix, uncached input is still about 48%. The cache-read price, which hosts set separately, decides what an agent pays. The DeepInfer pack shows the same split from DeepSeek’s side.

Posted isn’t paid. 15 of the 33 hosts list exactly DeepSeek’s peak rate, $0.30 per million input tokens and $1.20 per million output; 8 run discounts of 30% to 84%. OpenRouter now shows next to the posted prices what buyers actually pay on average, which caching and discounts pull well below the list.

What one agent request costs at each host

Headline prices mislead. Decart has the 2nd-lowest input-plus-output price but ranks 11th on an agent’s bill, because its cache reads cost $0.018 per million, three times DeepSeek’s peak rate. Alibaba Cloud drops from 10th to 19th the same way. Modal and Google Vertex charge $0.03 for cache reads and cost 3.0× DeepSeek’s off-peak rate per request.

Which models pay best

Frontier-size open models, on the newest racks. Kimi K3 has 2.8 trillion parameters, 104 billion of them active, and nine of the ten platforms that served it in its first weeks charged Moonshot’s $3 input, $15 output. That price holds because the economics only work at rack scale: on NVIDIA’s recipes a GB300 GPU earns $34.91 an hour on K3, a GB200 $12.58 and an H200 $1.35.

New models, for a few days. StepFun’s Step 5 Preview reached second place on OpenRouter in its first week, and K3 was live on nearly every major platform within 48 hours of its weights. The window before thirty hosts pile in is short.

The Flash tier is where the volume is. V4.1-Flash, GLM-5.3-Flash and MiniMax’s M3.1-Flash carry the most tokens, and their prices follow cost down: over about 90 days the cheapest input price fell 77% for V4.1-Flash and 90% for GLM-5.3-Flash. GLM-5.3-Flash at Z.ai’s list earns $15.05 an hour on a B200, slightly more than V4.1-Flash at DeepSeek’s off-peak rate. Margins in this tier go to whoever runs the rack-scale stack.

Hopper loses on every model here. An H200 can’t cover its rent on V4.1-Flash at DeepSeek’s off-peak prices, and NVIDIA doesn’t publish an H100 recipe for the model at all. Free endpoints, MiniMax’s M3.1-Flash and NVIDIA’s Nemotron 3 Ultra among them, carried 12.6T and 5.5T tokens in the week to October 26, free to users, so someone else pays.

Who keeps the margin

Reported figures. The measures differ, so compare rows with care.

• DeepSeek — Layer: Lab serving its own open weights · Measure: Theoretical cost-profit ratio, V3 and R1 on H800s priced at $2 an hour · Feb 2025 · Value: 545% · At R1 list prices; DeepSeek said actual revenue was far lower
• Nebius NBIS — Layer: GPU cloud with a token API · Measure: Adjusted EBITDA margin, AI business · Q2 2026 · Value: 50% · Mostly GPU rental; Token Factory is its fastest-growing line
• Fireworks — Layer: Inference platform, private · Measure: Gross margin · 2026 · Value: ~50% · Targeting 60%
• Cerebras CBRS — Layer: Inference cloud on its own chips · Measure: Core gross margin · Q2 2026 · Value: 41% · Cloud services 53% (core) in Q1
• Zhipu 2513.HK — Layer: Lab selling its open weights by API · Measure: Gross margin on cloud API · H1 2026 · Value: 24.6% · −0.4% a year earlier, after raising API prices about 101% this year
• MiniMax 0100.HK — Layer: Lab selling its open weights by API · Measure: Gross margin, whole company · H1 2026 · Value: 17.9% · Open-platform revenue up 703%

How the margin moved

2023 to 2025: compute got cheap. H100 rents fell from peaks near $8 an hour to $2.00 on Silicon Data’s index by December 2025. In February 2025 DeepSeek said its V3 and R1 fleet cost $87,072 a day at $2 an H800-hour against $562,027 of theoretical revenue, about $12.90 per GPU-hour, while warning that actual revenue was far lower.

2026: compute got scarce again. H100 and H200 contract prices rose about 40% between October 2025 and March 2026 as memory costs passed through. Silicon Data’s H100 index reached $2.88 on October 24, and Ornn’s settled at $3.17 on October 26, up 24.8% in a month.

Token prices split. The Flash tier fell on architecture: V4.1-Flash activates 8B parameters to read and 16B to write. The frontier tier went up: K3 launched at $3/$15, Zhipu raised API prices about 101% this year, and DeepSeek warned of a significant price increase. Revenue per GPU-hour on Blackwell racks is about three times DeepSeek’s 2025 figure even after its price cuts (the DeepInfer pack), so hosts running the rack-scale stack kept ahead of rising rents and hosts on smaller setups fell behind.

The clearest public read is Zhipu: its API sales went from a −0.4% gross margin to 24.6% in a year, with prices about doubled. Selling tokens is still a thin business for anyone without the cheapest capacity. Owning the rack, or the model, is where the margin is.

If the GPU curve inverts

GPU rents aren’t inverted now. Silicon Data says its modeled forward curve sloped down in 2025 and moved to contango this year. In mid-2026 its B200 curve was only about 8% lower at 36 months ($5.62 to $5.17), and H100 was flat past 15 months. CME has distributed that curve to subscribers since October 15, and the H100 and B200 futures it planned to list on October 5 were reported as still pending regulatory review. Our own index (tracker) started this week and doesn’t yet have enough term quotes to read a curve.

A steep inversion, a high near month and a low back end, would mean scarce capacity now and a supply wave or fast obsolescence later. A GPU-hour can’t be stored, so, as with power or freight, nothing ties the two ends together. Who gains, and who loses:

Winners and losers in a steeply inverted curve

Our assessment, roughly in order of how much each gains.

• Wins: Owners of installed GPUs that aren’t under contract — They sell at the expensive near end and stay off the cheap back end. Their cash margin is the spot rent less power.
• Wins: Hosts that own their GPUs — Token prices settle where hosts renting at prompt rates break even. Owners keep the difference, and the inversion widens it.
• Wins: Buyers who can wait — Training runs, RL and batch jobs buy the cheap back end. A long position in a deferred future gains as it rolls up toward spot, if spot holds.
• Wins: Holders of leases fixed near the 2025 low — They are long compute below market and can undercut renters or sublet where contracts allow.
• Wins: Capacity that can switch on now — Energized sites and the bottleneck parts, such as HBM. An inverted curve pays for speed.
• Wins: Brokers and market makers — Steep curves and wide gaps between chips, regions and indices bring hedging demand and spreads to capture.
• Loses: New builds that need contracts to borrow — Lower contracted rates mean less debt per GPU, which slows new supply and pulls the curve back.
• Loses: Hosts renting at the prompt price — They pay the scarce price for GPUs and sell tokens at prices set by owners.
• Loses: Long-dated bets on resale value — An inverted curve is the market pricing faster depreciation than six-year accounting lives.

How new hosts get in

Most of the 33 hosts on V4.1-Flash fit four groups: inference-engine and model companies (Modular, Relace, Morph, Decart, Reka, Cohere), GPU clouds selling tokens as well as GPU-hours (CoreWeave, Nebius, DigitalOcean, Crusoe, GMI, Prime Intellect), Chinese clouds (Baidu, Alibaba, SiliconFlow), and privacy-focused hosts (Phala, Venice).

Uptime first. Over the day to October 27, OpenRouter served V4.1-Flash 99.91% of the time; without its failover between hosts it would have been 87.81%. Hosts that stay up get the traffic. Then price: with inverse-square routing, a launch discount buys share fast. Pass the quality checks: OpenRouter publishes each host’s benchmark scores and has a routing mode for tool-calling accuracy, so heavy quantization shows. Be early, or fast: day-zero support catches a model before prices converge, and a speed tier sells at a premium (Baseten Fast posts $0.60/$2.40).

The arithmetic sets the floor. A host renting 8-GPU B200 boxes and pricing at DeepSeek’s off-peak rate needs 60% utilization to cover its rent. The hosts that last own their GPUs, fill idle ones, run the rack-scale stack, or have someone else paying.

Listed ways in

No listed company only rents GPUs and resells open-model tokens. The listed exposure is GPU clouds with token APIs on their own GPUs, Cerebras on its own chips, and the Hong Kong-listed labs selling their own open weights.

• Nebius NBIS — What it sells: Token Factory: per-token API on 60+ open models, on its own GPUs. Hosts V4.1-Flash on OpenRouter at $0.30/$1.20. · Latest figures: Q2 revenue $582M; AI business 50% adjusted EBITDA margin; bought Eigen AI for $643M
• CoreWeave CRWV — What it sells: Serverless Inference: per-token API on an open-model catalog, on its own GPUs. Hosts V4.1-Flash on OpenRouter at $0.15/$0.65. · Latest figures: Managed-inference booked ARR went from $1M to over $100M within months of launch
• Cerebras CBRS — What it sells: Fast inference cloud on its own wafer-scale chips. · Latest figures: IPO in May raised $6.4B; Q2 core cloud revenue $127.7M, up 287%; core gross margin 41%
• DigitalOcean DOCN — What it sells: Hosts V4.1-Flash ($0.225/$0.90) and Kimi K3 on OpenRouter.
• Zhipu 2513.HK — What it sells: Sells its own open-weight GLM models by API, including through OpenRouter. · Latest figures: API revenue RMB 825M in H1, up 2,736%, 86.5% of revenue; API gross margin 24.6%
• MiniMax 0100.HK — What it sells: Sells its own M-series models by API, including through OpenRouter. · Latest figures: Open-platform revenue up 703% in H1; gross margin 17.9%
• Alphabet, Alibaba, Baidu GOOGL · BABA · BIDU — What it sells: Host V4.1-Flash on OpenRouter: Vertex at the posted rate, Alibaba Cloud at 50% off, Baidu Qianfan at 84% off.

The pure-play hosts are private: Fireworks (over $1B annualized revenue, valued at $17.5B in July), Together ($8.3B), Baseten ($13B) and DeepInfra. Lambda, a GPU cloud, is raising up to $4B before a possible IPO at a $14.5B pre-money valuation.

What this is anchored on

Host prices as posted on October 27; rents as of October 20 (the DeepInfer pack) and the dates shown.

• Rack-scale yield — $38.96 per GPU-hour at DeepSeek's off-peak prices on one GB300 NVL72 rack, SemiAnalysis AgentX traffic; 954K tokens/s per GPU counting cached input · vLLM blog, Oct 23
• Recipe yields — Output tokens/s per GPU on the agentic trace: V4.1-Flash 1,155 (GB200), 1,088 (B200), 359 (H200), disaggregated · NVIDIA Dynamo, DeepSeek-V4.1-Flash recipe
• Recipe yields — Kimi K3: 91.3 (GB300), 32.9 (GB200), 3.54 (H200) output tokens/s per GPU; trace: 64K input, 400 output, 90% cache hits · NVIDIA Dynamo, Kimi-K3 recipe
• Recipe yields — GLM-5.3-Flash: 579 output tokens/s per GPU on four B200s · NVIDIA Dynamo, GLM-5.3-Flash recipe
• Hosts — 33 hosts for V4.1-Flash with input, output and cache-read prices; pooled availability 99.91% against 87.81% without routing; top apps are agent harnesses · OpenRouter, DeepSeek V4.1-Flash
• Routing — Default routing weights stable hosts by the inverse square of price · OpenRouter docs, provider routing
• Fees — No markup on host prices; 5.5% fee on credit purchases · OpenRouter FAQ
• Traffic mix — V4.1-Flash, week to Oct 14: 36.62T prompt and 807B completion tokens; Step 5 Preview reached #2 in its first week · Metaprompt, from OpenRouter's weekly table
• Traffic mix — Open-weight models carry about 60% of tokens from US-originating requests on OpenRouter, mostly Chinese models · The New Stack
• Prices — V4.1-Flash off-peak $0.15 input, $0.003 cached, $0.60 output; peak hours double · Apidog
• Prices — Kimi K3: 2.8T parameters, 104B active; providers' prices · DeepInfra
• Prices — Nine of ten K3 platforms at $3/$15 with $0.30 cached; most live within 48 hours of the weights · DEV Community
• Prices — GLM-5.3-Flash $0.15 input, $0.03 cached, $0.50 output · OrcaRouter
• Prices — DeepSeek warned of a significant price increase after launching V4 at very low rates · VKTR
• Price falls — V4.1-Flash cheapest input price down 76.7% over about 90 days, $0.30 to $0.07 · PricePerToken
• Price falls — GLM-5.3-Flash cheapest input price down 90% over about 90 days · PricePerToken
• Rents — H100 rents have come down from historical peaks near $8 an hour · Thunder Compute
• Rents — H100 $3.17 an hour on Oct 26, up 24.8% in a month · Ornn Compute Price Index
• Rents — Silicon Data H100 index $2.00 (Dec 9, 2025) to $2.88 (Oct 24, 2026); its forward curve moved from sloping down to contango · FourWeekMBA
• Rents — Mid-2026 forward curve: B200 about 8% lower at 36 months; H100 flat past 15 months · Silicon Data
• Rents — H100 and H200 contract prices up about 40% from Oct 2025 to Mar 2026 as HBM costs passed through · IntuitionLabs
• Futures — CME distributes Silicon Data's forward curves from Oct 15; H100 and B200 futures planned for Oct 5, pending regulatory review · Crypto Briefing; Spheron
• Owner cost — Owner cost per GPU-hour: B200 $1.95, GB200 $2.21, GB300 $2.65 · SemiAnalysis via InferenceX
• Margins — DeepSeek's Feb 2025 disclosure: $87,072 a day of GPU cost against $562,027 of theoretical revenue · South China Morning Post
• Margins — Zhipu cloud API gross margin −0.4% to 24.6%; API prices up about 101% this year · 36Kr
• Margins — Zhipu API revenue and share; Zhipu and MiniMax gross margins, H1 2026 · 36Kr
• Margins — MiniMax open-platform revenue up 703.1% · MiniMax interim results, HKEX
• Margins — Fireworks gross margin about 50%, targeting 60% · Sacra
• Companies — Cerebras Q2: core cloud revenue $127.7M, core gross margin 41%, $6.4B IPO · Cerebras 8-K
• Companies — Nebius Q2 revenue $582M; AI business 50% adjusted EBITDA margin · Zacks via CNHI
• Companies — CoreWeave managed-inference booked ARR from $1M to over $100M · Futurum
• Companies — Fireworks, Together and Baseten valuations and revenue · VC Cafe
• Companies — Lambda raising up to $4B at $14.5B pre-money · Let's Data Science

Method

Revenue per GPU-hour. For NVIDIA’s recipes: output tokens per second per GPU ÷ 400 output tokens per request × dollars per request × 3,600. Dollars per request: 64,000 prompt tokens, 10% at the uncached input price and 90% at the cache-read price, plus 400 output tokens. Full load means the recipe’s measured concurrency around the clock. NVIDIA’s output figures include reasoning tokens; its GLM and Kimi runs use a 15% subset of the trace and a simulated speculative-decoding acceptance rate.

The rack is vLLM’s figure on one GB300 NVL72 at DeepSeek’s off-peak prices, on SemiAnalysis’s AgentX traffic, a different trace from NVIDIA’s. The paler bars double the V4.1-Flash figures, since DeepSeek’s peak rates double every price.

Break-even utilization is rent ÷ revenue at full load. Rents are typical on-demand prices on October 20, as in the DeepInfer pack; committed rents are lower, so these overstate what a long-term renter needs. Owner costs are SemiAnalysis’s total cost of ownership from its 2025 survey.

Not counted: CPUs, networking, cache storage, staff, a GPU cloud’s own margin on top of the GPU, and capacity sitting idle between peaks. NVIDIA’s trace is more input-heavy than OpenRouter’s 45-to-1 mix.

Hosts are each host’s cheapest standard-routing endpoint on OpenRouter’s V4.1-Flash page on October 27, at discounted prices where a discount runs; list prices are kept in the data. Token counts for the week to October 26 are from our OpenRouter tracker’s daily top 50.

Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at 06.ai/trackers.

Not investment advice. Estimates are labeled as estimates.

DeepSeek open-sourced DeepInfer, the engine its own API runs on. Its benchmarks show a Blackwell GPU producing 26,300 output tokens a second at about 103 tokens a second per user, and DeepSeek says it earns $38 an hour per GPU at its off-peak prices, about three times its 2025 figure. Only Blackwell data-center GPUs and Huawei’s Ascend 950 can run it. Three days later vLLM published results within 6% of it, so the yield is now available to anyone with a Blackwell rack.

• 6.3× — Output per decode GPU vs 2025. 11,573 tokens/s at DeepSeek’s production point, against 1,850 for V3 on H800s, at 11.5× the speed per user
• $38 — Revenue per GPU-hour, off-peak. DeepSeek’s figure; vLLM gets $38.96 on agent traffic. $12.91 for V3 in 2025; a GB200 rents for $10.58
• 155T — Tokens a day through one cluster. 94% of them cached input, on our reconstruction; about 2,700 GPUs
• ≤6% — vLLM’s gap to DeepInfer on decode. Oct 23, on GB200 at batch 8 to 256; 954K tokens/s per GPU on agent traffic on a GB300 NVL72

Output per GPU falls as each user gets faster

What DeepSeek released

On October 20 DeepSeek put DeepInfer, the engine behind its own API, on GitHub under the Apache 2.0 license. It serves one model, V4.1-Flash, and runs only on NVIDIA’s Blackwell data-center GPUs (B200, B300 and their 72-GPU racks) and Huawei’s Ascend 950. Hopper GPUs such as the H100 and H200 fail its startup check, and AMD isn’t a build target. It is the offline core: the cache store, request routing and prefill/decode split a live service needs are left out, and the smallest efficient deployment is 64 prefill plus 64 decode GPUs. DeepSeek says the open build is within 5% of its internal one and that live traffic reaches about 80% of these benchmarks.

Does it change GPU usage?

Not for DeepSeek, which has run this stack since V4.1-Flash launched in September. For everyone else it set the bar, and the bar was cleared fast. On the dashboard we built on release day, DeepInfer’s curves implied about 2.0× the tokens per GPU vLLM had published on agent traffic a few days earlier (830K a second). On October 23 vLLM, the most widely used open-source engine, published decode results within 6% of DeepInfer’s across its batch sizes, using DeepSeek’s kernels, and 954K tokens a second per GPU on the same agent workload on a GB300 NVL72 rack, with the online serving DeepInfer lacks. That is 15% more than its October 17 figure, with a P90 speed of 196 tokens a second per user against a 100 target then. Anyone hosting V4.1-Flash on a Blackwell rack can now get close to DeepSeek’s own yield with an open engine, which cuts the GPUs a host needs for the same traffic.

Does it change token yields?

Against 2025, by a lot. At its production point DeepSeek gets 6.3× the output per decode GPU of its V3 fleet on H800s, at 11.5× the speed per user, and $38 of revenue per GPU-hour against $12.91, even after cutting prices by about three-quarters. Most of that arrived with the model: V4.1-Flash stores 890 bytes of cache per token, against 48,068 for V3.2. The engine is how DeepSeek cashes it in. But the yield depends on the speed sold: 26,300 tokens a second per GPU at 103 per user, falling to 26,300 at 103.

Do GPU types change differently?

Yes, and topology matters more than the chip. Decode wants 64 to 128 GPUs in one NVLink domain, which only GB200 and GB300 NVL72 racks offer. An 8-GPU B200 or B300 box keeps up at relaxed speeds and tops out at 239 tokens a second per user. Hopper and AMD gain nothing directly. Ascend 950 is a first-class target, which matters more for China’s chip mix than for anyone’s throughput.

Cost per million output tokens at market rental

The 8-GPU box is the cheaper way to buy output up to about 130 tokens a second per user; past that the rack wins, by a widening margin. The box costs more per output token than DeepSeek charges off-peak beyond about 225; the rack holds out to about 310. At equal speed the rack makes 1.1× the box’s output per GPU at 105 tokens a second per user and 6.4× at 235.

What DeepInfer changes, chip by chip

Per-GPU specs from vendor sheets and coverage, HBM as vendors list it. Scale-up domain is the most GPUs one interconnect domain holds. Rental is the typical on-demand price per GPU-hour on October 20. “What changes” is our assessment.

• GB200 NVL72 NVIDIA — DeepInfer: Runs · FP4 PFLOPS: 10 · HBM GB: 186 · Scale-up: 72 · Rent / hr: $10.58 · Full effect. The setup DeepSeek benchmarks: EP64 decode inside one 72-GPU NVLink domain, Mega MoE kernels, native FP4 weights.
• GB300 NVL72 NVIDIA — DeepInfer: Runs · FP4 PFLOPS: 15 · HBM GB: 279 · Scale-up: 72 · Rent / hr: $14.59 · Full effect expected; FlashMLA and DeepGEMM build for SM 10.3. Extra HBM matters less now that KV no longer binds; extra FP4 compute helps prefill. Rents for about 38% more than GB200 per GPU-hour.
• HGX B200 NVIDIA — DeepInfer: Runs · FP4 PFLOPS: 9 · HBM GB: 180 · Scale-up: 8 · Rent / hr: $8.02 · Partial. Runs DeepInfer, but the NVLink island caps decode at EP8: close to parity at relaxed speeds, falls behind fast. DeepSeek uses 8-GPU boxes for prefill.
• HGX B300 NVIDIA — DeepInfer: Runs · FP4 PFLOPS: 13.5 · HBM GB: 270 · Scale-up: 8 · Rent / hr: $9.99 · Partial, as B200. More HBM and FP4 compute don't lift the EP8 ceiling on fast decode; well suited to prefill.
• H200 NVIDIA — DeepInfer: No · HBM GB: 141 · Scale-up: 8 · Rent / hr: $5.33 · No direct effect. The engine's startup check rejects anything but SM 10.x and Hopper has no FP4 tensor cores. Gains arrive only as vLLM and SGLang port the ideas.
• H100 NVIDIA — DeepInfer: No · HBM GB: 80 · Scale-up: 8 · Rent / hr: $4.17 · No direct effect, as H200, with less HBM.
• H800 NVIDIA — DeepInfer: No · HBM GB: 80 · Scale-up: 8 · No direct effect. The chip DeepSeek served V3 and R1 on in 2025.
• H20 NVIDIA — DeepInfer: No · HBM GB: 96 · Scale-up: 8 · No direct effect; also the weakest compute in this table.
• MI355X AMD — DeepInfer: No · FP4 PFLOPS: 10.07 · HBM GB: 288 · Scale-up: 8 · Rent / hr: $5.60 · No direct effect. FP4-capable, but DeepInfer's kernels are CUDA or Ascend only; runs V4.1-Flash through vLLM, SGLang or ATOM.
• MI300X AMD — DeepInfer: No · HBM GB: 192 · Scale-up: 8 · Rent / hr: $3.65 · No direct effect; no FP4 matrix units, so FP4 experts are dequantized in software.
• Ascend 950 Huawei — DeepInfer: Runs · FP4 PFLOPS: 2 · HBM GB: 144 · Scale-up: 8,192 · Strategic. First-class target (DeepGEMM-Ascend, DeepEP-Ascend, TileLang). DeepSeek and Huawei built a 128-chip Ascend 950 supernode that fits the decode design; Huawei's SuperPoD scales to 8,192 chips. Peak FP4 per chip is a fifth of GB200's.

Most tokens are cache reads; most revenue isn’t

At those rates about 2,700 GPUs would carry the 155T tokens a day DeepSeek reports and bill about $0.9B a year off-peak, or $1.1B counting peak hours at double. Reported annualized revenue for all of DeepSeek was about $1.0B at the start of October, so the figures hang together. vLLM’s independent $38.96 on a GB300 NVL72 points the same way.

Memory no longer limits the batch

Then and now

DeepSeek’s February 2025 fleet averages for V3 and R1 on H800s against V4.1-Flash on DeepInfer at its production point, after the 80% factor. Different chips and price decks; V3 revenue is DeepSeek’s theoretical figure at R1 list prices.

Where the gain comes from

The model. V4.1-Flash keeps 890 bytes of cache per token, 3.9× less than V4-Flash, and activates 8B parameters when reading and 16B when writing. Its launch prices show where costs fell: cached input 57% cheaper than V4-Flash, uncached input 32%, output only 9%.

The engine. Wide expert parallelism is most of it. Bounded replay rebuilds only the last 140 tokens of sliding-window state, which DeepSeek says nearly halves prefill work. Confidence-scheduled drafting verifies fewer guessed tokens per step and adds 5% throughput at the production batch, 11% at the largest.

The hardware. The model’s expert weights are in FP4, which Blackwell, AMD’s MI355X and Ascend 950 multiply natively and Hopper does not. Decode wants a big NVLink domain, which is why the rack beats the box.

What it means

For GPU economics. A GB200 rents for about $10.58 an hour and, served this way, earns $38 at DeepSeek’s off-peak prices. Serving open models on Blackwell racks is profitable at market GPU prices, so long as the traffic is cache-heavy agent work, and vLLM now makes that available to every host. Third-party hosts charge about DeepSeek’s peak rate, so there is room for their prices to fall toward its off-peak ones (the token-price pack).

For older chips. Hopper rents for less but gets none of this directly, and it can’t do the FP4 math. The yield gap between a Blackwell rack and an H100 widens with each release like this one, which should weigh on Hopper rental prices over time.

For China. DeepSeek built Ascend 950 support in from the start and is reported to be planning a 160,000-chip Huawei data center. DeepSeek hasn’t published any Ascend throughput, which is the number to watch.

What this is anchored on

Decode and prefill points are read off DeepSeek’s benchmark charts; rental prices and DeepInfer figures as of October 20, vLLM’s as of October 23.

• DeepInfer — Engine, supported hardware, decode and prefill benchmark charts, $38 per GPU-hour, 155T tokens a day · DeepInfer repository
• DeepInfer — Production concurrency, draft verification table, balance factor, KV storage, minimum deployment · DeepInfer deployment Q&A
• V4.1-Flash — Parameters, 890 bytes of KV cache per token, KV series by generation · DeepSeek-V4.1-Flash technical report
• V4.1-Flash — KV-cache series by generation · Context Studios
• Prices — V4-Flash and V4.1-Flash off-peak and peak rates; peak hours a week · Apidog
• vLLM — Within 6% of DeepInfer on decode; 954K tokens/s per GPU on AgentX, GB300 NVL72; $38.96 per GPU-hour (Oct 23) · vLLM blog
• vLLM — 830K tokens/s per GPU on AgentX at 100 tok/s per user (Oct 17) · AICoder
• vLLM — 5× high-throughput gain on V4.1-Flash since day 0 · vLLM blog
• AgentX — Median input and output tokens, cache-hit rate · AgentX workload
• AgentPerf — Tokens/s per MW by system on V4.1-Flash · Artificial Analysis
• AgentPerf — MI355X, B200 and B300 results · CryptoBriefing
• Rental — Typical on-demand price per GPU-hour, Oct 20 · Savrn AI index
• Ascend 950 — Specs · TechRadar
• Ascend — DeepSeek’s Ascend tooling; planned Huawei data center; $1B run-rate · The Next Web
• DeepInfer — Release coverage · MarkTechPost
• DeepInfer — Release coverage · The Next Web
• V3 / R1, 2025 — DeepSeek’s Feb 2025 inference-system figures on H800s · Analytics Vidhya

Method

Reading the charts. DeepSeek’s decode chart defines output per GPU as per-user speed × requests per GPU, and our readings meet that within 0.5%. Decode output assumes 75% acceptance on 3 draft tokens, as DeepSeek’s chart does; only the step rate is measured. The EP64 curve ran inside one NVLink domain, so on a GB200 or GB300 NVL72 rack; DeepSeek doesn’t say which.

The production point is DeepSeek’s draft-verification table at 60 sequences per GPU, which it says is where production runs. “After the live-traffic factor” multiplies by 80%, DeepSeek’s ratio of live traffic to its perfectly balanced benchmarks.

Costs use typical on-demand rental per GPU-hour, not DeepSeek’s own costs, which it doesn’t disclose. HGX B200 output is the EP8 curve scaled by DeepInfer’s own peak-rate table (2,250 against 2,500 TFLOPS).

The revenue check assumes every GPU runs at the production point around the clock, with prefill at 32,768-token steps. The comparison with vLLM’s earlier result priced one agent request (131,072-token context, 4,096 uncached, 444 output tokens) in DeepInfer GPU-seconds; it set an offline benchmark against a live replay, so it was indicative, and vLLM’s like-for-like decode test on October 23 supersedes it.

What DeepSeek doesn’t disclose: which Blackwell part ran the benchmarks, whether the cluster behind the daily token figure is NVIDIA, Ascend or both, fleet size, the real cache-hit share, and any Ascend 950 throughput.

Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at 06.ai/trackers.

Not investment advice. Estimates are labeled as estimates.

Before AI, both clouds were slowing: Azure from 50% growth in 2021 to 26% in mid-2023, AWS from 40% to 12%. Now Azure grows 43% and AWS 37%. We estimate AI accounts for about 26 of Azure’s 43 points of growth this year and 17 of AWS’s 33. Renting out compute takes a lot of capital, and so far each dollar has earned more at Amazon than at Microsoft: in 2026, about 40 cents a year of AI revenue for each dollar of AI capex already in service at Amazon, and 26 cents at Microsoft. Count only cloud revenue, charge all the AI capex, value the stakes in OpenAI and Anthropic at zero, and the spending through 2027 earns about 12.8% a year at Amazon and 3.0% at Microsoft. Both numbers turn on how quickly capex starts earning, and on two labs that don’t yet make money paying their bills. These are the middle of our ranges.

• 26 of 43 pts — AI’s share of Azure’s 2026 growth about $39B of revenue, 32% of Azure, against a no-AI trend of 17% a year
• 17 of 33 pts — AI’s share of AWS’s 2026 growth about $31B, 18% of AWS, against 14.5%; in line with Amazon’s own AI run-rates
• 12.8% — Amazon’s return on AI capex through 2027 a year from AWS alone, stakes at zero. 8.6% charging the $68B it put into OpenAI and Anthropic; 7.2% if the labs pay half their bills from 2028
• 3.0% — Microsoft’s return on AI capex through 2027 from Azure alone; 6.0% counting the Copilots, 2.4% charging the money put into the labs, −6.0% if the labs pay half from 2028

Both clouds bottomed in 2023; AI turned them back up

What changed

Microsoft. Azure’s growth fell from 50% in the quarter to September 2021 to 26% in the quarter to June 2023, as customers cut their cloud bills. It has climbed since, to 43% in the quarter to June 2026 (42% on the definition Microsoft now reports), and Microsoft guides 44–45% in constant currency for the quarter to September. Azure passed $100B of revenue in the year to June. Microsoft says demand still exceeds the capacity it can bring online and expects to roughly double that capacity in two years.

Amazon. AWS slowed from 40% at the end of 2021 to 12% in mid-2023. It grew about 17% in the first half of 2025, then 20%, 24%, 28% and, in the second quarter of 2026, 37%, its fastest in 18 quarters. Amazon says it won’t have enough capacity for all of 2026’s demand and expects the same in 2027. The whole market is moving: Synergy Research puts cloud infrastructure spending up 43% in the second quarter, with generative-AI services up 165%.

Microsoft said how much of Azure’s growth was AI, until AI was 16 points of it

Microsoft’s points show AI rising from about 1 point of Azure’s growth in mid-2023 to 16 points by March 2025, when AI services were about 19% of Azure. The rest of Azure slowed as it did so, from 24 points of growth in the quarter to March 2024 to 17 a year later. Microsoft has since given only its AI business’s run-rate, which “surpassed an annual revenue run rate of $37 billion” in April 2026, Copilots included. At a 17% trend for the rest, AI adds about 20 points of Azure’s growth through 2025 and 26 by mid-2026, when it is about 34% of Azure. By the same chain, AI was already about 5% of Azure in 2022. Earlier capex paid for that, so for the returns we leave it, and its growth, out.

By our estimate AWS will spend about as much on capex this year as it takes in

The backlogs say the demand is there, and that much of it is two labs

Is growth accelerating?

Yes, at both, and capacity is the limit. Azure grew 39–40% for four quarters, then 43%, and Microsoft guides 44–45%. AWS has accelerated for five straight quarters. Microsoft’s commercial backlog reached $678B in June, up 84%; excluding OpenAI it grew 25%, and all of the quarter’s increase came from customers other than the frontier labs. AWS’s went from $195B to $496B in a year, with an average remaining life of 6.4 years, after OpenAI added $100B of commitments and Anthropic more than $100B.

By our estimate AI adds about $39B to Azure this year and $31B to AWS

Azure, 2026: $122B of revenue, $39B of it (32%) from AI at a 17% no-AI trend. AI adds about 26 of this year’s 43 points of growth.

By business

How we split AI from trend

Both trends are checked against what the companies themselves disclose. Azure, 17% a year from 2025: by Microsoft’s own split, Azure outside AI grew about 24% in 2024 and was down to 19% by early 2025. At 17%, Azure’s AI services ran at about $30B a year in early 2026; add the Copilots and that fits Microsoft’s AI run-rate of more than $37B in April. AWS, 14.5%: it grew 13% in 2023 and 19% in 2024, when AI was a small business. At 14.5%, AWS’s AI revenue ran at about $20B a year in the first quarter of 2026 and $32B in the second; Amazon says it was “over $15 billion in Q1 2026” and above $25B in the second quarter. Much above 16%, our method falls short of Amazon’s own figures, so the simulation drops draws that do. The rest of each gap we leave to things that aren’t AI.

At these settings AI adds 19 points to Azure’s growth in 2025 and 26 in 2026, 32% of this year’s revenue, and 5 and 17 points to AWS’s, 18%. AI’s points are the growth in AI’s revenue over last year’s revenue, so they and the rest add up to the year’s growth. Microsoft’s Copilots aren’t counted: Microsoft 365’s commercial cloud revenue has grown 14–16% a year in constant currency since mid-2024, no faster than before Copilot.

From the cloud alone, Amazon’s AI capex clears 10%; Microsoft’s falls well short

At a 60% cash margin and the base no-AI trend (Azure 17%, AWS 14.5%): Amazon 13.6% through 2026 and 12.8% with 2027, 8.1% all-in; Microsoft 4.1%, 3.0% and 0.7%.

How the return is counted

Returns. Cloud revenue needs capacity, so after 2026 we tie AI revenue to it. We assume capex starts earning two quarters after it is spent: servers earn within weeks of landing, data centers a year or more after work starts. On that basis each dollar of incremental capex brought in about 40 cents a year of AI revenue at Amazon in 2026 and 26 cents at Microsoft. Microsoft’s is lower, probably because more of its capex has gone into things Azure’s revenue doesn’t show: its own models and Copilots, capacity for OpenAI’s research, and data-center sites ahead of use. At Microsoft we also leave out the 10% of Azure’s AI revenue we estimate runs on capacity it rents from neoclouds (it paid CoreWeave about $3.4B in 2025), and that capacity’s cost. Each dollar of capacity keeps renting at that rate, less 2.5% a year as compute gets cheaper, at a 60% cash margin before depreciation. That is a little above AWS’s 52% (operating income plus depreciation) in 2025; AWS’s operating margin was 39% in the second quarter of 2026, up from 33% a year earlier.

Costs. All capex above a pre-buildout intensity: 14% of Microsoft’s no-AI revenue (its 2019–23 average, finance leases included) and 31% of AWS’s, about $449B at Microsoft and $367B at Amazon over 2023–27. Microsoft’s 2026 capex is its April guide, $190B, before it moved more data-center leases off capex. Servers, 60% of capex (Microsoft has said half to two-thirds), are replaced every 6 years at Microsoft and 5 at Amazon, at 75% of their cost, for the 75% of the fleet that rents to customers. Servers get bonus depreciation for tax (100% since 2025), data centers go over 20 years. Nothing is counted after 2040, though the data centers will still stand.

Results. Through 2026: 13.6% at Amazon, 4.1% at Microsoft. Adding 2027’s build: 12.8% and 3.0%, paid back in 2033 and 2037. The cloud’s return barely depends on whether 2027 is spent, since new capacity earns what the old does as long as it rents at today’s rates; if 2027’s rents for 20% less, the returns fall to 10.4% and 1.3%. The biggest unknown is how long capex takes to start earning, because that sets how much each dollar in service earns today. If capex earns from one quarter after it is spent, more of it is already in service, each dollar earns less, and the returns fall to 6.2% at Amazon and −1.3% at Microsoft; from three quarters, they rise to 20.3% and 7.5%. Every ten points of margin move the returns four to five points. On the same question, Meta’s ads return about 18% and Alphabet’s ads and Cloud 13% (the ad-payback pack).

Payback comes in 2033 at Amazon and 2037 at Microsoft

If the labs can’t pay, Microsoft’s return goes negative and Amazon’s drops below 10%

Two customers, and two stakes

The labs are much of the AI revenue. Microsoft says nearly 90% of Microsoft Cloud’s $214B of revenue in fiscal 2026 came from customers other than frontier-model companies. That leaves the labs about $22–26B, against our estimate of about $29B for all of Azure’s AI revenue that year, so we put them at 80% of it; OpenAI was 45% of Microsoft’s commercial backlog at the end of 2025. At Amazon, OpenAI’s and Anthropic’s commitments are about half the backlog, and we use 50%. The labs pay with money they raise: OpenAI’s own plan has it burning about $278B of cash over 2026–30 (the lab-funding pack), though both labs’ revenue is compounding fast (the lab-revenue pack). If from 2028 they paid only half their bills, and the capacity they left were re-let at half price, Microsoft’s return through 2027 would fall to −6.0% and Amazon’s to 7.2%.

The stakes are where the payoff might be. This pack counts them at zero. Amazon carried its Anthropic notes and preferred stock at about $190B at the end of June, against $18B invested, and has booked about $80B of pre-tax gains on them since early 2025; it has also put $50B into OpenAI, which committed another $100B to AWS. Microsoft’s roughly 27% of OpenAI was valued at about $135B in October 2025; OpenAI has since raised money at $852B. Charge the cash put into the labs and value the stakes at nothing, and the returns through 2027 are 8.6% at Amazon and 2.4% at Microsoft; all-in, with running costs and Microsoft’s neocloud contracts too, 8.1% and 0.7%. Count the stakes at their marks and the AI build has already paid for a good part of itself, on paper.

Amazon’s build clears 10% in about two-thirds of our draws; Microsoft’s in fewer than one in ten

When does the cash come back?

AI capex through 2027 clears a 10% return in 64% of our draws for Amazon and 8% for Microsoft, and clears zero in 97% and 72%. All-in, it clears 10% in 36% and 3%. The medians, 12.5% and 3.0%, sit within a point of the base case. Andy Jassy’s case is that the returns come later: “The FCF and ROIC for these investments are cumulatively quite attractive a couple years after being in service.” Amazon’s free cash flow was −$7.6B over the twelve months to June. Microsoft’s was $19.6B in the June quarter, down 23% on a year earlier, and it guides capex up again in fiscal 2027.

What to watch

Microsoft, October 28. Azure against the 44–45% guide, now on the narrower definition; capex against “over $50 billion” for the quarter and any figure for fiscal 2027; whether the backlog keeps growing outside OpenAI.

Amazon, October 29. AWS against analysts’ 30–35%; capex against $220B for the year and a first word on 2027; the backlog after the OpenAI and Anthropic expansions; the AI run-rate, last above $25B.

The labs’ funding. OpenAI’s next round and Anthropic’s IPO decide whether the bills behind half of AWS’s backlog, and most of Azure’s AI revenue, get paid. In our base case a lab dollar is worth the same as any other; the labs chart shows what happens when it isn’t.

What this is anchored on

No-AI trends, AI’s share of above-trend revenue, the lag, the second half of 2026, 2027 capex, margins, replacement, rented capacity, the labs’ share of AI revenue, timing of the all-in costs and the returns are our estimates, built from the sources below.

• Microsoft FY26 Q4 — Revenue $90.0B, +18%, and $331.8B for the fiscal year; Azure and other cloud services +43%; Azure revenue passed $100B for the first time; free cash flow $19.6B, down 23% · Microsoft FY26 Q4 results (8-K), Jul 29
• Microsoft FY26 Q4 call — Capex $41B, about two-thirds short-lived assets (GPUs and CPUs), including $5.6B of finance leases; data-center useful life extended from 15 to 25 years, so more future data-center leases will be operating leases, outside capex; calendar 2026 capex about $175B on that basis, plans otherwise unchanged; fiscal Q1 2027: Azure about 45% in constant currency, capex over $50B; commercial RPO $678B, +84%, +25% excluding OpenAI, all of the quarter’s increase from customers outside frontier-model companies; Microsoft Cloud gross margin 65%; over 30M paid Microsoft 365 Copilot seats; demand exceeds capacity; capacity to roughly double in two years · Microsoft FY26 Q4 call, via The Motley Fool
• Microsoft FY26 Q4 call — Nearly 90% of fiscal 2026 Microsoft Cloud revenue came from customers outside frontier-model companies; a $3.2B gain from the Anthropic investment · MarketBeat, Jul 29
• Microsoft, Sep 2026 — Azure revenue in dollars on a narrower definition: fiscal 2025 $72.6B (quarters $16.0B, $17.3B, $18.5B, $20.7B), fiscal 2026 $101.9B ($22.4B, $24.1B, $26.0B, $29.4B); June 2026 quarter +42% on this basis; fiscal Q1 2027 guide 44–45% in constant currency · Microsoft FY27 segments and metrics (8-K), Sep 2026
• Microsoft FY26 Q3 — AI business “surpassed an annual revenue run rate of $37 billion”, up 123% · Microsoft FY26 Q3 results (8-K), Apr 29
• Microsoft FY26 Q3 call — Capex $31.9B; calendar 2026 capex about $190B, including about $25B from higher component prices · Sherwood News, Apr 30
• Microsoft FY26 Q2 call — “Approximately 45% of our commercial RPO balance is from OpenAI,” of $625B · Microsoft FY26 Q2 call, via roic.ai
• Microsoft AI points — Azure growth included 13 points from AI services, which grew 157% (December 2024 quarter) · Microsoft FY25 Q2 call, via MarketBeat
• Microsoft AI points — 16 points from AI services in the March 2025 quarter · Microsoft FY25 Q3 call, via MarketBeat
• Microsoft AI points — Roughly 12 points from AI services in the September 2024 quarter, “similar to last quarter” · Microsoft FY25 Q1 call, via Nasdaq
• Microsoft AI points — 8 points in the June 2024 quarter; 7 and 6 in the March 2024 and December 2023 quarters; roughly 3 in September 2023 and 1 in June 2023 · Microsoft earnings calls, Jul 2023–Jul 2024
• Microsoft FY24 recast — Azure growth for fiscal 2024 restated to 31%, 33%, 35% and 34% by quarter · Microsoft metric recast (8-K), Aug 2024
• Microsoft capex — Capex including finance leases by quarter, $6.6B (September 2022) to $41B (June 2026); finance leases $11.1B, $6.7B, $4.7B and $5.6B in the quarters of fiscal 2026, mostly for large data-center sites · Microsoft earnings calls, fiscal 2023–26
• Microsoft baseline — Cash capex from $13.9B in fiscal 2019 to $28.1B in fiscal 2023, about 12% of revenue; in fiscal 2023, $31.9B including finance leases. Our 14% baseline adds finance leases · Microsoft results and calls, fiscal 2019–23
• Microsoft leases — Data-center leases signed but not yet commenced of $329.1B at June 30, 2026, starting from fiscal 2027 through fiscal 2033 · btw.media, on the fiscal 2026 10-K
• Microsoft–OpenAI — Microsoft holds about 27% of OpenAI Group PBC, valued at about $135B; OpenAI contracted an incremental $250B of Azure services; Microsoft loses its right of first refusal on compute · Microsoft, Oct 28, 2025
• Microsoft–OpenAI — Microsoft stops paying OpenAI a revenue share; OpenAI’s payments to Microsoft continue through 2030 under a cap; Microsoft’s license becomes non-exclusive · OpenAI, Apr 27
• Microsoft–OpenAI — $13B of funding commitments to OpenAI, $11.8B funded at Mar 31, 2026 · Microsoft 10-Q, quarter to Mar 31
• Microsoft–Anthropic — Anthropic committed to buy $30B of Azure compute; Microsoft to invest up to $5B in Anthropic · Microsoft, Nov 18, 2025
• Microsoft neoclouds — Dedicated GPU capacity for Microsoft worth about $17.4B through 2031, up to about $19.4B with additional services · Nebius (6-K), Sep 2025
• Microsoft neoclouds — A five-year, $9.7B AI cloud contract with Microsoft · IREN (8-K), Nov 2025
• Microsoft neoclouds — Neocloud contracts above $60B: Nscale about $23B, Nebius up to $19.4B, IREN $9.7B, Lambda and CoreWeave; mostly five-year terms · Introl, citing Bloomberg, Nov 2025
• Microsoft neoclouds — The Nebius capacity is for Microsoft’s internal teams building models and a consumer assistant · Tom’s Hardware, citing Bloomberg, Oct 2025
• CoreWeave — About 67% of 2025 revenue of $5.1B came from Microsoft · CoreWeave 10-K, 2025
• Amazon Q2 2026 — AWS revenue $42.2B, +37% (36.7%), operating income $16.6B; “Exceeded a $25 billion annual revenue run rate for AWS’s AI business”, and the same for Amazon’s chips business; Q3 guidance: revenue $197–202B, operating income $22.5–26.5B; free cash flow −$7.6B over twelve months · Amazon Q2 2026 results (8-K), Jul 30
• Amazon Q2 2026 — Performance obligations, primarily AWS, about $496B with a 6.4-year average remaining life; OpenAI expanded its $38B commitment by $100B over 8 years, Anthropic its own by more than $100B over 10 years; $50B invested in OpenAI ($21.3B of it after June 30); $8B of Anthropic notes plus $10B of preferred, carried at about $190B · Amazon 10-Q, quarter to Jun 30
• Amazon Q2 2026 call — Cash capex $53.1B in the quarter; about $220B for 2026, up from about $200B on the higher cost of memory; “we will still not have enough capacity to meet all the demand we have in 2026”, and likely the same in 2027 · Amazon Q2 2026 call, via The Motley Fool
• Amazon letter — “AWS’s AI revenue run rate is over $15 billion in Q1 2026”; “The FCF and ROIC for these investments are cumulatively quite attractive a couple years after being in service” · Andy Jassy, 2025 letter to shareholders, Apr 9
• Amazon Q1 2026 — AWS revenue $37.6B, +28%; chips business above a $20B run-rate; backlog about $364B · Amazon Q1 2026 results (8-K), Apr 29
• Amazon 2025 — AWS revenue $128.7B, operating income $45.6B; 2025 cash capex $128.3B; about $200B of capex planned for 2026 · Amazon Q4 2025 results, Feb 5
• Amazon segments — AWS net additions to property and equipment $24.8B, $53.3B and $96.5B in 2023–25; AWS depreciation and amortization $21.5B in 2025 · Amazon 10-K, 2025 (segment note)
• Amazon segments — AWS net additions to property and equipment $13.1B, $16.5B, $22.0B and $27.8B in 2019–22 · Amazon 10-Ks, 2020 and 2022 (segment notes)
• Amazon servers — Useful life of a subset of servers and networking equipment cut from six to five years from January 2025 · Amazon 10-K, 2024
• Amazon–OpenAI — $15B invested, $35B more on conditions; the $38B AWS agreement expanded by $100B over 8 years, about 2 GW of Trainium · OpenAI, Feb 27
• Amazon–Anthropic — Anthropic commits more than $100B over ten years to AWS, up to 5 GW; Amazon invests $5B, up to $20B more · Anthropic, Apr 20
• AWS Q3 — Analysts expect AWS growth of 30% to 35% in the third quarter; Amazon reports October 29 · Benzinga, Oct 22
• Microsoft FY27 Q1 — Results after the close on October 28 · Microsoft, via Finviz
• Cloud market — Q2 2026 cloud infrastructure spending $143.4B, +43%, the fastest in eight years; GenAI-specific services +165% · Synergy Research, Jul 30
• OpenAI — Plan of July 2026: negative free cash flow of about $278B over 2026–30 · FT, via Reuters, Sep 18
• OpenAI — Round closed March 31 at $122B and an $852B valuation · TechCrunch, Mar 31
• Anthropic — At least $518B of compute commitments over about a decade, including Amazon $110B and Microsoft $31.4B · Reuters, on the draft prospectus, via Zawya
• History — Microsoft quarterly revenue, Azure growth, capex and RPO, fiscal 2022–26; AWS quarterly revenue 2019–2026 · Microsoft and Amazon earnings releases and filings

Method

Clock. Reported through June 2026: Microsoft’s fiscal 2026 and Amazon’s second quarter, both mapped to calendar years. Azure’s revenue is Microsoft’s new dollar series for fiscal 2025–26; earlier quarters are backed out from the growth it reported. The second half of 2026: Azure up 44% and AWS up 32.5%. Capex 2026: Microsoft $190B (its guide before the lease change), Amazon $220B plus finance leases with 75% to AWS; 2027 our estimates.

AI’s share. Azure: AI services as Microsoft split them through March 2025, less the AI Azure already had in 2022 grown with the rest of Azure; after that, 90% of revenue above a 17% trend. AWS: 90% of revenue above 2023’s revenue grown at 14.5% a year, consistent with Amazon’s AI run-rates. Trends hold through 2027 and fade to 6% by 2035.

Returns. Cash flows from 2022 to 2040, no terminal value. In: AI revenue at a 60% cash margin. Through 2026 it is the estimate above (at Microsoft, less the 10% we put on rented capacity); after that, each dollar of incremental capex earns what 2026’s AI revenue implies per dollar of capex then in service (26 cents at Microsoft, 40 cents at Amazon), from two quarters after it is spent, fading 2.5% a year. Out: capex above 14% of Microsoft’s no-AI revenue and 31% of AWS’s; servers, 60% of capex, replaced every 6 or 5 years at 75% of their cost for the 75% of the fleet that rents to customers; cash tax at 19% and 17%, with servers getting first-year bonus depreciation (80% in 2023, 60% in 2024, 100% since; the rest over five years) and data centers over 20 years, losses offsetting other income. All-in adds the cash put into the labs (Amazon: $8B of Anthropic notes over 2023–25, $10B of preferred and $50B into OpenAI in 2026; Microsoft: about $9B of its 2023 OpenAI commitment over 2023–25, our timing, and up to $5B into Anthropic), half of Microsoft’s neocloud contracts (about $75B over 2026–33, our timing), and running costs of 3% a year of the cost of the fleet that doesn’t rent to customers, while its servers last. The labs cases cut the labs’ payments from 2028 and re-let what they stop paying for at half price. The base case is the middle of every range in the simulation.

What’s left out. The stakes in OpenAI and Anthropic, and OpenAI’s revenue share to Microsoft. Microsoft’s Copilots (6.0% if counted at $7–10B a year), Amazon’s use of AI in retail and ads, and every company model or assistant. Operating leases Microsoft signs from fiscal 2027 sit outside its reported capex; our 2026–27 figures keep the old definition. Monte Carlo: 8,000 draws per case with a fixed seed. Model and data: scripts/packs/cloud_payback.py in the site’s repository.

Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at 06.ai/trackers.

Not investment advice. Estimates are labeled as estimates.

AI accelerator shipments roughly double from about 14M packages in 2026 to 26.5M in 2030, and HBM can’t grow fast enough to give each one the memory on today’s roadmaps. At announced specs, HBM supply falls 18% short of demand in 2027 and 29–38% short in 2028–30. To balance, the average accelerator can carry about 270 GB of HBM in 2027 and 430 GB in 2030, growth of 17% a year against 28% on the roadmaps. The trims now being reported, from Rubin Ultra’s cut to 8-Hi stacks on next year’s ASICs, do roughly that; price does the rest.

• ~270 GB — HBM per accelerator, 2027 the fleet average supply can support. Rubin and B300 carry 288 GB
• ~430 GB — HBM per accelerator, 2030. +17% a year from 2027, against +28% on announced roadmaps (~690 GB)
• 29% short — At announced specs, 2028. HBM supply against demand: 18% short in 2027, 38% by 2030. At the trimmed specs, within ±3% every year
• −22% to +52% — Biggest swing, 2028 unit volumes, from the vendors’ deal books to a power-limited build, at trimmed specs

At announced specs, HBM supply falls 18–38% short from 2027; the reported trims close the gap

Base units, base supply: at announced specs supply is 18% short in 2027 and 38% short in 2030; at trimmed specs 2% short in 2027 and 2% short in 2030.

What the balance says

How we count. Supply is HBM wafer starts times the good terabytes each wafer yields. Demand is units times HBM per package, bought about four months before the accelerator ships, plus about 6% for stacks lost in 2.5D assembly and physical capacity above the usable figure. The lead matters when growth is steep: purchases run 37% above the bits inside 2026’s shipments, because 2027’s step-up is bought early, and 13–17% above in 2027–30.

2026 is roughly balanced (−2%). That fits the price signal: HBM contract prices paused or slipped in 2026 while commodity DRAM rose about sixfold. In 2025, supply (about 2.8 EB) ran well ahead of the bits shipped inside accelerators. The excess went into inventory across the chain: Nvidia’s own rose from about $10B to $21B over its fiscal 2026 and reached $31.6B by July 2026.

2027 is the crunch at announced specs. Demand jumps with TPU 8, Trainium 3–4, MI450, MTIA and the first Rubin Ultra builds. Supply grows only 54%: wafer input rises 31% and good bits per wafer 18%, less than node shrinks alone would give because the HBM4 die is larger for its doubled interface. The announced specs would need 7.3 EB against about 6.0 EB available.

2028–30 turns on two things we can’t know yet: how fast memory per package grows again, and how many packages the power grid can absorb. New cleanrooms lift supply 33% in 2028 and about 24% a year after. If memory per package grows about 17% a year, the market stays balanced; if it follows the announced roadmaps, supply falls more than a third short.

How this compares with published views

Supply ÷ demand − 1. Goldman sits between our two cases; J.P. Morgan’s deeper −16% came in September, after the first trims were reported.

The commodity DRAM squeeze is the same story one step removed. With nearly every new big-three wafer going to HBM, conventional bit growth has to come from node shrinks and CXMT. AI then takes a second bite through CPU memory: Vera’s LPDDR5X, AMD Helios’s 36 TB of DDR5 per rack, and agentic CPU servers. Counting HBM at its wafer cost, AI systems take roughly 27% of DRAM wafers in 2026 and about 43% in 2027, by our estimate from the rack configurations below. Conventional contract prices compounded roughly sixfold from 3Q25 to 4Q26, with 4Q26 still up 10–15% (TrendForce). IDC expects 2026 PC shipments to fall 11.3% and smartphones 13%. Published 2027 DRAM sufficiency ranges from −1.9% (DS Investment) to −5.9% (Goldman); Micron calls 2027–28 “tighter than 2026”.

Is Nvidia really cutting memory per GPU?

Partly. No shipping data-center GPU has lost HBM: Rubin keeps 288 GB, the same as Blackwell Ultra. But Nvidia is reported to be halving Vera’s CPU memory for 2027 builds and cutting Rubin Ultra’s HBM by 75–81% per package against its announced spec. HBM per GPU has stopped growing this generation, and per compute die it is set to fall in the next.

• Rubin (VR200), shipping since Aug 2026 — What changed: Still 288 GB of HBM4 on eight 12-Hi stacks. Listed bandwidth fell from 22 TB/s (CES) to 19.2 TB/s, no reason given · Status, Oct 2026: Nvidia’s page. A ZDNet Korea report of 8-Hi orders was contradicted by Counterpoint, which says most Rubin stays 12-Hi
• Vera CPU (LPDDR5X on SOCAMM2) — What changed: 1.5 TB → 768 GB per CPU; ~54 → ~28 TB per NVL72 rack · Status, Oct 2026: Reported for 2027 supply plans by TrendForce, Daishin and SemiAnalysis (June). Nvidia calls 768 GB a “flexible SKU”; its spec still says “up to 1.5 TB”
• Rubin Ultra (2027–28) — What changed: Announced: 1 TB of HBM4E on 4 dies. Reported: ~192–256 GB of 8-Hi HBM4 on 2 dies · Status, Oct 2026: SemiAnalysis; The Information. Nvidia: “our roadmap is intact”; no confirmation
• Rubin CPX (128 GB GDDR7) — What changed: Absent from Nvidia’s October Vera Rubin page; the SRAM-based Groq 3 LPX rack appears instead · Status, Oct 2026: No source ties this to the GDDR7 shortage
• GeForce RTX 50 — What changed: Fewer cards (a 30–40% cut reported for the 16 GB RTX 5070 Ti and 5060 Ti in 1H26), not less memory per card · Status, Oct 2026: Nvidia: “memory supply is constrained”

Per compute die, the cut is real. Rubin carries 144 GB per die; the reported Rubin Ultra carries 96 GB per die, a third less, against 256 GB per die in the original plan. Half of the package-level drop comes from going from four dies to two, which sources put down to packaging and rack issues rather than memory.

It is not just Nvidia. AMD reportedly built a 288 GB MI450 for Meta beside the 432 GB MI455X. TrendForce says GPU and ASIC makers are “considering reducing HBM capacity per device” and will evaluate 8-Hi stacks in 2027. The tell: 8-Hi stacks now carry a 10–30% per-gigabit premium over 12-Hi, the reverse of normal pricing.

Why. SemiAnalysis gives three reasons for Rubin Ultra. There are not enough HBM wafers to build 12-Hi stacks for the logic and CoWoS capacity Nvidia has secured for 2027, and 8-Hi yields more stacks per wafer. HBM cost jumps in 2027: GF Securities estimates memory would have reached about 29% of a Vera Rubin system’s bill of materials without the Vera cut, against Nvidia’s preferred 20%. And workloads are shifting: a 576-GPU domain of 192 GB parts (about 110 TB) still dwarfs a 20.7 TB NVL72.

Nvidia has guided its gross margin down from 75% to 71–72% by the quarter ending January 2027 on “extreme pricing conditions in memory”. Its supply and capacity purchase commitments, mainly memory, rose from $119B to $279B in one quarter. The logic is to protect units by trimming memory per unit: TrendForce describes the Vera cut as “lowering memory per module while increasing total module shipments”, a supply-driven compromise, not a drop in demand. We model the reported trims as the base case and the announced specs as the alternative.

Spec trims cut HBM demand by 1.2 EB in 2027 and 3.5 EB in 2028, mostly at Nvidia

Supply supports about 270 GB per accelerator in 2027 and 430 GB in 2030

How much memory per accelerator the market can afford

To balance, fleet-average HBM per accelerator can be about 270 GB in 2027, 320 GB in 2028 and 430 GB in 2030: growth of about 17% a year, against 28% on the announced roadmaps. That is feasible. 288 GB stays the flagship through 2027, next-generation parts move to 8-Hi stacks, and capacity beyond that moves off HBM. It is a fleet average, so a 432 GB part has to be offset by a 144–216 GB part elsewhere.

• 2027 (~270 GB). Rubin, TPU 8i and the 288 GB MI450 sit at the top of the affordable range. The 432 GB MI455X and the 384–512 GB MTIA 500 fit only if Trainium 3 (144 GB), TPU 8t (216 GB) and Maia 200 (216 GB) keep the average down.
• 2028 (~320 GB). Rubin Ultra at 192–256 GB on two dies fits. The announced four-die, 1 TB version would absorb about a fifth of 2028 supply on about 15% of units.
• 2029–30 (~380–430 GB). Room returns for Feynman-class parts at 400–600 GB, as long as custom ASICs stay near 300–420 GB.

Per compute die, that means about 100–130 GB on Rubin Ultra (Rubin: 144 GB) and roughly 200 GB per die by 2030.

Workloads can live with it. MLPerf Inference v6.1 compares GB300 (278 GB) with GB200 (186 GB) at the same 8 TB/s. A 671B model and small models that can’t be capacity-bound gained the same ~1.29×, so the gain from capacity itself was about zero. Our modelling of a one-third cut, 288 GB to 192 GB through 8-Hi stacks at unchanged bandwidth:

• 0–10% throughput loss for today’s dominant pattern: 0.7–2T-parameter MoE models on 72 or more GPUs with wide expert parallelism, 32K–400K context. Weights take about 20–45 GB per GPU at that width, depending on precision and replication; the rest is KV cache.
• 15–30% loss for 10T-class models, sparse-attention long-context agents, or idle sessions parked in HBM. About 45% in the worst combination.
• 0–12% for training.
• Concurrency falls more than throughput. Sessions per GPU drop 36–41%, because weights are a fixed charge, but per-user decode speeds up, so tokens per GPU fall only 5–11%. Token-priced services barely notice; session-priced agent hosting does, unless KV moves to another tier.

The hard floor is about 155 GB per GPU, below which a 10T-parameter model at FP8 no longer fits in a 72-GPU domain (about 90 GB at FP4). These figures assume FP4 weights, FP8-or-smaller KV cache and an off-HBM KV tier, all shipping on Blackwell and Rubin.

The economics push the same way. At 2027 HBM4 prices of $32–40/GB, 288 GB costs about $9–12K per package and 1 TB about $33–41K. Morgan Stanley puts memory at 25–30% of a Vera Rubin rack’s cost, up from 5–10% in earlier racks. A 12-Hi to 8-Hi switch cuts bits per GPU by a third; with 8-Hi’s 10–30% per-Gb premium, HBM spend per GPU falls 13–27%.

The technology is not the limit; wafers are. Per-stack bandwidth depends on interface width and pin speed, not height, so a shorter stack keeps bandwidth: B200 to B300 proved the inverse, 192 GB to 288 GB on the same eight stacks at 8 TB/s. Packaging allows far more than the market can feed: 12 stacks × 48 GB = 576 GB in 2027, and 20 × 64 GB ≈ 1.3 TB by 2028–29 on TSMC’s roadmap.

Units nearly double from 2026 to 26.5M in 2030; custom ASICs are about 60% of them from 2027

Demand: units, memory per unit, and each generation’s appetite

Accelerator shipments roughly double from about 14M packages in 2026 to 26.5M in 2030 in our base case, and custom ASICs (TPU, Trainium, MTIA, OpenAI’s chip) make up about 60% of units from 2027. At announced specs, the HBM inside each year’s shipments would grow sixfold, from 2.9 EB in 2026 to about 17.5 EB in 2030; at the specs now shipping or reported, about fourfold, to 11.1 EB.

Units. Our 2026–27 paths follow vendor-level evidence: Nvidia’s guide for about 70% revenue growth in its fiscal 2028 (February 2027 to January 2028) at a much higher price per package (a Vera Rubin rack costs roughly twice a GB300 rack), Morgan Stanley’s 2027 CoWoS demand (2.7M wafers, Nvidia about 45%), TPU forecasts of 5–8.8M in 2027, and the AMD, Broadcom and Anthropic deal books. After 2027 units grow 12% in 2028 and about 8% a year after. A capex- and power-constrained top-down view gives far fewer, about 15–17M packages a year through 2030; it is our low case. The compute-balance pack shows why it matters: see “How this squares with the compute-balance pack” below.

Nvidia generations: memory per package, per rack, and over each generation’s life

• Hopper H100 / H200 — Volume years: 2023–25 · HBM per package: 80 / 141 GB · Rack fast memory (HBM + CPU LPDDR): 0.64–1.1 TB per 8-GPU HGX · Lifetime HBM, our estimate: ~0.5 EB (2024–26)
• Blackwell B200 / GB200 — Volume years: 2025 · HBM per package: 186–192 GB · Rack fast memory (HBM + CPU LPDDR): 30.4 TB per NVL72 (13.4 + 17) · Lifetime HBM, our estimate: ~0.6 EB
• Blackwell Ultra B300 / GB300 — Volume years: 2H25–27 · HBM per package: 288 GB · Rack fast memory (HBM + CPU LPDDR): 37 TB per NVL72 (20 + 17) · Lifetime HBM, our estimate: ~1.8 EB
• Rubin VR200 — Volume years: 2H26–28 · HBM per package: 288 GB · Rack fast memory (HBM + CPU LPDDR): 48–75 TB per NVL72 (20.7 + 28–54) · Lifetime HBM, our estimate: ~2.6 EB of HBM plus 3.4–6.8 EB of LPDDR
• Rubin Ultra — Volume years: 2H27–29 · HBM per package: 1 TB on 4 dies (announced); ~192–256 GB on 2 dies (reported) · Rack fast memory (HBM + CPU LPDDR): 147 TB of HBM per Kyber rack (announced) · Lifetime HBM, our estimate: ~2.9 EB at announced spec vs ~1.1–1.5 EB reported
• Feynman — Volume years: 2028– · HBM per package: “Custom HBM”, capacity undisclosed

Lifetime totals use our unit mix (Hopper 5.2M, Blackwell 3.3M, Blackwell Ultra 6.4M, Rubin 8.9M and Rubin Ultra 5.7M packages) and are rounded. Rubin Ultra counts as two-die equivalents.

The Rubin row is the one to notice. Its LPDDR5X attach, even at the reported halved level, matches or exceeds its HBM in bytes, which is why the Vera cut matters for commodity DRAM.

Everyone else’s current parts

• AMD — Now shipping: MI355X, 288 GB; MI455X, 432 GB HBM4 (since late Q3 2026) · Next: MI500 (2027); memory not yet disclosed · A 288 GB MI450 reportedly built for Meta
• Google — Now shipping: TPU v7 Ironwood, 192 GB; TPU 8i, 288 GB HBM3E (since Sep 2026) · Next: TPU 8t, 216 GB HBM3E · Anthropic: 1 GW in 2026, 5 GW of 8i in 2027
• AWS — Now shipping: Trainium 3, 144 GB · Next: Trainium 4, ~288 GB (HBM4-class) · Anthropic up to 5 GW; OpenAI 2 GW from 2027
• Meta — Now shipping: MTIA 300, 216 GB · Next: MTIA 400/450, 288 GB; MTIA 500, 384–512 GB · The largest among announced 2027 custom ASICs
• Microsoft — Now shipping: Maia 200, 216 GB · Next: Maia 300 in design
• Huawei — Now shipping: Ascend 950PR / 950DT, 128 / 144 GB · Next: Ascend 960DT, 288 GB (Q1 2027) · Domestic HBM; see China below

By HBM bits, the 2027 split is about 37% Nvidia, 31% Google, 13% AWS, 9% AMD and 9% other ASICs in our base case, close to Morgan Stanley’s 37/36/12% for Nvidia, Google and AMD.

After M15X and P4, new memory fabs add wafers from mid-2027; Terafab’s DRAM points to about 2030

Supply: where the HBM bits come from

HBM output roughly doubles in two years, from about 3.9 EB in 2026 to 6.0 EB in 2027 and 8.0 EB in 2028 in our base case, but only by taking about 30% of the big three’s DRAM wafers. Apart from SK hynix’s M15X and Samsung’s P4 phase 2, both producing since the first half of 2026, no new fab adds volume before mid-2027.

HBM eats wafers more than twice as fast as it makes bits. HBM takes 18%, 22% (vendor capacity figures imply about 25%) and 30% of the big three’s DRAM wafer starts in 2025, 2026 and 2027, but yields only 8%, 9% and 13% of DRAM bits (TrendForce). Micron puts HBM3E at about 3× the wafer area per bit of DDR5 and says the ratio rises each generation. We estimate 650–850 good GB per HBM wafer start in 2026, about twenty 36 GB stacks. The consequence: non-HBM wafer starts at Samsung, SK hynix and Micron stay roughly flat at 1.26–1.33M a month through 2027. Nearly every new wafer goes to HBM.

Capacity by vendor, thousand HBM wafer starts a month

Q2 2026 HBM revenue share: SK hynix 50%, Samsung 33%, Micron 18% (Counterpoint).

From wafers to bits. Our base case runs average HBM wafer input from about 450k a month in 2026 to 590k in 2027, 710k in 2028 and 975k in 2030, with good output per wafer rising from 0.72 TB to 1.05 TB as the industry moves to 1c/1d nodes and 32 Gb dies. That lands close to the published anchors, 1.5 EB in 2024 and about 3.0 EB in 2025 (TrendForce), about 4 EB in 2026, and TrendForce’s +50–60% for 2027: we get 1.5, 2.8 and 3.9 EB and +54%.

Beyond 2027 no vendor publishes HBM capacity. Our 2028–30 path assumes HBM keeps about a third of a big-three DRAM base growing toward about 3M wafer starts a month by 2030 (Citrini puts Samsung and SK hynix alone at 2.3–2.6M). The low and high cases sit about 19% below and 23% above the base in 2028, and 28% below and 33% above in 2030.

Price is what forces the shift. 2026 was an HBM pricing pause: HBM3E sold near $1.50/Gb (about $12/GB) while conventional contract prices compounded about sixfold from 3Q25 to 4Q26, to roughly $2/Gb. By Q1 2026 a wafer of 64 GB server DIMMs earned more than a wafer of HBM. At a 3× wafer ratio, a $2/Gb commodity bit implies about $6/Gb for an HBM bit to break even per wafer. That is why 2027 HBM4 quotes sit at $4–5/Gb (about $32–40/GB, or $1,150–1,450 per 36 GB stack) and blended 2027 forecasts run from +54% (J.P. Morgan) to +121% (TrendForce). The HBM price is now hostage to the commodity price.

Most of it is already sold. Micron has $32B of customer commitments and has contracted the “vast majority” of its 2027 HBM. Samsung has signed five-year agreements with its top five data-center customers and plans to cover 60–70% of capacity; SK hynix has long-term deals with about ten customers. A buyer without a contract in 2027 has very little to buy.

At 128K tokens, 2026’s flagship models need 5–50× less KV cache than Llama 3.1 405B

KV cache: the demand driver and the levers

KV cache is what now fills HBM, and it is the most movable part of memory demand. In a typical 2026 serving setup KV takes about 60% of a 288 GB GPU and weights about 15% (LMSYS). Attention design changes KV per token by more than 50× and precision by up to 3.5×, and warm context can be parked in LPDDR, CXL memory or flash. Net: KV growth raises total memory demand but pushes the marginal byte off HBM.

Sizing. KV per token is 2 × layers × KV heads × head size × bytes for standard attention; latent, compressed and linear-attention designs cut it further. At FP8, a 128K-token context ranges from 34 GB on Llama 3.1 405B to 0.7 GB on DeepSeek-V4-Pro, a model four times larger. FP4 KV fits 1.78× more than FP8 but is still experimental, costing 1.6 points on SWE-bench Verified in one test.

Why it keeps growing. Google processed more than 3.2 quadrillion tokens a month in May 2026, against 480 trillion a year earlier. Agents make prompts long and repetitive: 94–98% of coding-agent prompt tokens are reused prefixes, and inputs outnumber outputs 131 to 1. DeepSeek-V4-Pro agent inputs run at a median of 88K tokens and 675K at the 99th percentile; Nvidia sizes Rubin for 128K–400K cached tokens per user.

Where it can live. The tier ladder runs from HBM (the active batch) to CPU LPDDR, optical and CXL pools, flash and network storage (table below). Tiering is now standard in vLLM, SGLang, TensorRT-LLM and Nvidia’s Dynamo. On coding-agent traces, a shared KV store lifted throughput 3.8× and cut time-to-first-token 46× by raising cache hits from 2% to 92%.

• CPU LPDDR. Vera carries 768 GB–1.5 TB per two GPUs over a 1.8 TB/s link. This is Nvidia’s main answer, and why the reported LPDDR cut matters.
• CXL pools. Samsung’s CMM-D pool ran KV offload at about 92% of local-DRAM throughput at TP=8; Alibaba’s CXL pool cut time-to-first-token about 90% against an RDMA store.
• Flash. Nvidia’s BlueField-4 context-memory tier is due in 2H 2026 with “petabytes” per pod and a claimed 5× tokens/s, still unmeasured in public. Citi estimates about 16 TB of NAND per Rubin GPU, about 115 EB in 2027 or 9% of global NAND.
• High Bandwidth Flash. A 512 GB-per-stack design, samples in 2027 and volume in 2028 per analysts, for read-mostly weights, not active KV.

Shrinking it. DeepSeek-V4 stores 8.7× less KV than V3.2 at 1M tokens, Kimi K3 and Qwen3.8 cut per-token KV 2–3× against their predecessors, and Nvidia credits part of GB300’s 1.6× software gain in MLPerf to lower KV precision. Hardware moves the same way: Groq 3 LPX racks run MoE decode from SRAM (500 MB per chip, 128 GB per rack, no HBM) while Rubin GPUs keep attention and KV. The GDDR7-based Rubin CPX no longer appears on Nvidia’s Vera Rubin page.

What it does to the balance. HBM capacity sets concurrency: GB300’s 1.5× HBM over GB200 gave 1.6× the decode batch, and a third less HBM removes 36–41% of resident sequences because weights are a fixed charge. Whether that costs tokens depends on reads. With dense attention, per-user decode speeds up and tokens per GPU fall only 5–11%; with sparse attention, or agent sessions parked between tool calls, the loss approaches the full 40%. Hence the design now shipping: size HBM for the active batch and stage warm context one tier down during tool-call pauses, which last about 5 seconds at the median.

Which tier depends on attention design. At 57 agents per GPU (Artificial Analysis’s measured rate for DeepSeek-V4-Pro on GB300) and 90th-percentile agentic context, an MLA model spills about 25 TB of warm KV per Rubin rack, which fits Vera’s LPDDR. A dense 405B-class model spills 80–280 TB, which needs flash, while DeepSeek-V4-class models fit in HBM (our estimates). So KV growth lands mostly in LPDDR, tightening commodity DRAM, and in NAND: enterprise SSD is now about 37% of NAND demand (CFM), and Citi sees a 6.1% NAND shortfall in 2027.

Our trimmed-spec case assumes that tiering works. If agent hosting is priced per session and insists on HBM-resident context, buyers will pay for taller stacks again, and demand moves toward the announced-roadmap case.

Marvell’s Photonic Fabric and pooled memory

Optical pooling slows the growth of HBM per GPU from 2028; it does not cut it before then. On an accelerator designed around a pool, it could replace 10–20% of HBM capacity in 2028 and 25–35% by 2030. Across the whole fleet that is about 2% in 2028 and 3–8% in 2030 (our estimates).

What Marvell bought. Marvell closed its $3.25B purchase of Celestial AI on February 2, with up to $2.25B more tied to revenue. Photonic Fabric links chips over light at about 2.4–3.1 pJ/bit and adds roughly 200–300 ns of latency; on-package HBM is under 100 ns. The memory module fronts 2 TB of DDR5 with two HBM3E stacks (72 GB) acting as a cache. A 2U appliance of 8–16 modules holds about 33 TB, shared by up to 16 servers as far as 50 m away.

Where it stands. The memory module taped out in 2026, is in bring-up, and samples to “multiple customers” in 2027. Every disclosed win so far is optical interconnect, not pooled memory: a Tier 1 hyperscaler’s XPU scale-up network and co-packaged optics added to the AWS agreement. Marvell’s revenue targets, a $500M run-rate by November 2027–January 2028 and $1B by November 2028–January 2029, are interconnect too. Amazon holds a warrant for 1,045,171 Marvell shares at $87.00 that vests with its Photonic Fabric purchases through 2030, the clearest signal of who adopts first.

It replaces capacity, not bandwidth. The appliance switch moves about 115 Tbps (about 14 TB/s), roughly one Rubin GPU’s HBM4 bandwidth. Shared across 64–128 accelerators, that is 0.1–0.2 TB/s each, about 1% of Rubin’s 19–22 TB/s, while offering 250–500 GB of capacity each. So the pool can hold only bytes touched rarely: warm or parked KV cache, cold MoE experts and adapters, optimizer state, checkpoints. KV for sequences in the active decode batch, and hot weights, are read every token step and stay in HBM.

The lever it unlocks is stack height. HBM capacity comes bundled with bandwidth: drop a stack and you lose both. Shortening stacks keeps bandwidth (12-Hi to 8-Hi cuts GB by a third, 16-Hi to 12-Hi by a quarter), and a pooled tier is what lets a designer take that cut. That bounds displacement near 25–35% per accelerator.

• On-package HBM — Example: Rubin, 288 GB of HBM4 · Added latency: under 100 ns · Bandwidth per accelerator: 19.2–22 TB/s · Status, Oct 2026: shipping 2026
• CPU-attached LPDDR — Example: Vera, 768 GB–1.5 TB of LPDDR5X per 2 GPUs · Added latency: local DRAM · Bandwidth per accelerator: ~0.6 TB/s (1.2 TB/s per Vera) · Status, Oct 2026: ships with Vera Rubin
• Optical memory pool — Example: Marvell Photonic Fabric appliance · Added latency: ~200–300 ns · Bandwidth per accelerator: 0.1–0.2 TB/s shared; 1.8 TB/s per dedicated chiplet · Status, Oct 2026: samples 2027
• CXL switch pool — Example: Marvell Structera S 30260 (48 TB), Astera Leo, Samsung CMM-D · Added latency: ~460–750 ns · Bandwidth per accelerator: ~26–36 GB/s per device path · Status, Oct 2026: expanders shipping; switches sampling
• RDMA / flash context tier — Example: Nvidia ICMS (BlueField-4), Enfabrica EMFASYS · Added latency: microseconds · Bandwidth per accelerator: ~0.1 TB/s (800 Gb/s) · Status, Oct 2026: ICMS due 2H 2026

Who adopts, and why the effect is muted. Nvidia, the largest HBM buyer, is building its own non-optical tiers (Vera LPDDR5X, ICMS, the Enfabrica team it paid over $900M for) while its announced Rubin Ultra still lists 1 TB of HBM per package (reportedly cut to 192–256 GB). That leaves custom accelerators (AWS, Google, Meta, Microsoft) and UALink systems as the early market; optical NVLink for Feynman in 2028 is reported but unconfirmed. Vendors sell pooling as a 2–5× throughput gain, mostly from simulation, not as HBM savings, and cheaper tokens can raise total HBM demand, so the industry-level effect is smaller than the per-accelerator one and could be zero. The tell that this is changing would be a named memory-appliance customer or purchase order; none has been disclosed.

China: CXMT, YMTC, Huawei and export controls

China narrows the global commodity DRAM shortfall but barely touches the HBM balance. Hanwha estimates CXMT shrinks the 2026 global DRAM deficit from 15.5% to 9.4%, and 2027’s from 12.0% to 5.6%. Almost none of that reaches HBM, and most of it stays inside China.

CXMT is big in wafers, smaller in bits, and not cheap. It runs about 300–320k wafer starts a month across two Hefei fabs and one in Beijing, heading to about 420k in 2027 and 500–600k in 2028. That is about 17% of the top five makers’ wafer starts but only 6.7–9% of bits, because its nodes lag; it took 9.5% of DRAM revenue in Q2 2026, fourth globally. It prices at parity with Samsung, earned a net margin of about 52% in the first half of 2026 (RMB77.6bn on RMB150.3bn of revenue), and is booked through 2027. Its July STAR Market IPO raised RMB57.9bn ($8.6B). So it relieves volume, not price.

It mostly displaces imports. Goldman puts CXMT at 41% of China’s DRAM demand in 2026 and about 50% by 2028. The bits that frees for Korean and US vendors elsewhere are small next to industry bit growth of 22–24% a year, and Korean HBM conversion is driven by HBM demand regardless.

China’s HBM is a closed, starved loop. CXMT makes HBM3 8-high at about 25% overall yield (SemiAnalysis, June) and began small-scale HBM3E in September; it has no HBM4. Huawei’s pre-control stockpile of about 13M stacks (11.4M from Samsung) runs out in 2026. Epoch AI’s model has about 1.5M Ascend chips in 2026, but only about 240k with domestic HBM, about 30–35M GB or 0.03 EB, while about 750k use non-HBM memory. Ascend prices rose 20–50% in two months; the 950DT now lists above RMB250k. DeepSeek’s order for 160k 950DTs alone would take about 70% of 2026’s domestic HBM.

Export controls cap China’s demand on the Korean and US balance. The December 2024 rule bars HBM above 2 GB/s/mm² of bandwidth density; equipment controls cover every CXMT node. Nvidia’s H200 is licensed, with 25% to the US Treasury and at most 75k units per buyer, but Beijing’s own gatekeeping held deliveries to about 20k by August against more than 2M ordered. Even the full licensed ceiling, about 750k H200s at 141 GB, is about 0.1 EB. China’s draw on Korean and US HBM is policy-gated, small, and could swing either way.

The Korean fabs in China are frozen at their nodes. Samsung’s and SK hynix’s China fabs lost fast-track (validated end-user) status on December 31, 2025 and run on annual licences that bar upgrades. Wuxi is about half of SK hynix’s DRAM capacity, so its HBM growth has to come from Korea and the US while Wuxi drifts toward commodity output that competes with CXMT.

YMTC matters more for flash than DRAM. It is now third in NAND by shipments, with 14–16.5%. Omdia has it adding about 486k NAND wafers in 2027 against about 75k each for Samsung and SK hynix, useful for the SSD tier that holds offloaded KV cache. Its DRAM entry, about 50k wafers, is immaterial before 2028.

Oversupply risk: low through 2027, rising after. China’s DRAM capacity could reach about 1.4M wafer starts a month by 2030, and domestic DUV scanners arrive in 2027–28 (Citrini). An AI air pocket would hit DDR5 and LPDDR5X first, CXMT’s core products. Citrini still models a global DRAM deficit of about 25% in 2030.

Dates to watch. The US affiliates rule returns on November 10 unless extended; the HBM temporary licence expires on December 31; the Korean fabs in China need 2027 licences; and CXMT’s Entity List designation is approved but unpublished.

Terafab and other new capacity

Terafab adds essentially no memory before 2030; through this forecast it matters as a buyer, not a supplier. Our base case puts its DRAM output at about 10,000 wafer starts a month in 2030, roughly 0.1% of global DRAM bits, and no HBM.

What is actually committed. Musk floated a “Tesla TeraFab” in November 2025, described it in January 2026 as “logic, memory and packaging”, and launched it on March 21 with talk of 100,000 wafer starts a month initially and 1,000,000 eventually. Intel joined in April, and the first full-scale phase is to run Intel’s 14A logic process. Two things are firm. A research fab at Giga Texas of “a few thousand” wafers a month, about $3B, starts operating in the first half of 2027; its first test chip is SRAM, due in May–June 2027. And a $16.8B Phase 1 at Grimes County runs from December 1, 2026 to the end of 2028, within up to $119B of phases through 2036. On October 20 Musk targeted volume logic production for Q3–Q4 2028, “with big error bars”. No DRAM capacity, node, process licensor or memory partner has been disclosed; the memory evidence is job postings since August for sub-20nm DRAM process integration and HBM/LPDDR interface design.

Why memory comes late. Incumbents with their own process take about 3.7 years from groundbreaking to first DRAM wafers: Micron’s Idaho ID1 broke ground in October 2023 and outputs first wafers in mid-2027, with meaningful output in 2028. CXMT, the last new entrant, took about 3.5 years to first low-volume output and 8–10 years to a mid-single-digit share. Phase 1’s $16.8B buys roughly one memory fab ($15–20B) or half a leading-edge logic fab, and it is going to logic. Without a licensed DRAM process, first product wafers point to about 2030; HBM, which needs best-in-class DRAM yields plus stacking, comes after that.

Our scenarios. Base (35%): about 2,000 DRAM wafer starts a month in 2029 and about 10,000 in 2030, about 0.1% of 2030 DRAM bits. High (10%), with a licensor signed and tools ordered by mid-2027: about 20,000 and 50,000, about 1% of bits and under 1% of HBM. Low (55%): engineering lots only. For scale, CXMT alone adds about 100,000 wafer starts a month in 2027, twice the high case for 2030.

Net effect: neutral to tighter. Whatever Terafab makes is captive, so it displaces Tesla and SpaceX purchases rather than adding merchant supply. Meanwhile Musk’s companies plan about 13 GW of compute by end-2027, including a 440,000-GPU Vera Rubin cluster by mid-2027, roughly 0.13 EB of HBM4 on its own. Musk himself calls memory “the biggest constraint”. The one real effect before 2030 is on bargaining: a credible in-house threat can weigh on 2028–30 contract pricing before any bits exist.

Other new capacity is incumbents, financed by customers. Outside China there is no new DRAM producer before 2030. SoftBank and Intel’s Saimemory plans Z-Angle Memory prototypes by March 2028 and commercialization in its fiscal 2029, with no capacity disclosed. Buyer money instead funds incumbents’ cleanrooms. Micron has 26 strategic customer agreements covering more than 35% of revenue through 2030, and $32B of customer commitments, mostly prepayments. SK hynix is moving to 3–5-year contracts with 10–30% upfront.

Unit volumes, supply and memory per package decide the cycle

Supply vs demand (supply ÷ demand − 1) in each scenario. “Low” units are the capex- and power-limited view, about 15–17M packages a year; “high” takes the vendors’ deal books at face value, about 25M in 2027 and 39M by 2030.

• Base — Units: base · Specs: trimmed · HBM supply: base · 2027: −2% · 2028: +3% · 2030: −2%
• Roadmaps hold — Units: base · Specs: announced · HBM supply: base · 2027: −18% · 2028: −29% · 2030: −38%
• Deal books delivered — Units: high · Specs: trimmed · HBM supply: base · 2027: −22% · 2028: −22% · 2030: −33%
• Supply disappoints (HBM4 yields, cleanroom slips) — Units: base · Specs: trimmed · HBM supply: low · 2027: −16% · 2028: −17% · 2030: −29%
• Supply surprises (Samsung, new fabs on time) — Units: base · Specs: trimmed · HBM supply: high · 2027: +12% · 2028: +27% · 2030: +30%
• Power- and capex-limited build — Units: low · Specs: trimmed · HBM supply: base · 2027: +37% · 2028: +52% · 2030: +52%
• Power-limited build, roadmaps hold — Units: low · Specs: announced · HBM supply: base · 2027: +15% · 2028: +4% · 2030: −6%

One input at a time, 2028

2028 supply vs demand at trimmed specs when one input changes and the rest stay at base. Labels: the input’s setting and the result. Whole unit cases swing it further, −22% to +52%.

Reading the scenarios

The base case sits on a knife-edge. Moving to the low or high supply case (about −19% or +23%), or ±15% on memory per package, swings 2028 anywhere from 10–17% short to 21–27% long. Three variables decide the cycle: total unit volumes, which power limits; good bits per HBM wafer (HBM4/4E yields, Samsung’s ramp); and whether 8-Hi stacks stick.

Reading the camps. The “oversupply by 2028” view (Bloomberg Intelligence, DS Investment, Morgan Stanley’s Micron model with fiscal 2029 EPS down about 40%) needs either the power-limited build or the supply upside. The “short through 2030” view (SK hynix, Micron, Citrini’s 25% DRAM deficit in 2030) needs the announced roadmaps or the deal books. Our base is neither: within ±3% every year, and only because memory per package stops compounding at its old rate.

How this squares with the compute-balance pack. The compute-balance pack found that from 2027 power and money, not chips, limit what gets switched on: about 38 GW of all-in capacity in 2027, rising to 51 GW in 2030, while the chipmakers could deliver chips for 47–80 GW. On the compute-balance pack’s basis (a data center draws about twice its chips’ rated power) and our estimate of 1.4–1.7 kW of chip power per package, our base units need about 55 GW in 2027 and 68–93 GW a year in 2028–30, 7–17% more than the compute-balance pack’s chip path. Its switched-on path sits below even our low unit case (41–61 GW). In that low case HBM runs 37–52% long at trimmed specs in 2027–30, and the industry could afford the announced roadmaps instead: supply would sit within −6% to +15% of demand. If chip purchases fall to the compute-balance pack’s switched-on path, the surplus is larger still.

Asymmetry. Our base needs more new AI capacity a year after 2027 than the compute-balance pack expects to be switched on, so the unit risk leans down. The downside case for memory makers, a power-limited build with trimmed specs (+52% in 2028), arrives just as Yongin, P5, ID2 and Hiroshima add wafers. The upside needs the announced roadmaps or deal-book volumes, either of which leaves 2028 22–29% short.

What to watch

The next six months settle more of this than the last six: 2027 HBM4 contract prices, whether 8-Hi sticks on Rubin Ultra and the 2027 ASICs, and how many gigawatts actually get switched on. 2027 memory earnings are largely contracted; the open question is 2028.

• Oct 29 — SK hynix and Samsung Q3 results · 2027 capex, HBM wafer adds, HBM4 share at Nvidia, long-term-agreement coverage
• November — Nvidia fiscal Q3 2027 results · Rubin ramp, gross margin path, supply commitments; any confirmation of Rubin Ultra’s 192/256 GB configuration or Vera’s 768 GB SKU
• Nov 10 and Dec 31 — US affiliates rule returns; HBM temporary licence expires · China’s access to HBM and tools; 2027 licences for Korean fabs in China
• 4Q 2026 — 2027 HBM4 contract settlements · HBM4 near $4–5/Gb validates the top of the +54–121% range; HBM4 well below $4/Gb, or blended pricing nearer +54%, means buyers had more leverage
• 1Q 2027 — First 2027 conventional DRAM contract print · A rise under 5% a quarter favours the balance-by-late-2027 camp
• March 2027 — Nvidia GTC · Feynman’s memory spec and Rubin Ultra’s final configuration
• Mid-2027 — M15X at full rate; Yongin fab 1, Micron ID1 and Tongluo first output · Whether 2028 supply arrives on time
• 2027 — Photonic Fabric memory samples; HBF samples; first named pooling customer · Whether capacity moves off HBM after 2028
• Rolling — Gigawatts switched on against SemiAnalysis’s ~51 GW for 2027; hyperscalers’ 2027 capex guides · Which unit path, base, power-limited or deal-book, is real

HBM makers (SK hynix, Samsung, Micron). Pricing power peaks in 2027, when HBM must reprice toward commodity-DRAM parity at about 3× the wafer cost per bit. Micron’s $32B of customer commitments and Samsung’s five-year agreements, planned to cover 60–70% of capacity, make 2027 visible; the equities already discount a 2028 turn (SK hynix about 57% off its June peak, at about 2.7× 2027 earnings per BofA). Our base case says the 2028 risk comes from the demand side, trimmed specs plus a power-limited build, more than from new fabs.

Nvidia. Memory is the main gross-margin headwind (75% to 71–72%) and the reason for server price increases of more than 15%. Trimming memory per GPU protects unit volume; per-die HBM falls on Rubin Ultra.

Custom-ASIC chains (Broadcom, MediaTek, Alchip, Marvell). They take most of 2027’s unit growth and are still mostly on HBM3E, which keeps HBM3E tight through 2027.

Capacity tiers. LPDDR/SOCAMM suppliers, enterprise SSD and NAND, and CXL silicon (Astera Labs, Marvell, Montage) are where KV-cache growth shows up first. Optical pooling (Marvell) and High Bandwidth Flash (SanDisk, SK hynix) come in 2028 and later.

China and Terafab. CXMT relieves volume, not price, and stays mostly domestic; China’s demand for Korean and US HBM is set by policy, not need. Terafab is a buyer through 2030, with bargaining value before it has any bits.

What this is anchored on

Supply, demand, memory per package, the balance and the scenarios are our estimates, built from the sources below as of October 25, 2026.

• Nvidia — Rubin: 288 GB of HBM4 per GPU, listed bandwidth 19.2 TB/s (22 TB/s at CES); Vera “up to 1.5 TB” of LPDDR5X; sized for 128K–400K cached tokens per user · Nvidia Vera Rubin NVL72 page
• Nvidia — GB300 NVL72: 20 TB of HBM (about 278 GB per GPU) and 17 TB of LPDDR5X per rack · Nvidia GB300 NVL72 page
• Nvidia — Rubin Ultra announced with 1 TB of HBM4E on four compute dies; Kyber NVL144 · TechNews, GTC 2026, Mar 18
• Nvidia — Q2 gross margin 75.0%; Q3 guide 74% ±50 bp; inventory $31.6B · Nvidia Q2 FY27 8-K
• Nvidia — CFO Colette Kress on “extreme pricing conditions in memory”: gross margin troughs at 71–72% in the quarter to January 2027; supply commitments from $119B to $279B · Nasdaq / Zacks, Sep 3
• Nvidia — Supply and capacity purchase commitments, mostly memory, from $119B to $279B in one quarter; Morgan Stanley’s 2027 HBM split of 37/36/12% for Nvidia, Google and AMD · XenoSpectrum
• Rubin Ultra — Moves to 8-Hi stacks: not enough HBM wafers for 12-Hi, HBM cost, and a 576-GPU domain of 192 GB parts (~110 TB) · SemiAnalysis, Sep 13
• Rubin Ultra — Configurations tested with as little as 192 GB, and a step back to HBM4 · The Information, via Tom’s Hardware
• Rubin — Most Rubin stays on 12-Hi HBM4 · Counterpoint, via Sammy Fans, Oct 8
• Vera — LPDDR5X per Vera CPU cut from 1.5 TB to 768 GB in 2027 supply plans, “lowering memory per module while increasing total module shipments” · TrendForce, Jun 10
• Vera Rubin cost — Without the Vera cut, memory would be ~29% of a ~$2.1M VR200 system’s bill of materials, against Nvidia’s preferred ~20% · GF Securities, via TrendForce, Jul 28
• Vera Rubin cost — Memory up from ~5–10% to 25–30% of a rack’s bill of materials; server prices up more than 15% · Morgan Stanley and Bloomberg, via VKTR
• GeForce — RTX 5070 Ti and 5060 Ti 16 GB supply cut 30–40% in 1H26 against 1H25 (Board Channels, Benchlife) · TweakTown, Dec 17, 2025
• GeForce — Nvidia: “memory supply is constrained”; all GeForce SKUs still shipping · Tom’s Guide, Jan 22
• HBM per device — GPU and ASIC makers “considering reducing HBM capacity per device”, evaluating 8-Hi in 2027; 8-Hi at a 10–20% per-Gb premium; blended 2027 HBM price +121% · TrendForce, Sep 29
• HBM price — HBM4 12-Hi about $2/Gb in 2026 and $4–5/Gb in 2027; 8-Hi at a 20–30% premium · Citi, via StockHub, Oct 8
• HBM price — HBM3E near $1.50/Gb under 2026 contracts · Korea Herald
• HBM balance — HBM sufficiency −16% across 2026–28; HBM price +54% in 2027 · J.P. Morgan, via Zhitong / Webull, Sep
• HBM balance — HBM sufficiency −5.4%, −6.0% and −4.3% in 2026–28; DRAM −5.0%, −5.9%, −3.9% · Goldman Sachs, via Tiger, Jun
• HBM balance — SemiAnalysis’s HBM balance; Samsung P5 HBM output in 2028 · Luminix
• DRAM balance — DRAM sufficiency −5.0%, −1.9% and +0.5% in 2026–28 · DS Investment & Securities, Jul 20 (PDF)
• HBM supply — HBM 18%, 22% and 30% of the big three’s DRAM wafer starts in 2025–27, and 8%, 9% and 13% of bits · TrendForce, via I-Connect007
• HBM supply — HBM3E takes about 3× DDR5’s wafer area per bit, and the ratio rises each generation · Micron, via XenoSpectrum
• HBM supply — Samsung HBM wafer input ~180k a month on average in 2026 and ~250k in 2027 · TrendForce, Sep 29
• HBM supply — Micron adds 60k HBM wafers a month, reaching ~100k by year-end · TrendForce, Sep 4
• HBM supply — DRAM and TSV (HBM) wafer capacity by vendor, 2025–26 · TrendForce, FMS 2025
• HBM supply — HBM bit shipments: 1.5 EB in 2024, about 3.0 EB in 2025 · TrendForce, FMS 2025
• HBM supply — Q2 2026 HBM revenue share: SK hynix 50%, Samsung 33%, Micron 18% · Counterpoint, via MT, Oct 19
• DRAM capacity — 2030 capacity: Samsung 1.14–1.45M and SK hynix 1.18M wafer starts a month; China ~1.4M; domestic DUV in 2027–28 · Citrini, via Tom’s Hardware, Jul 15
• DRAM price — Conventional contract prices up about sixfold from 3Q25 to 4Q26; 4Q26 +10–15% · TrendForce, Sep 30
• Contracts — Micron: 26 strategic customer agreements covering more than 35% of revenue through 2030, $32B of customer commitments; ID1 and Tongluo output mid-2027, ID2 and Japan late 2028, New York 2030 · Micron Q4 FY26 call, via Benzinga, Sep 30
• Contracts — SK hynix moves to 3–5-year agreements with about 10–30% upfront; Samsung requires agreements of at least three years · TrendForce, Apr 9
• New fabs — SK hynix: M15X mass production accelerated; Yongin phase-1 cleanroom opens early 2027 · Edaily, Jul 29
• New fabs — SK hynix Yongin fab 2: groundbreaking Jul 2027, first cleanroom Jun 2029 · The Motley Fool, Aug 15
• New fabs — SK hynix’s Yongin fab 1 at ~150k wafer starts a month by end-2028 · CSIS
• New fabs — Samsung P4 adds about 60k wafer starts a month through Q2 2026; P5 operating in 2028 · KB Securities, via JoongAng Daily, Jan 4
• Units — Anthropic: 1 GW of Ironwood TPUs in 2026 and 5 GW of TPU 8i in 2027; TPU 8i production shipments started · Broadcom fiscal Q3 call, via Webull, Sep 3
• Units — OpenAI: 2 GW of Trainium from 2027 · 247 Wall St., Jul 21
• Units — 2027 CoWoS demand of about 2.7M wafers, Nvidia ~45% · Morgan Stanley, via KuCoin
• Power — New data-center IT capacity outside China ~25 GW in 2026 and ~51 GW in 2027 · SemiAnalysis, via Superpower Daily
• Power — From 2027 power and money cap what is switched on: about 38 GW all-in in 2027 rising to 51 GW in 2030 · 06.ai compute-balance pack
• Other chips — MTIA 300, 216 GB; MTIA 400/450, 288 GB; MTIA 500, 384–512 GB · Tom’s Hardware
• Other chips — TPU 8t, 216 GB of HBM3E; TPU 8i, 288 GB · StorageReview
• Other chips — MI455X with 432 GB of HBM4 in the Helios rack · AMD ROCm blog
• Other chips — Maia 200, 216 GB · HotHardware
• Other chips — Trainium 3, 144 GB of HBM3E · Electronic Specifier
• Other chips — Trainium 4 doubles Trainium 3’s memory and quadruples its bandwidth · The Next Platform
• Workloads — MLPerf Inference v6.1: GB300 and GB200 gain the same ~1.29× on models that can and can’t be capacity-bound · Nvidia MLPerf results
• Packaging — TSMC: 9.5-reticle CoWoS in volume in 2027 with 12 or more HBM stacks · TrendForce, Apr 24, 2025
• Packaging — TSMC: CoWoS with 20 HBM stacks in 2028 and up to 24 in 2029 · TrendForce, May 14
• KV cache — GB300 serving DeepSeek-R1: KV pool ~61% of HBM, weights ~14%; 1.5× the HBM gave 1.6× the decode batch · LMSYS, Feb 19
• KV cache — FP4 KV holds 1.78× more than FP8; −1.6 points on SWE-bench Verified in one test · LMSYS, Sep 16
• KV cache — DeepSeek-V4 stores 8.7× less KV than V3.2 at 1M tokens · vLLM, Apr 24
• KV cache — Coding-agent traces: 94% prefix-cache hits, 131:1 input to output, 5.2 s median pause; a shared KV store lifted throughput 3.8× and cut time-to-first-token 46× · vLLM × Mooncake, May 6
• KV cache — Agent calls hit 85–97% cache after the first; agent teams 97.2% · Nvidia Dynamo blog
• KV cache — DeepSeek-V4-Pro agent inputs: median 88K tokens, 675K at the 99th percentile · SemiAnalysis AgentX, Aug 24
• KV cache — 57.5 concurrent agents per GPU for DeepSeek-V4-Pro on GB300 (Artificial Analysis AgentPerf) · TensorRT-LLM tech blog 26
• KV cache — Model configs: Llama 3.1 layers and KV heads · Meta llama-models
• KV cache — Kimi K3: ~27 KB of MLA KV per token, ~434 MB of recurrent state per request · LMSYS, Jul 27
• KV cache — Qwen3.8: 23 grouped-query layers with 4 KV heads of 256, plus linear-attention state · SGLang cookbook
• KV cache — GLM-5.2: a measured FP8 KV page of 2.95 MiB per 64 tokens · TensorRT-LLM tech blog 30
• KV cache — Google: more than 3.2 quadrillion tokens a month in May 2026, against 480 trillion a year earlier · Shacknews
• Lower precision — Part of GB300’s 1.6× software gain credited to lower KV-cache precision · Nvidia blog
• Groq 3 LPX — SRAM decode racks: 500 MB per chip, 128 GB per rack, paired with Rubin GPUs · Nvidia technical blog
• Tiers — KV offload to a CXL pool at ~92% of local-DRAM throughput · Samsung CMM-D white paper
• Tiers — Alibaba’s CXL pool cut time-to-first-token ~90% against an RDMA store · arXiv
• Tiers — BlueField-4 context-memory storage, 2H 2026, “up to 5×” tokens/s · NAND Research
• Tiers — ~16 TB of NAND per Rubin GPU; ~115 EB in 2027, ~9% of global NAND · Citi, via Wccftech
• Tiers — High Bandwidth Flash: samples 2027, production 2028 · TrendForce, Aug 14
• Tiers — Enterprise SSD 37% of 2026 NAND demand · CFM, via TechNews, Apr 7
• Tiers — NAND shortfall of 6.1% in 2027 · Citi, via Wallstreetcn, Sep 21
• Tiers — Structera: 48 TB behind one CXL switch · Futurum
• Marvell — Celestial AI acquisition closed Feb 2: $3.25B plus up to $2.25B tied to revenue · Marvell
• Marvell — Photonic Fabric memory module: 2 TB of DDR5 behind 72 GB of HBM3E; a 2U appliance of ~33 TB for up to 16 servers · Celestial AI, HOTI 2025
• Marvell — Memory module samples to “multiple customers” in 2027; interconnect revenue targets · Marvell Investor Day transcript
• Marvell — Amazon warrant for 1,045,171 shares at $87.00, vesting with Photonic Fabric purchases through 2030 · Marvell 8-K
• Marvell — Photonic Fabric link energy and latency · Celestial AI, Hot Chips 2025
• China — CXMT cuts the global DRAM deficit from 15.5% to 9.4% in 2026 and from 12.0% to 5.6% in 2027 · Hanwha, via XenoSpectrum
• China — CXMT capacity and output · TrendForce
• China — CXMT to double DRAM output · DigitalToday
• China — CXMT begins HBM3E production · JoongAng Daily
• China — CXMT capacity sold out through 2027 · Guru3D
• China — CXMT 41% of China’s DRAM demand in 2026, ~50% by 2028 (Goldman) · DigitalToday
• China — ~1.5M Ascend chips in 2026, ~240k with domestic HBM · Epoch AI
• China — Ascend prices up 20–50%; 950DT above RMB250k · TrendForce
• China — Huawei’s pre-control stockpile of ~13M HBM stacks · The Decoder
• China — H200 licence terms: 25% to the US Treasury, at most 75k units per buyer · Introl
• China — Chinese firms ordered more than 2M H200s (Reuters) · Fintool, Jan 18
• China — About 20k H200s delivered, to ByteDance and Tencent, by August · Spheron
• China — Samsung and SK hynix China fabs lose validated end-user status on Dec 31, 2025 · Federal Register
• China — YMTC adds ~486k NAND wafers in 2027 (Omdia) · FNNews
• Terafab — Launched Mar 21 with talk of 100,000 wafer starts a month initially, 1,000,000 eventually; Texas filings put Phase 1 construction from Dec 1, 2026 to end-2028 · Wikipedia, citing filings and the Austin American-Statesman
• Terafab — $16.8B first phase at Grimes County, to run Intel’s 14A process · Tom’s Hardware, Aug 7
• Terafab — Volume logic production targeted for Q3–Q4 2028, “with big error bars” · Reuters, via The Star, Oct 21
• Terafab — Hiring a memory design interface architect with HBM and LPDDR experience; research fab to start operating in 1H 2027 · Seoul Economic Daily, Oct 23
• Terafab — Hiring for sub-20nm DRAM process integration · TrendForce, Aug 7
• Terafab — CXMT adds about 100,000 wafer starts a month in 2027 · TrendForce, Sep 24
• Terafab — About 13 GW of compute by end-2027, including a 440,000-GPU Vera Rubin cluster · DiarioBitcoin
• Terafab — Musk calls memory “the biggest constraint” · Seoul Economic Daily, Aug 8
• New entrants — Saimemory: Z-Angle Memory prototypes by Mar 2028, commercialization in FY2029 · SoftBank
• Markets — SK hynix down ~57% from its June peak at ~2.7× 2027 earnings · BofA, via Tiger, Oct 16
• Markets — Micron EPS of $225 in fiscal 2028 and $134 in fiscal 2029 · Morgan Stanley, via Tiger, Oct 2
• Markets — IDC expects 2026 PC shipments to fall 11.3% · IDC, Mar 12

Method

The balance. A calendar-year model of HBM from SK hynix, Samsung and Micron against the HBM inside the accelerators that use it. Supply = average HBM wafer starts a month × 12 × good TB per wafer start. Demand = K × [(1 − L) × C(y) + L × C(y + 1)], where C(y) is the HBM inside the year’s shipments (units × GB per package, summed over families), L = 0.35 is the share of a package’s HBM bought in the calendar year before it ships (stacks move through CoWoS, boards and buffers a quarter or more ahead), and K = 1.06 covers assembly losses and physical capacity above the usable figure. Supply vs demand = supply ÷ demand − 1. Supportable GB per package = supply ÷ (K × [(1 − L) × U(y) + L × U(y + 1)]), with U the packages shipped; the announced and trimmed averages use the same weights. China’s domestic HBM and accelerators sit outside the balance; licensed H200s are inside Nvidia’s units. LPDDR, GDDR and flash effects are discussed, not modelled.

• HBM wafer starts, thousand a month — 2026: 450 (420–480) · 2027: 590 (540–640) · 2028: 710 (620–820) · 2029: 840 (700–1,000) · 2030: 975 (780–1,180)
• Good TB per wafer start — 2026: 0.72 (0.70–0.78) · 2027: 0.85 (0.80–0.90) · 2028: 0.94 (0.87–1.00) · 2029: 1.00 (0.91–1.08) · 2030: 1.05 (0.95–1.15)

Supply inputs. Vendor wafer figures run to 2027 (TrendForce, ETNews); after that HBM keeps about a third of a big-three DRAM base growing toward ~3M wafer starts a month by 2030. Good TB per wafer is calibrated to TrendForce’s bit totals, then rises 5–10% a year with 1c/1d nodes and 32 Gb dies, less than node gains alone because HBM4E and HBM5 dies are larger.

Units. 2026–27 bottom-up from Nvidia’s fiscal 2028 guide, Morgan Stanley’s 2027 CoWoS demand, TPU forecasts and the AMD, Broadcom and Anthropic deal books; then about 12% in 2028 and 8% a year after. Low is the capex- and power-limited top-down view; high takes the deal books at face value. We count packages; Rubin Ultra counts in two-die equivalents so units compare across generations (Nvidia’s GTC 2025 names counted dies as GPUs; its current names, like ours, count packages).

Memory per package. “Announced” uses vendor specs where they exist (Rubin Ultra 1 TB on four dies, MI455X 432 GB, MTIA 500 up to 512 GB) and extends their trend where they don’t, putting Feynman-class parts at about 0.8–0.9 TB per two-die package in 2029–30. “Trimmed” uses shipping or reported configurations: Rubin 288 GB, Rubin Ultra 192–256 GB on two dies, a 288 GB MI450, 8-Hi stacks on the 2027 ASICs, and Feynman near 400–460 GB. Per-family values are in the data under the memory-per-package chart.

KV cache. Bytes per token from each model’s configuration at FP8 (2 × layers × KV heads × head size for standard attention; 576 bytes per layer for latent attention), times 131,072 tokens, plus the fixed recurrent state of linear-attention hybrids. GLM-5.2 uses a measured FP8 page; DeepSeek-V4-Pro is as deployed (FP8 attention, FP4 indexer, vLLM’s formula). The throughput, concurrency and spill figures are our estimates from published serving configurations.

The compute-balance pack cross-check. Chip power per package is our estimate by family and year (Nvidia 1.4 kW in 2026 rising to 2.6 kW by 2030; TPUs, Trainium and other ASICs about 0.8–1.3 kW; AMD 1.4–2.2 kW), times units, times two for the whole facility, as in the compute-balance pack.

Known limits. Supply after 2027 is our estimate: no vendor publishes HBM capacity that far out, and the low and high cases are the honest range. HBM4 and HBM4E bits per wafer are undisclosed; a 10% miss in good TB per wafer moves the balance about 10 points. Units after 2027 are set by power, not chips. 2025 doesn’t reconcile: supply of about 2.8 EB exceeds the 2.0 EB our lead implies was bought; we read the gap as inventory across the chain, and some may be units we undercount. Units and specs don’t respond to price in the model; in practice price clears the market, so a shortfall here is pressure on price and specs, not a shortage that will occur.

Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at 06.ai/trackers.

Not investment advice. Estimates are labeled as estimates.

Before the AI buildout Meta’s revenue grew about 17% a year and Google’s ads about 15%; in 2022–23 both averaged single digits. Meta now grows 26% on more than three times its 2019 revenue. We estimate AI accounts for about 15 of Meta’s 27 points of ad growth this year, about $55B, and about 5 of Google’s 14. Count only those gains, charge every dollar of AI capex, superintelligence labs included, and Meta’s spending through 2026 earns about 35% a year: core ads pay for the moonshot’s capex. The 2027 step-up is the bet. If it earns nothing new, Meta’s return falls to about 18%, or 12% once spending outside capex is counted. Alphabet’s, with only Cloud’s backlog to show for 2027, falls to about 13%. These are the middle of our ranges.

• 15 of 27 pts — AI’s share of Meta’s 2026 ad growth about $55B of revenue, 22% of the year’s ads, against a no-AI trend of 11% a year
• 35% — Meta’s return on AI capex through 2026 a year, from ads alone with every moonshot at zero. Paid back in 2029; the ad gains could carry 2.5× that capex at a 10% hurdle
• 18% — Add 2027’s ~$235B, no extra payoff. 12% counting cloud leases, AI talent and the moonshot fleet’s running costs; 8% if about $250B more goes into 2028
• 13% — Alphabet, ads and Cloud, through 2027 ads alone 11%, Cloud alone 15%; 23% on capex through 2026

Meta grows half again as fast as before the buildout; Google’s ads are back to their old pace

What changed

Meta. Revenue grew 17.5% a year from 2019 to 2023 and averaged about 7% in 2022–23. It grew 22% in 2025 and 26% in the third quarter of 2026, on more than three times its 2019 revenue. In constant currency, growth peaked at 29% in the first quarter.

Alphabet. Outside Cloud it grows about 15%, its 2019–23 pace and twice its 2022–23 rate. Search accelerated from 10% in early 2025 to 19% in the first quarter of 2026 and 17% in the second; about 3 points of the first quarter’s company-wide growth was currency, against 1 in the second. YouTube ads grow about 13%, well below their 2019–23 pace. Cloud is the outlier: 28% growth in early 2025, 82% now, with a $514B backlog.

Capex went from a fifth of Meta’s revenue to over half

Google Cloud is still accelerating; Meta peaked in the first quarter

Is growth accelerating?

Meta: no, it has leveled off high. Constant-currency growth rose from 19% in early 2025 to 29% in the first quarter of 2026 and has eased by about a point a quarter since; the fourth-quarter guidance midpoint is about 25%. Against two years earlier Meta is steady at 24–26% a year, so most of the slowdown is tougher comparisons.

Alphabet: only Cloud. Its ad businesses accelerated through 2025 and have leveled off at about 15% outside Cloud. Cloud has accelerated every quarter, from 28% to 82%. The Street expects Alphabet’s total growth to ease to about 22% in the third quarter, on consensus revenue of $124.6B.

At Meta, price is doing the work

Impressions grew 19% in the first quarter of 2026 and 10% in the third, as daily users grew 3%. The average price per ad went the other way, from 6% in the fourth quarter of 2025 to 14%, with no currency effect in the third quarter. Higher prices on slowing volume are consistent with better targeting and conversion, though tighter ad supply would push prices up too. Meta credits ad performance, along with macro conditions and currency, and said on its second-quarter call that ranking advances lifted Facebook ad clicks 8.3% and conversions 15.7%. Volume is the risk: WPP Media expects social ad growth to slow from 2027 as engagement plateaus and chatbots compete for time.

By our estimate AI adds about $55B to Meta’s ads this year, $30B to Google’s and $38B to Cloud

Meta ads, 2026: $250B of revenue, $55B of it (22%) from AI at a 11% no-AI trend. AI adds about 15 of this year’s 27 points of growth.

By business

How we split AI from trend

No business has a clean control group, so the trend is a judgment. Meta, 11% a year: user growth of 3–5%, more ad load and new surfaces such as WhatsApp and Threads, and ad prices rising with the market. WPP Media puts global ad growth at 8.9% this year excluding US political advertising, and the US at 11.9%. Meta’s own 2019–23 pace, 17.5%, includes a pandemic boom on half today’s base. Google’s ads, 7.5%: their 2022–23 pace was 6.6%, Search’s 8.4%. Without AI, Search would arguably have lost queries to chatbots, so this may understate what AI is worth to Google. Cloud, 22%: it grew 26% in 2023 and was slowing with scale. The rest of each gap we leave to macro, currency and products that aren’t AI.

At these settings AI adds about 9 and 10 points to Meta’s growth in 2024 and 2025 and 15 in 2026, 22% of this year’s ad revenue; about 3–5 points a year to Google’s ads, 9% of this year’s; and 37% of Cloud’s revenue. For 2027 we assume Meta’s ads grow 18%, Google’s 11% and Cloud 55%. AI’s points are the growth in AI’s revenue over last year’s revenue, so they and the rest add up to the year’s growth.

Ads alone pay for Meta’s AI capex through 2027, even if superintelligence earns nothing

At a 11% no-AI trend with 80% of the gap credited to AI: 35% on capex through 2026, 18% adding 2027 and 8% adding 2028; all-in 27%, 12% and 3%.

How the return is counted

Returns. AI’s share of above-trend ad revenue at a 75% cash margin, after the cost of serving ads and before depreciation, taxed at 15%. Capex earns from the year after it’s spent, so 2027’s ad gains count against capex through 2026. After 2027 the gain is held as a share of the business, since a better ad system multiplies whatever advertisers spend, and fades 1.5% a year: no further share gains from AI.

Costs. All capex above 17% of no-AI ad revenue: $187B over 2022–26, and $201B more if 2027 comes in at $235B. Meta hasn’t guided 2027; in July the buy-side expected $225–250B. The 17% baseline is close to Meta’s 2021 capex (16% of revenue) and below its 2019–23 average (about 21%), so it charges more to AI than a higher baseline would; at 21% the return through 2026 would be 46%. We put servers at 60% of capex, Alphabet’s split. They are replaced every six years at 75% of their cost for half the fleet, the part the ad gains need; the rest belongs to the moonshots and earns nothing. Servers are expensed at once for tax, data centers over 20 years. Nothing is counted after 2040.

Results. Through 2026: 35%, paid back in 2029. At a 10% hurdle the ad gains could carry 2.5× that capex. With 2027 at $235B and nothing extra to show: 18%, paid back in 2032; each further $10B in 2027 takes about 0.6 points off. With another $250B in 2028: 8%, paid back in 2036. All-in: 27%, 12% and 3%. In our base case another year of moonshot spending, 2028, is what takes the return below 10%; all-in, 2027 alone brings it close. At the low end of our ranges 2027 can do it on its own, as the grid above and the simulation below show.

Spend through 2026 pays back in 2029; each further year of moonshot capex pushes payback out three to four years

At Alphabet, Cloud carries the 2027 build; the ad case turns on what Search would have done without AI

At a 7.5% no-AI trend for Google’s ads and a 50% Cloud margin: ads and Cloud together return 23% on capex through 2026 and 13% through 2027; ads alone 11% and Cloud alone 15% through 2027.

Alphabet’s two legs

Ads. Charged half the capex, Google’s AI ad gains return 33% on spend through 2026 and 11% through 2027. That rests on something nobody can observe: at a 9.5% no-AI trend the 2027 case returns −1%, at 5.5% 21%. At a 10% hurdle the ad gains alone could carry 105% of all Alphabet’s incremental capex through 2026, but only 53% through 2027.

Cloud. Revenue grew 82% to $24.8B in the second quarter, with an operating margin of 35.6%. We use a 50% cash margin on its AI revenue, before depreciation; Cloud’s figures include Gemini’s API, compute rented to AI labs and TPU sales. Charged the other half of the capex, Cloud returns 11% through 2026, when its revenue stops at the end-2026 run-rate, and 15% through 2027, as the 2027 build converts backlog: compute-rental economics, steady but not spectacular. Alphabet says renting third-party capacity as a bridge while it builds its own will weigh on Cloud’s margins.

Together 23% through 2026 and 13% through 2027, paid back in 2034. The 2027 build pays well only if Cloud keeps compounding beyond its backlog or the models earn money of their own.

Spend through 2026 very likely pays; the 2027 build is likely at Meta, closer to a coin flip at Alphabet

When does the cash come back?

Spend through 2026 clears a 10% return in all of our 8,000 draws for Meta and 93% for Alphabet. Add 2027 with nothing extra from the moonshots and it clears in 93% of draws for Meta, 61% all-in, and 66% for Alphabet. The medians, 34%, 17% and 11% for Meta and 22% and 13% for Alphabet, sit within a point of the base case.

A vintage that earns its cost of capital still doesn’t reach shareholders while the race is on. Meta bought back $26.2B of stock in the first nine months of 2025 and none so far in 2026; its free cash flow was −$2.1B in the third quarter, and it is creating a non-voting Class C share, partly to raise capital. Alphabet bought back no stock in the first half of 2026 and raised about $106B: $56B of debt, $30B of common stock and $19B of preferred. Cash comes back when capex plateaus, and that runs against success: Zuckerberg’s stated fallback if superintelligence takes longer is to put the compute into the core business and, at worst, slow new building for a while. In our judgment, buybacks back at 2025’s pace by the end of 2028 are about a 30% chance at each company, likelier if the moonshots disappoint than if they work.

What to watch

Alphabet’s third quarter, October 28. Search against 17% and Cloud against 82%. A Cloud slowdown with capex still rising is the case where the 2027 build doesn’t pay.

Meta’s 2027 capex. Meta hasn’t guided yet. Each $10B above $235B with nothing extra to show costs about 0.6 points of return on the program.

Price against volume at Meta. Price per ad growing faster than impressions is the ranking gains showing up. Both slowing together would mean the gains are fading.

Legal costs. Meta’s settlement with 52 state attorneys general provides for payments of up to about $18.6B, and more youth-related trials are scheduled. None of this is in the model.

What this is anchored on

No-AI trends, AI’s share of above-trend revenue, margins, 2026–27 estimates beyond guidance, capex baselines, replacement, off-capex spending and the returns are our estimates, built from the sources below.

• Meta Q3 2026 — Revenue $64.60B, up 26% reported and in constant currency; ad impressions +10%, average price per ad +14%; daily users 3.64B, +3%; capex $34.91B; free cash flow −$2.12B; no buybacks in 2026; Q4 guidance $73–76B with about 1 point of currency headwind; 2026 capex $135–145B; a new non-voting Class C share, partly for raising capital · Meta Q3 2026 results (8-K), Oct 21
• Meta Q3 2026 — AI investments already deliver substantial incremental ad revenue; MSL investments will take longer to return than the core business; buybacks of $26.2B in the first nine months of 2025 · Meta Q3 2026 results (8-K), Oct 21
• Meta legal — An agreement with 52 state attorneys general providing for payments of up to about $18.59B; youth-related trials scheduled in the coming months · Meta Q3 2026 results (8-K), Oct 21
• Meta Q2 2026 — Revenue $60.80B, +28% (+27% in constant currency); impressions +14%, price per ad +12%; capex $31.08B · Meta Q2 2026 results (8-K), Jul 29
• Meta Q2 2026 call — Price per ad up on ad performance, better macro conditions and currency; ranking advances lifted Facebook ad clicks 8.3% and conversions 15.7% (no period given) · MarketBeat, Jul 29
• Meta Q1 2026 — Revenue $56.31B, +33% (+29% in constant currency); impressions +19%, price per ad +12%; ad revenue $55.02B; capex outlook raised to $125–145B · Meta Q1 2026 results (8-K), Apr 29
• Meta 2025 — Q4 revenue $59.89B, +24% (+23% in constant currency); ad impressions +18%, price per ad +6% in Q4; 2025 revenue $200.97B; 2025 capex $72.22B · Meta Q4 2025 results, Jan 28
• Meta 2025 — Q3 revenue $51.24B, +26% (+25% in constant currency); impressions +14%, price per ad +10% · Meta Q3 2025 results (8-K)
• Meta 2025 — Q2 revenue $47.52B, +22% reported and in constant currency; impressions +11%, price per ad +9% · Meta Q2 2025 results (8-K)
• Meta 2025 — Q1 revenue $42.31B, +16% (+19% in constant currency); impressions +5%, price per ad +10% · Meta Q1 2025 results (8-K)
• Meta 2027 capex — No 2027 guidance; the buy-side expected $225–250B · Dolphin Research, Jul 29
• Meta off-capex — Most of 2026’s expense growth from infrastructure, including third-party cloud spend · DCD, Jan 29
• Meta off-capex — Meta funds its 20% share of Hyperion’s remaining construction as an investment, not capex · DCD, Oct 30, 2025
• Meta strategy — If superintelligence takes longer, the extra compute goes to the core business; at worst Meta slows new building for a while · Reuters, via The Financial Express
• Alphabet Q2 2026 — Revenue $119.8B, +24% (+23% in constant currency); Search & other $63.27B, +17%; YouTube ads $11.06B, +13%; Network $7.30B; Cloud $24.77B, +82%, operating income $8.81B; capex $44.9B; no buybacks in the first half; first-half proceeds of $56.2B of debt, $30.5B of common stock and $19.1B of mandatory convertible preferred · Alphabet Q2 2026 results (8-K), Jul 22
• Alphabet Q2 2026 — Cloud backlog $514B · Zacks, via CNHI
• Alphabet capex — 2026 capex guidance raised to $195–205B; Q2 capex $44.9B, about 60% of technical infrastructure in servers · W.Media, Jul 23
• Alphabet capex — Another meaningful step-up in capex in 2027 · Zacks, via Nasdaq
• Alphabet capex — Analysts’ 2027 capex projections above $300B · Pepperstone, Jul 21
• Alphabet compute — About half, or just over half, of 2026’s ML compute goes to Cloud · Alphabet CFO, Morgan Stanley TMT conference, Mar 3
• Alphabet Q1 2026 — Revenue $109.9B, +22% (+19% in constant currency); Search $60.40B, +19%; YouTube ads $9.88B, +11%; Network $6.97B; Cloud $20.03B, +63% · Alphabet Q1 2026 results (8-K), Apr 29
• Alphabet 2025 — Q4: Search $63.07B, +17%; YouTube ads $11.38B, +9%; Cloud $17.66B, +48%. Full year: revenue $402.8B; capex $91.4B · Alphabet Q4 2025 results (8-K), Feb 4
• Alphabet Q3 — Reports October 28; consensus revenue $124.6B · Finviz, Oct 21
• Alphabet Cloud — Alphabet plans to rent more third-party capacity in Q3 as a bridge while it builds its own, and says the cost will weigh on Cloud’s operating margin · EFE, via El Nuevo Diario, Jul 22
• Ad market — Global ad revenue +8.9% in 2026 excluding US political advertising, US +11.9%; social media growth slowing from 2027 as engagement plateaus and AI chatbots compete · WPP Media, via exchange4media, Jun 16
• History — Annual revenue by segment, capex and quarterly revenue, 2017–2025 · Meta and Alphabet annual reports (10-K) and earnings releases
• History — Alphabet segment revenue and capex, 2017–2025 · Alphabet investor relations

Method

Clock. Reported through Meta’s third quarter of 2026 and Alphabet’s second. Meta’s 2026 is nine months reported plus the midpoint of its fourth-quarter guidance, $74.5B, of which about $72.3B ads. Alphabet’s second half is our estimate: Search up 15% and 14%, YouTube ads 12% and 11%, Network down 2%, Cloud $27.0B and $30.5B. 2027: Meta ads +18%, Google ads +11%, Cloud +55% and +30% in 2028 as the backlog converts, both only when 2027’s capex is charged; otherwise Cloud stays at its end-2026 run-rate, $122B. Capex earns from the year after it’s spent.

AI’s share. Each business against a trend from 2023: Meta’s ads 11% a year, Google’s ads (Search, YouTube, Network) 7.5%, Cloud 22%, held through 2027 and fading to 4% (Cloud 6%) by 2035. AI is credited with 80%, 75% and 90% of the gap. AI points of growth are the growth in AI’s revenue over the year before’s revenue. After 2027 the ad gains are held as a share of the business and fade 1.5% a year; Cloud’s are held in dollars from its last estimated year (2028 with 2027’s build, 2027 without) and fade 2.5% a year. Holding Meta’s gain flat in dollars instead would cut the return through 2026 to 29%.

Returns. Cash flows from 2022 to 2040, no terminal value. In: AI’s share of above-trend revenue times a cash margin before depreciation (Meta ads 75%, Google ads 70% after traffic acquisition and YouTube content costs, a judgment, Cloud 50%). Out: capex above a pre-buildout intensity (17% of Meta’s no-AI ad revenue; 11% of Alphabet’s no-AI revenue), counted through 2026 or 2027 (2028 in one Meta case); servers, 60% of capex, replaced every 6 years at 75% of their cost for the share of the fleet the gains need (half of Meta’s, half of Alphabet’s outside Cloud, all of Cloud’s); cash tax at 15% (Alphabet 16%) with servers expensed at once and data centers over 20 years, losses offsetting other income. Alphabet’s incremental capex is split half to Cloud. All-in Meta cases add $15B a year of AI spending outside capex in 2026–31 and $7B after, plus running costs for the moonshot share of the fleet at 3% of its capex a year while its servers last. The base case is the middle of every range in the simulation.

What’s left out. Any revenue from the moonshots: Meta AI and business agents, Meta selling compute, frontier models beyond what Cloud sells, Gemini subscriptions, Waymo. The value of what AI defended rather than added, likely large for Search. Running costs of Alphabet’s non-Cloud moonshot compute, spending on AI researchers beyond Meta’s all-in cases, legal costs, and currency beyond what reported growth shows. The returns are on the AI program alone, not the companies. Monte Carlo: 8,000 draws per case with a fixed seed; the ranges are in the chart’s note. Model and data: scripts/packs/ai_payback.py in the site’s repository.

Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at 06.ai/trackers.

Not investment advice. Estimates are labeled as estimates.

The compute-demand pack sized the compute AI demand needs to the end of 2027. Here we add the supply side, carry demand to 2030 and compare the two at the end of each year. In our base case the shortage all but closes during 2027, when about 38 GW of new capacity comes online, if power keeps pace with capex. From 2028 demand pulls ahead again: what customers and labs plan to spend on compute grows faster than the power and money for new capacity, and the market ends 2030 21% short. Power and money, not chips, limit what is switched on from 2027; training budgets and capex decide whether, and how far, the market runs short.

• +38 GW — Switched on in 2027 of AI data-center capacity, all-in, limited by power and money. The chips the industry could deliver would fill about 47 GW
• −1% — End of 2027 supply about level with demand, from 19% short at the end of 2026, if power keeps pace with capex (15% short at our power research’s high case)
• −21% — End of 2030 short again. Demand grows 1.9–2.4× a year in 2028–30, supply 1.8–2.2×, in compute terms
• −69% to +43% — Biggest swing, 2030 if training and research grow at our high or low path instead of the base

Short in 2026, about even at the end of 2027, short again by the end of 2028

Base demand, base supply, demand 20% above supply today: 1% short at the end of 2027 and 21% short at the end of 2030.

Needed: gigawatts of that year’s hardware it would take, on top of everything installed before, for supply to meet demand at year-end. A surplus or shortfall carries into the next year.

What the base case says

Short through 2026. Today’s market is short. One-year H100 rental contracts rose about 40% between October and March and the B200 rental index is up about 28% this year. Every large cloud says demand exceeds what it can supply, and NVIDIA says its supply supports about 70% revenue growth in its year to January 2028, against demand growth nearer 100%. We assume demand at today’s prices runs 20% above what is being served. The fourth quarter adds about 7 GW, mostly Blackwell racks and new TPUs, while inference demand grows about 56% in the quarter, so the year ends 19% short.

2027 brings supply about level with demand. About 38 GW comes online, 60% more than the 24 GW switched on in 2026. Most of it is Vera Rubin racks and Google’s eighth-generation TPUs, plus Blackwell racks delivered late in 2026. Each gigawatt does about 4.6× the inference work of today’s average and 2.2× the training work. Supply grows about 3.1× during the year in compute terms and demand about 2.6×, so by December supply is about level with demand (−1%): 37.9 GW switched on against 38.3 GW needed. Month by month, with last year’s late deliveries switched on first, the market comes within 3 points of balance in September 2027 and stays there until August 2028. That fits Goldman’s view that supply and demand rebalance in the second half of 2027 at the earliest and more likely in the first half of 2028; Amazon expects to be short through 2027. It needs about 38 GW of power in 2027, above the top of our power research’s range: its high case is 30 GW in 2027 and 40 GW in 2028. At that high case, 2027 ends 15% short.

Then demand pulls ahead. In compute terms supply grows 2.2×, 1.9× and 1.8× in 2028–30, and demand 2.4×, 2.1× and 1.9×. Capacity switched on rises only from 43 GW to 51 GW a year, because AI capex grows 11–18% a year on street estimates. Meanwhile what customers spend on inference grows 35–70% a year in our base case, and training budgets keep growing. The market ends 2028 9% short and 2030 21% short. By 2030, 86 GW of new capacity would be needed against 51 GW switched on, because each year’s shortfall carries into the next. Either capex keeps rising faster than analysts expect, or prices rise until demand fits. The demand and supply cases move this a long way; see the controls above and “What tips the balance” below.

From 2027 the chips could outrun power and money

Supply: chips, power and money

The chips keep coming. Epoch AI counts 11.3M H100-equivalents delivered in the first half of 2026. We estimate 33M for the year, above the compute-demand pack’s 24–27M, which extended the first half’s pace: NVIDIA’s guidance implies a much larger second half, with Vera Rubin, rated at 8.8 H100-equivalents a chip, shipping from the fourth quarter. In 2027 the chipmakers could deliver about 100M as Rubin ramps, Google ships its eighth-generation TPUs and AMD its MI455X racks; NVIDIA would be about 56% of that and Google 27%. NVIDIA says supply limits its revenue at least through January 2028, Broadcom expects its AI revenue to double to about $115B in fiscal 2027 and again in 2028, and TSMC says AI demand still exceeds what it can make. Rated at about 26 GW, 2027’s chips need about 52 GW to run; allowing a quarter to install them, about 47 GW could be switched on during 2027. After 2027 we let accelerator spending rise from about $1T to $1.25T a year and each generation’s gains do the rest: 170M H100-equivalents in 2028 and 350M in 2030.

Power doesn’t keep up. About 24 GW came online for AI in 2026, by our estimate from chip deliveries; we assume power for about 27 GW was ready. For 2027 Morgan Stanley expects hyperscalers and neoclouds, 60–95% of new capacity, to add 32 GW. Goldman counts 2.7 times as much US capacity scheduled for 2027 as for 2026 but expects only half to 60% of it on time. Neither says whether it counts IT power or the whole facility. Trackers found most US projects due in 2027 not yet under construction in April, new gas turbines are being booked for delivery as late as 2031 and large transformers take more than three years. We expect about 38 GW of all-in capacity to come online in 2027, rising to 58 GW by 2030. That is within the range those forecasts imply once they are put on an all-in, worldwide basis, and above the top of our own power research’s range through 2029. Its central estimate, which applies project trackers’ delay rates, is about 23 GW in 2027 rising to 35 GW in 2030, with a high case rising from 30 GW in 2027 to 60 GW in 2030. It starts from a lower 2026, about 16 GW, partly because it takes Epoch’s end-2025 installed base at 2.5× chip power where we use 2×. Our path leaves chips for about 9 GW more than can be switched on in 2027; a consensus estimate cited by Elon Musk in August put that at about 15 GW.

Money binds first. AI capex is about $1T in 2026 (Goldman’s estimate). On street estimates for the hyperscalers, Oracle, SpaceX and the neoclouds, plus China and other buyers, it reaches about $1.4T in 2027 and $2.05T by 2030, growing 11–18% a year after 2027. All-in, a gigawatt of IT power costs about $45B to build today (Epoch puts a GB200 site at $38B, NVIDIA says $50–60B), rising to about $52B by 2030 as chips carry more value per watt. That pays for about 38 GW in 2027 and 51 GW in 2030, at or below our power path, so money is the limit in the base case from 2027. Both sit well below what the chipmakers could ship. The money includes the chips, so when it binds, fewer chips get bought than could be made, or chips wait for power. If these estimates are right, from 2027 new AI capacity is limited less by the fabs than by power and the money to build.

A gigawatt switched on in 2030 could do 19× the inference work of today’s average

Why gigawatt totals mislead

Today’s fleet is still about a fifth Hopper. On agentic inference a GB300 rack does about 12.6× an H200’s work per MW, and Vera Rubin about 2.1× a GB300’s (SemiAnalysis, measured on pre-release software). So the capacity switched on in 2027 does about 4.6× today’s average per gigawatt at inference, and 2030’s about 19× if Rubin Ultra and Feynman deliver what their roadmaps suggest. Training gains ride on FLOPs per watt, which improve about 30% a year: 2.2× today’s average for 2027’s capacity and 5.0× for 2030’s. These are peak 8-bit FLOPs, the way Epoch counts them. The training throughput our efficiency research expects newer chips to achieve is lower for capacity switched on through 2029; counted that way, 2027 would end 11% short and 2030 25% short. That research labels its tables by install year. We read them as each year’s deliveries, because its 2026 mix includes Rubin, which only ships from the fourth quarter; read as install years, 2027 would end 3% long and 2030 17% short.

Cheaper inference also raises demand. Capacity switched on in 2028 does about 58% more inference work per dollar than 2027’s, and about 60% and 52% more again in 2029 and 2030. If prices follow costs, each dollar customers spend buys that much more, so after 2027 inference demand rises with hardware gains just as supply does. Faster hardware therefore moves the balance mainly through training, which we count in FLOPs, and through the years to 2027, when demand follows usage.

Who gets the new hardware matters too. We split each year’s hardware between inference and training in the same proportion. Newer hardware’s edge is larger at inference (4.6× today’s average in 2027, against 2.2× at training). If it went to inference first and older chips did the training, the market would serve about 6–7 points more of demand each year. If the newest went to training first, as frontier labs train on their newest clusters, it would serve about 8–12 points less.

After 2027 demand grows faster than supply

Demand to 2030

Inference to 2027 is the compute-demand pack’s: OpenRouter’s compute on the AI-demand pack’s three paths. To scale it to the whole market we compared OpenRouter’s growth with six disclosed token counts over the same windows. The market grew at 0.32–0.86 of OpenRouter’s pace, measured in doublings, with a median of 0.63; we use 0.6. We apply it to OpenRouter’s compute rather than its tokens, because OpenRouter’s mix is probably moving to small, cheap models faster than the market’s; applied to tokens, demand at the end of 2027 would be about 28% lower. That puts the market’s inference demand at 4.3× today’s by the end of 2027, about 2.8× during 2027 alone. The compute-demand pack’s headline, about 47 GW of new GB300 racks by December 2027, assumed the whole market grew at OpenRouter’s full pace. At 0.6 of it, measured as the compute-demand pack does from October’s average to December’s, inference needs about 15 GW of GB300s, or about 10 GW of the capacity switched on in 2027, before counting today’s shortage.

After 2027, spending decides. The AI-demand pack’s paths have OpenRouter’s growth halving every three, six or twelve months. Carried past 2027, the base path has OpenRouter’s compute growing 65% in 2028, 14% in 2029 and 3% in 2030, which describes one marketplace maturing rather than the market. Instead we let inference demand follow what customers spend on it: 1.7× in 2028, 1.5× in 2029 and 1.35× in 2030 in the base case. That is slower than the leading labs plan to grow. OpenAI’s plan takes revenue from $36B in 2026 to $350B in 2030, about 1.8× a year, and Anthropic’s internal forecast of $190–200B in 2028 implies about 1.9× a year from an estimated $55B in 2026. Each dollar buys as much more inference as the capacity switched on that year delivers per dollar, 1.5–1.6× more each year. So inference demand grows about 2.7× in 2028, 2.4× in 2029 and 2.1× in 2030, reaching 57× today’s level by 2030.

Training and research use about 60% of today’s chips, and the compute-demand pack didn’t project them. We let them grow 2.4× in 2027, slowing to 1.6× in 2030, measured in 8-bit FLOPs: 19× today’s by 2030. The anchors: OpenAI’s compute roughly quadrupled in each of 2024 and 2025 (Epoch AI); it ended 2025 with 1.9 GW and targets 30 GW by 2030. Anthropic, at about 2 GW early in 2026, has placed about 5 GW of demand with Broadcom for 2027 and 10 GW for 2028 and could pass 30 GW by 2030 (22V Research). The power behind frontier training runs grows 2.2–2.9× a year (Epoch AI). We assume market-wide training grows more slowly than these labs’ plans: they use less than half of the world’s AI compute.

The market grows at about 0.6 of OpenRouter’s pace

What tips the balance

What would change the numbers

Training and research budgets. The biggest lever. If they grow only 4.5× by 2030, supply ends 2030 43% ahead of demand, a glut. If they grow 66×, the market is 69% short. Much of this is spent by a few labs that are still raising the money.

Power and money. If capacity added stays near 30 GW a year, 2030 is 43% short. At our power research’s central estimate it is 47% short, and at its high case 24% short. If power follows SemiAnalysis’s path, about 90 GW a year by 2029, with capex to match, supply runs 25% ahead by 2030. Chips aren’t what holds supply back in our base case: if power and money kept up with every chip the industry could deliver, supply would run 17% ahead by 2030.

Inference spending. If what customers spend on inference grows only 1.2×, 1.1× and 1.0× in 2028–30, the market stays near balance (+5% in 2030). If it grows 2.35×, 2.0× and 1.7×, faster than even OpenAI’s plan, 2030 is 49% short.

Hardware. If Rubin Ultra slips (SemiAnalysis reported in July that its 144-GPU rack moves to 2028), Feynman disappoints and the 2026–27 mix underdelivers too, 2030 is 43% short. If they beat their roadmaps, it is 11% long. Because cheaper inference also raises demand, hardware moves the balance mostly through training and the years to 2027. Software gains on existing chips help a little: the efficiency research’s central case, software gains of 30% a year against workloads 20% heavier, would leave 2030 18% short instead of 21%.

Who gets the new hardware. If the newest hardware went to inference first, where its edge is largest, 2030 would be 14% short; if it went to training first, 30% short.

Today’s shortage. With no shortage today, supply leads by 19% at the end of 2027 and 2030 ends 6% short. If demand runs 30% above supply today, 2027 ends 8% short.

Inference to 2027. The pace against OpenRouter and inference’s share of today’s chips move 2027 between 12 points up and 13 points down. Letting OpenRouter’s growth fade, as the AI-demand pack’s base path does if carried to 2030, would leave 22% more compute than demand by 2030.

What to watch

Gigawatts switched on in 2027. Our base case has about 38 GW. Hyperscalers report capacity added each quarter, and turbine and transformer deliveries lead them. At our power research’s high case, 30 GW, the market ends 2027 15% short.

Capex plans for 2028–30. Street estimates have AI capex growing about 18% in 2028 and 11% by 2030. If lab revenue grows as planned, that is too little; capex rising faster than that is how the gap would close.

Lab revenue against plan. OpenAI’s and Anthropic’s revenue, funding rounds and IPOs decide both what customers spend on inference and what labs spend on training, the two largest demand inputs after 2027.

Rubin in the field. SemiAnalysis measured Rubin at about 2× GB300 per MW on pre-release software. Production results, and when Rubin Ultra ships, set how much each 2027–28 gigawatt is worth.

Rents. In September the B200 rental curve priced the 12-month forward about 12% below spot: the market expects some easing within a year, consistent with supply catching up during 2027.

What this is anchored on

Supply, power, capex and demand paths, efficiency by year, the shortage today and the balance are our estimates, built from the sources below. OpenRouter data is licensed under CC BY 4.0.

• Today’s fleet and demand — AI chips operating, their power, the inference share, AgentX throughput by system and OpenRouter’s compute paths to December 2027 · 06.ai compute-demand pack
• OpenRouter paths — Tokens and spend to December 2027, with growth halving every 3, 6 or 12 months · 06.ai AI-demand pack
• OpenRouter tokens — Daily tokens by model, January 2025 to October 2026, for the pace comparison · OpenRouter rankings dataset (CC BY 4.0)
• Chips delivered — 20.3M H100-equivalents delivered by end-2025; 4.98M and 6.33M in the first two quarters of 2026, rated at 2.7 and 3.2 GW · Epoch AI chip-sales data (via the compute-demand pack)
• Chips by designer — Q2 2026 share of compute sold: NVIDIA 63%, Google 25%, Amazon 5%, AMD 4%, Huawei 3% · Epoch AI, updated Oct 19
• Chip ratings — Vera Rubin 8.84 H100-equivalents per chip; TPU v8i 5.1; B300 2.53 · Epoch AI chip-types data, Oct 14
• Facility power — A data center draws about twice its AI chips’ rated power · Epoch AI chip-sales data
• NVIDIA — A supply-constrained outlook of about 70% revenue growth in the year to January 2028; supply the bottleneck at least through then · CFO Dive, Aug 27
• NVIDIA — At least $1T of Blackwell and Rubin demand and orders through 2027 · DCD, GTC, Mar 16
• NVIDIA — A GW costs about $50–60B; NVIDIA’s content per GW about $18B for Hopper, $25B for Blackwell, $40B for Vera Rubin; supply supports about 70% revenue growth in fiscal 2028, against demand growth nearer 100%; top-5 hyperscaler capex near $800B in 2026 and $1.3T in 2027 · NVIDIA Q2 FY27 call, Aug 26
• Broadcom — AI revenue outlook about $115B in fiscal 2027 and $230B in fiscal 2028 · Bloomberg, via The Star, Sep 4
• TSMC — AI demand still exceeds what TSMC can supply · TSMC Q3 call, via AP, Oct 15
• Packaging — TSMC CoWoS capacity about 130K wafers a month at end-2026, doubling by 2028 · TrendForce, Sep 14
• Memory — HBM bit shipments up 50–60% in 2027, still short of demand · TrendForce, Aug 4
• Google TPUs — TPU shipments estimated at 4.5M in 2026 and 8.8M in 2027 · GF Securities, via 24/7 Wall St, Aug 25
• AMD — Helios racks shipping from late Q3 2026; 6 GW for OpenAI, 6 GW for Meta and up to 2 GW for Anthropic · Converge Digest, Jul 24
• Power — Hyperscalers and neoclouds add 32 GW in 2027 and 39 GW in 2028, 60–95% of new capacity worldwide · Morgan Stanley, via Zhitong, Oct 5
• Power — US data-center capacity scheduled for 2026 13.6 GW and 2027 36.3 GW; only 50–60% assumed on time · Goldman Sachs, via OilPrice, Aug 22
• Power — About 30 GW of AI compute added in 2026, 50 GW in 2027, 70 GW in 2028 and 90–100 GW in 2029 · Dylan Patel (SemiAnalysis), via The Neuron, Aug 31
• Power — A consensus estimate that about 15 GW of AI compute made in 2027 can’t be switched on in 2027 · Elon Musk, via 24/7 Wall St, Aug 29
• Power — At least 16 GW of large projects slated for 2026 worldwide, about 5 GW under construction in February; 30–50% unlikely to finish on time · Sightline Climate, via Latitude Media, Feb 24
• Power — More than 60% of US projects due in 2027 not under construction in April · SynMax, via the FT (Longbridge), Apr 17
• Equipment — A heavy-duty GE Vernova gas turbine ordered now arrives in 2031; Mitsubishi Heavy’s new orders deliver in 2028–30 · OilPrice, Aug 22
• Equipment — Generator step-up transformers take more than 160 weeks; one utility is buying about five years ahead · Wood Mackenzie, via Reuters, Jul 9
• Capex — Global AI investment about $1T in 2026 · Goldman Sachs, Aug 7
• Capex — Alphabet, Microsoft, Amazon, Meta, Oracle and SpaceX: about $800B in 2026, $1.2T in 2027 and $1.4T in 2028 · Goldman Sachs, via Crypto Briefing, Sep 25
• Capex — Hyperscaler, Oracle, SpaceX and neocloud capex to 2030 · 06.ai hyperscaler, neocloud and leveraged-middle packs
• Cost per GW — $38B per GW of IT power for a GB200 NVL72 site · Epoch AI, May 14
• Cost per GW — $34.6–39.5B per GW by architecture, from OpenAI’s custom chip to Vera Rubin · Bernstein, via Phemex, Oct 10
• Inference per MW — Vera Rubin NVL72 59.4M tokens/s per MW on AgentX at 100 tokens/s per user, 2.09× GB300 NVL72 (pre-release software) · SemiAnalysis InferenceX, Sep 14
• Inference per MW — Vera Rubin about 2× GB300 per MW up to 100 tokens/s per user on DeepSeek-R1 · SemiAnalysis, Jul 23
• Rubin — 35 PF dense NVFP4 and 17.5 PF dense FP8 per GPU · NVIDIA Vera Rubin NVL72 page
• Rubin Ultra — Kyber NVL144 rack reported slipping more than a year, to 2028 · SemiAnalysis, via MLQ, Jul 6
• Hardware trends — Chip performance per dollar up about 49% a year since 2023 · Epoch AI, Aug 13
• Hardware trends — Peak FLOP/s per watt of ML GPUs up about 1.28× a year · Epoch AI, Aug 2024
• Tokens — Google: 480T tokens a month in May 2025, more than 3.2 quadrillion in May 2026 · Google I/O, May 19
• Tokens — Google: about 1.3 quadrillion tokens a month in October 2025 · Google, via Robonomics
• Tokens — Google’s models: 7B tokens a minute via direct API use in October 2025 · Alphabet Q3 2025 call, Oct 29, 2025
• Tokens — Google’s model APIs: about 22B tokens a minute in July 2026 · Alphabet Q2 call, via MarketBeat, Jul 22
• Tokens — OpenAI’s API: 6B tokens a minute in October 2025 · TechCrunch, Oct 6, 2025
• Tokens — OpenAI’s APIs: more than 15B tokens a minute as of March 31 · Quartz, via Yahoo Finance, Mar 31
• Tokens — Doubao: more than 30T tokens a day in September 2025 · Volcano Engine, via KrASIA, Oct 21, 2025
• Tokens — Doubao: more than 180T tokens a day in June 2026 · Volcano Engine, via Diario Bitcoin, Jun 23
• Tokens — China: about 100T tokens a day at end-2025, more than 140T in March 2026 · National Data Administration, via China Daily, Mar 25
• Lab revenue — OpenAI plans revenue of $36B in 2026 rising to $350B in 2030, and about $856B on compute and infrastructure · FT, via Reuters, Sep 18
• Lab revenue — Anthropic’s internal forecast of about $190–200B of revenue in 2028 · Reuters, via The Next Web, Aug 17
• Lab compute — OpenAI ended 2025 with about 1.9 GW of compute, more than triple a year earlier · DCD, Jan 21
• Lab compute — OpenAI targets 30 GW by 2030 · YourStory, Apr 23
• Lab compute — Anthropic at about 2 GW early in 2026; about 5 GW of demand placed with Broadcom for 2027 and 10 GW for 2028; over 30 GW by 2030 · 22V Research, Sep 4
• Lab compute — OpenAI’s compute about 0.39M H100-equivalents in 2024 and 1.74M in 2025 · Epoch AI, Sep 9
• Training — The power drawn by frontier training runs grows 2.2–2.9× a year; the largest could draw 4–16 GW by 2030 · Epoch AI and EPRI, Aug 2025
• Shortage — H100 one-year rental contracts from about $1.70 an hour in October 2025 to $2.35 in March 2026 · SemiAnalysis, Apr 2
• Shortage — B200 rental index up about 28% this year to September 1; 12-month forward about 12% below spot in September · Silicon Data
• Shortage — Not enough capacity to meet demand in 2026, likely the same in 2027 · Amazon Q2 call, via The Motley Fool, Jul 30
• Shortage — Supply and demand rebalance in the second half of 2027 at the earliest, the first half of 2028 more likely · Goldman Sachs, via KuCoin, Sep 15

Method

Units and clock. Supply and demand are both counted in gigawatts of today’s hardware: the work one gigawatt of the fleet operating in September 2026 does, for inference (AgentX tokens, the compute-demand pack’s measure) or for training (peak 8-bit FLOP/s, Epoch AI’s H100-equivalents). Today’s fleet is 31.6M H100-equivalents, about 40 GW all-in, 40% of it serving inference. A gigawatt installed later counts for its multiple of today’s average. Today is October 1, 2026: the fleet is the chips delivered to June 2026, demand is measured from October 1 (OpenRouter’s October figure covers October 1–18), and every balance is at December 31. The balance is the largest share of both demands the installed capacity can serve at once, with every year’s hardware split between inference and training in the same proportion. All power figures are all-in, at the facility, about 1.3× IT power; several forecasts we cite don’t say which they count.

Chips. Our estimate of what the designers Epoch tracks could deliver: 33, 100, 170, 250, 350M H100-equivalents in 2026–30 (range 185–550M in 2030), rated at 15, 26, 34, 38, 41 GW. 2026 builds on Epoch’s first half and NVIDIA’s guided second half; 2027 on NVIDIA’s guidance, Broadcom’s outlook, TPU and AMD unit estimates and CoWoS and HBM capacity; 2028–30 on accelerator spending of $1T, $1.15T, $1.25T and about 30% more H100-equivalents per dollar each year. Chips switch on a quarter after delivery, at twice their rated power: 30% of a year’s deliveries come in its fourth quarter and switch on the next year. We split 2026’s second half between the third and fourth quarters 9:13, as for its H100-equivalents; the fourth quarter’s heavier Rubin mix would put a little more power in the third.

Power and money. Power: 38, 46, 52, 58 GW in 2027–30, our estimate from analysts’ forecasts put on an all-in, worldwide basis (Morgan Stanley, Goldman). Our power research’s own synthesis is lower: 23, 28, 31, 35 GW, high case 30, 40, 50, 60 (sensitivities). Low 30, 33, 35, 37; high 50, 70, 95, 105 (SemiAnalysis to 2029; 2030 ours). Money: AI capex of $1T, $1.4T, $1.65T, $1.85T, $2.05T in 2026–30, from the street capex in the hyperscaler, neocloud and leveraged-middle packs plus China’s largest platforms, sovereigns and others (86–88% AI), divided by $45, 48, 50, 51, 52B per GW of IT power and multiplied by 1.3 for facility power. The money includes the chips. What comes online each year is the lowest of chips, power and money; chips beyond that aren’t counted.

Each year’s hardware. Inference: AgentX tokens per MW for each year’s deliveries, 0.8, 1.7, 2.7, 4.5, 6.9× a GB300 NVL72 rack in 2026–30, our estimate anchored on SemiAnalysis’s measurements (Vera Rubin 2.09×) and roadmaps; low 0.55, 1.2, 1.8, 2.7, 3.8; high 1, 2.5, 4.4, 7.8, 13. Divided by today’s fleet average, 0.33× GB300. Training: H100-equivalents per GW from the chips’ rated power, against today’s 625 W per H100-equivalent; low and high from our range for training gains per MW. Capacity switched on in a year blends the previous year’s fourth-quarter deliveries with its own first three quarters, by power. Our efficiency research labels its tables by install year; we read them as deliveries (a sensitivity). Software stays at today’s level: we assume its gains offset heavier models and longer contexts (the efficiency research’s central case nets about 8% a year; a sensitivity).

Demand. Inference to December 2027: the compute-demand pack’s OpenRouter compute paths raised to the power 0.6, the market’s pace (calibrated above; 0.45–0.75 as a sensitivity), read at December 31. After 2027: what customers spend on inference grows 1.7, 1.5, 1.35× a year in the base case (1.2, 1.1, 1 low; 2.35, 2, 1.7 high), times the gain in inference work per dollar of the capacity switched on that year (1.58, 1.60, 1.52× in the base case, from the same hardware and supply cases). Training and research: 2.7, 2.4, 2.1, 1.8, 1.6× a year in 2026–30 (2.2, 1.7, 1.4, 1.3, 1.2 low; 3.2, 3.2, 2.8, 2.5, 2.2 high), the first applied to October–December only. Today’s shortage, demand 20% above supply (0–30%), scales demand in every year.

Needed GW is the capacity of that year’s hardware that, added to everything installed before, makes supply equal demand at year-end. Utilization stays at 60% throughout, as in the compute-demand pack; the table gives the utilization at which supply would just meet demand. We keep today’s chips in service and don’t model refreshes: replacing Hopper inside existing buildings would add supply without new power, though not without money. Disclosed token counts use different definitions; each is compared only with itself over time.

Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at 06.ai/trackers.

Not investment advice. Estimates are labeled as estimates.

OpenRouter’s 23.8T tokens a day would need about 60 MW of today’s hardware at benchmark throughput. In the AI-demand pack’s base case its tokens grow 22× by December 2027, but at today’s prices the compute they need grows only about 10×, because cheap models keep taking share. If the whole market’s inference demand grew at that pace, about 7.2× a year, it would need about 47 GW of new GB300-class capacity by the end of 2027, or 22 GW if it were all Vera Rubin. The AI chips sold in all of 2026 add roughly 10–11 GW for inference.

• 60 MW — OpenRouter today all-in, on today’s mix of hardware at benchmark throughput. On GB300 racks: 20 MW, about 9,300 GPUs
• 10× — OpenRouter’s compute, Dec 2027 today’s, for 22× the tokens and 4.2× the dollars in the AI-demand pack’s base case. About 195 MW of GB300s
• +47 GW — The market at OpenRouter’s pace of new GB300-class inference capacity by Dec 2027, about 22 million GPUs, 43 GW of it during 2027. +11 GW at half the pace
• 10–11 GW — 2026’s new chips, for inference of 25–28 GW added in all, on mixed hardware: 24–27M H100-equivalents, more than the 20.3M delivered in all years through 2025 (our estimate from Epoch AI data)

A token isn’t a token

Four things set what a token costs

The model. Compute per token rises with the parameters a model uses for each token more than with its total size, though total size and the length of the conversation matter too. On one B200, gpt-oss-120b, which uses 5.1B parameters per token, makes 3.8B tokens a day. DeepSeek-R1, at 37B, makes 0.40B. MiniMax-M2.5 uses a quarter of GLM-5’s active parameters and makes 3.9× its tokens. The rule is rough: Llama 3.3 70B, which uses all 70B of its parameters for every token, makes more than DeepSeek-R1. We size every model by its active parameters.

The hardware. Serving DeepSeek V4 Pro to coding agents at 100 tokens/s per user, a megawatt of GB300 NVL72 racks serves 13× the tokens of a megawatt of H200s running at 8-bit precision, and 4.1× those of B200s. Vera Rubin, now shipping, doubles GB300 again. Racks of 72 GPUs win on large models because they spread the model’s experts and its memory of the conversation across many chips. At full load, a gigawatt of GB300s would get through about 100 days of OpenRouter’s October volume each day, if every token were a DeepSeek V4 Pro agent token.

The kind of token. Output is 2.4% of OpenRouter’s tokens but about 34% of the GPU time behind them. New input is 19% of the tokens and 55% of the time. Cached input, the context agents send again at every step, is 79% of the tokens and about 11% of the time. These weights follow the labs’ own pricing: output costs about five times new input, and a cache read a tenth to a twentieth of it; we use a twentieth.

Speed and memory. Serving each user faster means smaller batches. DeepSeek-R1 on a B200 makes 6,799 tokens/s per GPU at 30 tokens/s per user and 2,685 at 100. A model that doesn’t fit on one server is worse: at 50 tokens/s per user, DeepSeek-R1’s best H100 setup made 74 tokens/s per GPU, against 4,610 on a B200.

On OpenRouter, 12% of the tokens use about half the compute

OpenRouter in GPUs

At SemiAnalysis’s measured throughput and 60% utilization, October’s 23.8T tokens a day would need about 20 MW of GB300 racks, about 9,300 GPUs. On H200s they would need 250 MW. On a mix of hardware like the industry’s installed base today, part Hopper, part Blackwell, part TPU and other chips, they would need about 60 MW. Every megawatt here is all-in: chips, servers, networking and cooling. Real deployments run older software than a tuned benchmark, which pushes the need up; how closely the benchmark’s traffic matches OpenRouter’s moves it either way (see Method).

The compute isn’t where the tokens are. Open models that use 20B parameters or fewer per token, led by DeepSeek V4.1 Flash with 21% of all tokens, are 50% of the tokens, 17% of the compute and 4% of the dollars. Closed models above that size, among them Claude Opus 5.5 and Sonnet 5.5, GPT-6 Sol and Astra and Gemini 3.8 Flash, are 12% of the tokens, 64% of the dollars and, by our price-based estimate, 49% of the compute. Opus 5.5 alone is 3% of the tokens and about 18% of the compute. Those compute shares rest on what we assume about the labs’ margins: at gross margins of 45–75% instead of 60%, the larger closed models are 38–56% of the compute.

Tokens 22×, compute 10×, dollars 4.2×

OpenRouter in 2027

In the base case OpenRouter handles 513T tokens a day by December 2027, 22× October’s, while spend grows only 4.2×, so the average price of a token falls 5.1×. Holding each model’s price fixed, as we do, that requires cheaper models to keep taking share: in the mix that reproduces the AI-demand pack’s numbers, DeepSeek V4.1 Flash’s tokens grow 48× and Claude Opus 5.5’s 1.5×. Because the cheap tokens need little compute each, compute grows 10×, from 20 MW of GB300s to 195 MW, about 84,000 more GPUs. On today’s mix of hardware it would be 600 MW. If prices fall model by model instead, as GPT-5.6 Luna’s did when OpenAI cut them 80% this summer, compute grows closer to tokens.

The low path needs 4.1× today’s compute (80 MW of GB300s) and the high path 23× (460 MW). Even the high case is under half a gigawatt of GB300s. OpenRouter on its own is a rounding error in a year that adds tens of gigawatts. What matters is what it says about everyone else.

If the whole market grew at OpenRouter’s pace

New inference capacity the whole market needs, cumulative from October 2026, in GW of all-in power, if its inference demand grows at a share of OpenRouter’s projected compute growth in the AI-demand pack’s base case. It starts from about 16 GW of chips doing inference today and assumes the new capacity is all GB300 or all Vera Rubin. The band is our estimate of the inference share of the chips added in 2026, a mix of hardware that serves fewer tokens per MW than GB300 racks.

At the default setting (Market grows at of OpenRouter’s projected monthly rate 100%), from the interactive version:

• New capacity by Dec 2027: +47 GW
• GPUs: 22 million
• In 2027 alone: +43 GW
• Market inference demand vs today: 10×

From OpenRouter to the market

OpenRouter is a small slice. On matching dates its tokens were about 3% of Google’s across all its products, 1.3% of a national total for China that TechNode reported for June without a source or method, and 0.6–2.2% of published estimates for the whole world. Its share of the compute is smaller still, about 0.4% of the chips doing inference by our estimate. Partly that is because its tokens are cheap: half come from small open models that coding agents call over and over. Partly it is because benchmarks flatter real deployments: at benchmark throughput, today’s inference chips could serve about 6.2 quadrillion OpenRouter-like tokens a day, 6.3× the highest estimate for the world.

So we scale up from the hardware, not the tokens, and use the benchmarks only to compare one kind of hardware with another. Epoch AI estimates 31.6M H100-equivalents of AI chips operating by September: chips delivered through June, allowing about three months to install. At about 1.25 kW each all-in, that is 39.5 GW. If 40% of it serves inference, that is 16 GW. If the market’s inference demand grows the way OpenRouter’s compute does in the base case, 10× by December 2027, it needs about 160 GW of hardware like today’s. Built as GB300s, which serve about 3.1× as many tokens per MW as today’s average, the increase is about 47 GW, or 22 million GPUs. Built as Vera Rubin, it is about 22 GW. Almost all of it falls in 2027: 43 GW of GB300s between December 2026 and December 2027.

That takes the premise at face value, and the premise is aggressive. At full pace the market’s inference compute grows about 7.2× a year, roughly what Google’s tokens did in the year to May; over the same year OpenRouter’s tokens grew 12.5×. The installed base of AI chips grew 3.0× in 2025. At half OpenRouter’s growth rate, about 2.7× a year, close to the chip stock’s growth, market demand rises 3.2× by December 2027 and needs 11 GW of new GB300s. Whether the answer is closer to 11 GW or 47 GW depends on whether the rest of the market speeds up toward OpenRouter’s pace.

Three inputs we can’t pin down move the 47 GW: the share of today’s chips doing inference (30–50% gives 35–58 GW), how efficient today’s fleet is (0.75–1.3× our estimate gives 35–60 GW) and how many chips are installed (Epoch’s range, roughly 1.5× either way, gives 31–70 GW). Moving all three together gives 17–113 GW.

2026 is on track to add more AI compute than all earlier years combined

What 2026 added

Epoch AI’s chip-sales data counts 11.3M H100-equivalents sold in the first half of 2026, against 13.6M in all of 2025. If the second half runs 10–40% above the first, 2026 adds 24–27M, more than the 20.3M delivered in all the years through 2025. Those chips are rated at 12–14 GW. By Epoch’s rule of thumb a data center needs about twice its chips’ rated power once servers, networking and cooling are counted, so 2026’s chips need 25–28 GW all-in. NVIDIA’s data-center revenue was $164B in the first half of its fiscal 2027, against $194B in all of fiscal 2026, and Vera Rubin began production shipments in August. NVIDIA expects supply to stay the bottleneck at least through January 2028.

Not all of it serves inference. At 40%, 2026 gives inference about 10–11 GW, on a mix of hardware that serves fewer tokens per MW than GB300 racks. Demand at OpenRouter’s pace would need about 4.1× that in 2027 alone if it is built as GB300s, or 1.9× as Vera Rubin. At half the pace, 9.7 GW of GB300s in 2027 is about the inference share of a 2026-sized year, and more if built from 2026’s mix of chips. This pack is about demand; a later one will look at how much capacity 2027 can actually deliver.

What would change the numbers

The pace. It is the biggest lever. From a quarter of OpenRouter’s growth rate to all of it, the need by December 2027 runs from 4.0 GW to 47 GW of GB300s.

The mix. Our compute grows less than tokens because, at fixed prices, cheap models keep taking share. If instead the mix froze and every model’s price fell, compute would grow with tokens, 22×, and the market would need about 106 GW. Two other choices matter less. If the larger closed models’ tokens grew 3.5×, as fast as OpenAI’s and Anthropic’s revenue run-rates in the lab-revenue pack, the market would need 45 GW. If free tokens stayed flat instead of growing with the rest, 42 GW.

How big the closed models are. Nobody discloses them, so we infer their size from price, assuming the labs keep 60% of it as gross margin. At 45% the models are bigger, take more of today’s compute and grow more slowly: 41 GW. At 75%, 55 GW.

How compute scales with size. We make compute per token proportional to active parameters. Across InferenceX’s six B200 results it rises more slowly, roughly with active parameters to the power 0.89. On that basis small models cost relatively more, the shift toward them saves less, and the market would need 53 GW.

Software. Benchmarks move every month. InferenceX has measured gains of 1.5–2× from software alone on the same chips within weeks, on DeepSeek-R1 and on DeepSeek V4 Pro. Each doubling roughly halves the need; we hold today’s software fixed.

Speed. We measure at 100 tokens/s per user. Serving agents faster costs throughput steeply: on DeepSeek V4 Pro, a GB300 rack serves about a fifth as many tokens per MW at 150 tokens/s per user as at 100, while Vera Rubin holds up better. Faster agents would raise every figure here and widen Rubin’s lead.

Output. Agents write about 1% of their tokens. An output token costs about five times a new input token, so agents that write more would raise compute faster than tokens.

What to watch

Google’s next token count. Its third-quarter results, due in late October, are the cleanest read on whether the rest of the market is growing near OpenRouter’s pace.

Rubin’s share of 2027 deliveries. Each GW of Vera Rubin does the work of about two of GB300s, so the mix of what ships in 2027 roughly halves or doubles the gigawatts needed.

Cache rates. OpenRouter’s agent traffic went from about 60% cached in January to nearly 90% in September, which kept compute growth below token growth. It can’t rise much further, so from here more of token growth turns into compute growth.

What this is anchored on

Compute, power, GPU counts, tier shares, the 2027 mix and the market estimates are ours, built from the sources below. OpenRouter data is licensed under CC BY 4.0.

• OpenRouter tokens — Daily tokens for the 50 biggest models plus one row for the rest, October 1–18, 2026; list prices from its model catalog · OpenRouter rankings-daily dataset and models API. Licensed under CC BY 4.0
• Token shape — Agent API keys used about 7× the tokens of human keys in late September; agent tokens about 87% cached input, 12% new input and 1% output; human tokens about 12% output, mostly new input (the 20% cached is our assumption) · OpenRouter, “Cache is Crucial,” Oct 8
• Demand forecast — OpenRouter tokens and spend to December 2027: low, base and high paths · 06.ai AI-demand pack
• Lab revenue — OpenAI’s and Anthropic’s annualized revenue to December 2027 · 06.ai lab-revenue pack
• Price cuts — OpenAI cut GPT-5.6 Luna’s prices 80%, to $0.20 per million input tokens and $1.20 output · DevTk.AI, August 2026
• Agent throughput — Total tokens/s per utility MW serving DeepSeek V4 Pro at 100 tokens/s per user on the AgentX benchmark, interpolated between measured points: Vera Rubin NVL72 59.4M (pre-release TensorRT-LLM), GB300 NVL72 28.5M, B200 6.95M, B300 5.56M, H200 2.26M, MI355X 2.01M · SemiAnalysis InferenceX, Sep 14
• Agent traffic — AgentX replays Claude Code sessions; on DeepSeek V4 median 88K tokens in and 413 out per request, 73–96% served from cache · SemiAnalysis InferenceX, Aug 24
• Token accounting — AgentX totals include reused (cached) input tokens · InferenceX glossary
• Power basis — InferenceX’s power per GPU is an all-in utility envelope that includes cooling and data-center overhead (near 2.17 kW for a B200 in this post) · InferenceX, Jun 9
• Power per GPU — H100 and H200 1.37 kW, B200 1.71 kW, B300 1.90 kW, MI355X 2.09 kW · InferenceX chip pages, Oct 19
• Power per GPU — GB300 NVL72 12,350 tokens/s per chip and 5.83M per MW, so 2.12 kW each · InferenceX, DeepSeek-R1 GB300 vs MI355X
• Model benchmarks — Tokens/s per GPU at 8K in / 1K out and 50 tokens/s per user: gpt-oss-120b, Llama 3.3 70B and DeepSeek-R1 on H100, H200 and B200 · InferenceX run pages
• Model benchmarks — MiniMax-M2.5 on H100 and B200, 8K in / 1K out · InferenceX, May 26
• Model benchmarks — GLM-5 on H200 and B200, 8K in / 1K out · InferenceX, May 26
• Model benchmarks — Kimi K2.5 on H200 and B200, 8K in / 1K out · InferenceX, May 26
• GB200 vs GB300 — GB300 NVL72 made 1.2–2.8× GB200 NVL72’s tokens per GPU on DeepSeek V4 Pro at matched speeds (8K in / 1K out, runs a month after the model’s release) · InferenceX, May 27
• Model sizes — DeepSeek V4 Pro: 1.6T parameters, 49B active · Hugging Face blog
• Model sizes — DeepSeek V4.1 Flash: 552B parameters, 8B active on input and 16B on output; other open models from their model cards, except the estimates named in Method · OpenRouter model pages; Hugging Face
• AI chips delivered — 20.3M H100-equivalents delivered by end-2025; 31.6M operating by September 2026, from deliveries through June with a three-month lag to install · Epoch AI, Sep 16
• Power per chip — 94 tracked AI data centers: 14.8M H100-equivalents, 14.2 GW of IT power, 18.5 GW at the facility; about 44% of global AI compute · Epoch AI data-center hub, Oct 19
• Chips added — H100-equivalents and chip power by quarter, from estimated sales by chip · Epoch AI chip-sales data, Oct 14 (our sums)
• Facility power — A data center draws about twice its AI chips’ rated power · Epoch AI chip-sales data, FAQ
• Inference share — OpenAI’s 2025 compute split roughly evenly between research and inference · Epoch AI, Sep 9
• NVIDIA — Data-center revenue $193.7B in fiscal 2026; $75.2B and $89.0B in the first two quarters of fiscal 2027 · NVIDIA filings and CFO commentary
• NVIDIA — Production shipments of Vera Rubin began in August; supply expected to remain a bottleneck at least through the end of fiscal 2028, in January 2028 · NVIDIA Q2 FY27 earnings call, Aug 26
• Software gains — DeepSeek V4 Pro on B200: about 300K to nearly 500K tokens/s per MW in six weeks · InferenceX, Jun 9
• Software gains — DeepSeek-R1 on B200: up to 1.8× from one SGLang release · InferenceX, May 2
• Google — More than 3.2 quadrillion tokens a month across its products, May 2026, 7× a year earlier · Google, May 19
• Google API — About 22 billion tokens a minute through customers’ direct API use, Q2 2026 · Google, Jul 22
• OpenAI API — More than 15 billion tokens a minute, as of its March 31 funding round · CoinDesk, Apr 1
• ByteDance — Doubao over 180T tokens a day, June 23 · 10jqka, Jun 23
• China — National use over 500T tokens a day in June, reported without a source or method · TechNode, Aug 28
• World estimate — About 370T tokens a day, July · I/O Fund, Jul 30
• World estimate — More than 30 quadrillion tokens a month, late June · Exponential View, as relayed by CFA UK, Jul 5

Method

Tokens. OpenRouter’s rankings for October 1–18, the window of the AI-demand pack’s October level: 23.8T tokens a day across the 50 biggest models and a row for the rest. Following OpenRouter’s October 8 data, we take seven-eighths of tokens as agent traffic (87% cached input, 12% new input, 1% output) and one-eighth as human traffic (12% output, as OpenRouter reports, and by our assumption 20% cached and 68% new input). That is 79% cached, 19% new and 2.4% output. The AI-demand pack assumed 10% output and half of input cached to price tokens; its dollar total is pinned to Sacra’s revenue estimate, so we keep its dollars and use OpenRouter’s own split for compute.

Compute per token. The reference is DeepSeek V4 Pro, 1.6T parameters with 49B active, on AgentX at 100 tokens/s per user, which counts cached input in its totals. Each model’s compute per token is its active parameters over 49B. DeepSeek V4.1 Flash, which uses 8B on input and 16B on output, is weighted by where the GPU time goes. GPU time per token is 5 for output, 1 for new input and 0.05 for cached input. Taking AgentX’s traffic to be shaped like OpenRouter’s agent traffic, an average OpenRouter token takes 1.63× an AgentX token. With AgentX’s own mix, 0.5% output and 73–96% of input from cache, it would take 1.1–3.1×, putting OpenRouter at 39–116 MW on today’s mix of hardware. That moves OpenRouter’s megawatts but not the market figure, which scales the fleet. Capacity runs at 60% utilization, the figure SemiAnalysis uses for revenue per GW.

Closed models are sized from their input price: the median open model charges about $0.010 per million input tokens for each billion active parameters, and we assume closed labs keep a 60% gross margin against 30% for the hosts of open models. That puts Claude Opus 5.5 at about 221B active-equivalent parameters, a pricing proxy, not a disclosed size. Discounted versions take their model’s standard price: Muse Spark 1.3 Contributor is sized as Muse Spark 1.3. Open models’ sizes come from their model cards, except Space Bunny, GLM-5.3, GLM-5.2, MiMo-V2.5 and Nemotron 3.5 Super-VL, which are our estimates. The long tail takes the average of the rest.

Hardware and power. AgentX’s tokens per MW for each system, with InferenceX’s power per GPU from its chip pages (1.37 kW for H200, 1.71 kW for B200, 1.90 kW for B300) and its GB300 comparison (2.12 kW). InferenceX’s power is all-in, including cooling and other data-center overhead, so every megawatt and gigawatt here is all-in. A June InferenceX post used about 2.17 kW for a B200; we use the chip pages throughout. GB200 NVL72 is our estimate at 70% of GB300 per MW and 1.87 kW per GPU, from rack power of about 135 kW. Today’s fleet is our estimate of the installed base by share of power: Hopper 21%, Blackwell 8-GPU servers 16%, NVL72 racks 32%, TPUs 15%, other 16%. On AgentX that averages 9.3M tokens/s per MW, against 28.5M for GB300 and 59.4M for Vera Rubin.

The 2027 mix. Each month, every paid model’s tokens grow in proportion to its price raised to a negative power, with free tokens growing with the total. Each model’s price stays at its October level. We solve for the power and the scale that reproduce the AI-demand pack’s token and spend paths, then re-weight by compute per token. In the base case the power reaches 0.68 by December 2027: a model ten times cheaper grows about 4.8× faster.

The market. Epoch’s 31.6M H100-equivalents operating by September, at 1.25 kW each all-in (its tracked data centers: 18.5 GW at the facility for 14.8M), gives 39.5 GW; 40% serves inference (30–50%). The market’s demand grows at a set share of OpenRouter’s monthly compute growth, in logs: at 50%, a path that grows 10× becomes about 3×. New capacity needed is today’s inference fleet × (growth − 1) × today’s average tokens per MW ÷ the new hardware’s. The existing fleet stays in service. The range on today’s fleet efficiency also covers the gap between Epoch’s power basis and InferenceX’s. OpenRouter’s token shares use its average over the 30 days before each disclosure.

2026. Our sums of Epoch AI’s estimated chip sales by quarter, which reproduce its published total of 20.3M H100-equivalents at the end of 2025 and, through June 2026, the 31.6M it counts as operating by September. The second half is 10–40% above the first. All-in power is twice the chips’ rated power, Epoch’s rule of thumb for a data center’s total draw. InferenceX’s power per GPU runs 1.5–1.7× the rating for Blackwell chips; on that basis 2026 adds 19–24 GW all-in and 7–10 GW for inference. Epoch’s estimates span roughly 1.5× either way, and it counts chips when delivered, not when running.

Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at 06.ai/trackers.

Not investment advice. Estimates are labeled as estimates.

Our base case has Anthropic at about $265B and OpenAI at about $230B of annualized revenue in December 2027, up from $71B and $50B at the end of September. The ranges overlap almost completely. History says the bigger risk is that both numbers are too low.

• $72B — OpenAI, end of 2026. Range $66–79B. OpenAI expects at least $70B
• $100B — Anthropic, end of 2026. Range $88–114B. About $93B on OpenAI’s accounting
• $230B — OpenAI, end of 2027. Range $150–360B
• $265B — Anthropic, end of 2027. Range $185–400B. About $245B on OpenAI’s accounting

Rebuilt from the record, projected to 2027

The leaks agree with each other

Neither company reports its run-rate on a schedule. The figures arrive through funding announcements, investor updates and leaks, on different dates and sometimes on different definitions. We collected 145 of them. 80 give a run-rate or revenue for a specific period; dropping repeats of the same figure, later restatements and figures on a different basis leaves the 40 we fit. One smooth path through each company’s figures misses the typical Anthropic figure by 3% and the typical OpenAI figure by 1%; the largest miss is 17%. The run-rates, the quarterly revenue and Anthropic’s draft prospectus tell one story.

The OpenAI path implies about $10.1B of revenue in the third quarter, up 50% from the second. OpenAI hasn’t reported it yet.

Anthropic slowed as OpenAI sped up

The curves crossed in the summer

On the fitted path, Anthropic’s run-rate grew about 44% a month in February and 49% in March and was down to about 3% by September; only the third-quarter total pins down how August and September split. Its revenue rose 45% in the third quarter, after 145% in the second. OpenAI went the other way: about 3–4% a month in March and April, 17–19% from July to September. The Information reported that OpenAI’s run-rate grew 70% from the end of June to the end of September, and Bloomberg that enterprise sales were driving most of the growth.

What a simple rule would have said

Trend: the last 12 months’ growth continues. Fading: that growth halves over 18 months. Each rule was run at every month-end since March 2024 (January 2025 for Anthropic) and scored against every fitted run-rate and quarterly revenue figure dated up to 18 months later, as forecast ÷ outcome − 1. The table shows the median size of the miss and its usual direction.

OpenAI followed the rule; Anthropic broke it twice

The simplest forecast is that the next 12 months grow like the last 12. Run at every month-end with only the figures public at the time, it called OpenAI well: its typical miss was 8% four to seven months out and 19–24% further out, on the high side. On Anthropic it missed by about a third four to 12 months out and by 74% beyond a year, on the low side, because Anthropic kept accelerating. Then it broke the other way. Run today, it puts Anthropic at about $174B at the end of 2026, far above the roughly $100–110B in recent reporting.

The version that has growth halving over 18 months did worse than the plain rule beyond three months for both companies, because neither has slowed on a twelve-month view yet. Anthropic has slowed for a few months before: its monthly growth fell from about 40% to about 12% during 2025, then picked up again in early 2026.

OpenAI’s plans land; Anthropic’s haven’t

OpenAI hit its 2024 revenue plan exactly and its 2025 plans within 3–13%. On its own latest forecast of about $35B, it will beat the 2026 plans it had in 2025 by 17–19% (about 25% on our estimate). Only its earliest pitch was far off: in late 2022 it told investors it would make $1B in 2024, and it made $3.7B. Anthropic’s projections a year out have been too low, mostly by a lot: its 2025 base case was less than half what it booked, and in October 2025 its internal base case for its 2026 run-rate was $20B. Its short-range calls held up: it projected $9B for the end of 2025 and got $9B, and its second-quarter projection was within 6%. Its one miss on the high side was its first: a $200M pace by the end of 2023, against $87M at the start of 2024.

OpenRouter tracks Anthropic, with big swings in share

A read on Anthropic, not on OpenAI

Spend on Anthropic’s models through OpenRouter has run at 1.5–2.1% of its revenue each quarter since the end of 2025. Scaling each quarter’s OpenRouter spend by the previous quarter’s ratio would have called Anthropic’s revenue within 12–17% in four of the last five quarters and been 29% off in the fifth (the two 2025 quarters are our estimates from annual revenue). It can’t do the same for OpenAI, whose revenue comes mostly from ChatGPT and from businesses buying directly, which OpenRouter doesn’t see.

The swing factor is Anthropic’s share. It held at 54–62% of OpenRouter’s dollars from January to July, fell to 42% in August as its own spend there dropped, and to 28% in September as OpenAI’s GPT-5.6 and GPT-6 models took share. It recovered to 37% in October after Anthropic released Opus 5.5 and Sonnet 5.5 in late September and Haiku 5.5 in early October. Opus 5.5 alone is drawing more spend than Opus 5 ever did, at lower prices: $4 and $20 per million input and output tokens, against $5 and $25. Revenue moves less than this share does: in August and September, Anthropic’s spend through OpenRouter ran about 30% below July’s while its run-rate kept rising.

The AI-demand pack’s base case has spend through OpenRouter growing 4.2× from October to December 2027. If Anthropic held its October share, that would put its run-rate near $350B; if its share settled at 30%, about $285B; if it kept falling at this year’s pace, about $180B. Our base case of $265B implies a share near 28%, about where it was at its September low.

How fast growth fades decides 2027

Run-rate in December 2027 if each company’s growth rate entering 2027 in our base case (about 13% a month for OpenAI and 11% for Anthropic) halves at different speeds.

The forecast

End of 2026. OpenAI expects to reach or exceed a $70B run-rate. It beat its end-2025 target by 7%, and the trend rule says $70B. We use $72B. For Anthropic, October’s spend through OpenRouter implies a run-rate of about $84B at this year’s average quarterly ratio (1.8%), and about $80B if October behaves like June, the last month with a launch-driven jump of the same size. We take $83B. Growth of about 10% a month from there reaches $100B in December, with about $22B of fourth-quarter revenue. That is at the top of what the FT reported after Anthropic’s investor day, an annualized figure below $100B, and short of the $110B analysts said was possible.

2027. For OpenAI we assume its growth rate in December 2026, about 13% a month, halves every 18 months. That gives about $230B in December 2027. For Anthropic we take the midpoint of two methods. The same rule, applied to its growth of about 10% a month late in 2026, gives $245B. OpenRouter gives $285B if Anthropic’s share of spend there settles at 30% while the market follows the AI-demand pack’s base case. The midpoint, $265B, is what you get if Anthropic enters 2027 growing about 11% a month and that rate halves every 18 months. The ranges come from varying the year-end level, the starting growth rate and the fade, from halving every 6 months to every 4½ years. On OpenAI’s accounting Anthropic would be about $245B, a little ahead of OpenAI.

For OpenAI, history leans to the upside: a rule that had its growth halving over 18 months ran 23–35% low eight to 18 months out. For Anthropic it cuts both ways: rules built on its trailing growth were far too low until mid-2026 and are far too high now. Scale argues for caution. Each company would add roughly $155–165B of run-rate in 2027, more than either has added in a year so far, and both depend on compute arriving on schedule. Anthropic’s own IPO projection, $190–200B of revenue in 2028, implies about $140B at the end of 2027 if its growth fades as in our base case. It has beaten every longer-range projection it has made since the start of 2025.

What to watch

Anthropic on OpenRouter in November. If spend holds near the $4.5–5M a day it has averaged since October 6, the October jump was real demand and $100B by December looks low. If it falls back by a third or more, as it did in late July, $85–90B is more likely.

OpenAI’s third-quarter revenue. Our fitted path says about $10.1B. Much less would mean the run-rate figures ran ahead of revenue.

Anthropic’s prospectus. If Anthropic goes ahead with its IPO, the public filing should show quarterly revenue on one definition, next to the run-rates it has been quoting.

What this is anchored on

Run-rates and revenue are as reported; fitted paths, back-tests, OpenRouter spend and forecasts are our estimates. All 40 figures used in the fit are listed below with their sources.

• OpenAI — About $50B run-rate at the end of September on OpenAI’s own (net) basis; the ~$70B figure came from investors applying Anthropic’s gross method; expects to reach or exceed $70B by the end of 2026 · Bloomberg, via Bloomberg Línea, Oct 9
• OpenAI — About $50B at the end of September, less than the ~$70B previously indicated to investors · Reuters, citing the FT, via The Standard, Oct 8
• OpenAI — Run-rate grew 70% from the end of June to the end of September (The Information, citing the CFO); enterprise sales driving most of the growth (Bloomberg) · The Next Web, Oct 9
• OpenAI — Full-year 2026 revenue of around $35B · Bloomberg, via AI Weekly
• OpenAI — Annual recurring revenue “nearing $70 billion” · Axios, Sep 29
• OpenAI — More than $40B run-rate in mid-August · Bloomberg, via investingLive, Aug 13
• OpenAI — $2B of revenue a month (March 2026) · OpenAI, Mar 31
• OpenAI — $21.4B run-rate at the end of 2025; more than $25B at the end of February · The Information, via Techloy, Mar 5
• OpenAI — Revenue $5.7B in Q1 2026 · The Information, via The Next Web, Jun 17
• OpenAI — Revenue $6.7B in Q2 2026 · Wall Street Journal, via AI Weekly, Aug 18
• OpenAI — Audited revenue: $3.7B in 2024, $13.07B in 2025 · Where’s Your Ed At (verified by the FT), Jun 15
• OpenAI — Projected revenue of $12.7B for 2025 and $29.4B for 2026 (March 2025) · Bloomberg, via Advisor Perspectives
• Anthropic — Revenue $16.8B in Q3 2026, up about 45% from Q2 · Financial Times, via The Next Web, Oct 14
• Anthropic — Revenue $4.7B in Q1 2026 and $11.6B in Q2 · Reuters, via BusinessDay, Oct 15
• Anthropic — Run-rate of $65B at the end of July · Bloomberg, via Bloomberg Government, Aug 17
• Anthropic — Run-rate of $14B · Anthropic, Series G announcement, Feb 12
• Anthropic — Run-rate above $30B, up from about $9B at the end of 2025 · Anthropic, Apr 6
• Anthropic — Run-rate crossed $47B in early May · Anthropic, Series H announcement, May 28
• Anthropic — Revenue of about $386M in 2024 and $4.59B in 2025 (draft IPO prospectus) · Reuters, via TradingKey, Sep 29
• Anthropic — Revenue of nearly $4.6B in 2025; cloud-marketplace sales booked in full, with the partners’ cut as an expense · Reuters, via Khaleej Times, Sep 30
• Anthropic — Internal base case of $20B of annualized revenue for 2026, best case $26B (October 2025) · Reuters, via Yahoo Finance, Oct 15, 2025
• Anthropic — Projected 2026 revenue of $18B (January 2026) · The Information, via Investing.com, Jan 27
• Anthropic — 2028 revenue of roughly $190–200B underpinning the IPO valuation · Reuters, via Yahoo Finance, Aug 14
• Anthropic — Annualized revenue expected to stay below $100B · Financial Times, via Yahoo Finance, Oct 15
• Anthropic — Analysts who met executives: run-rate “could reach $110 billion by the end of the year” · Reuters, via Calcalist, Oct 15
• Anthropic — About $74B annualized in October (third-party estimate) · SemiAnalysis, via Implicator, Oct 15
• OpenRouter — Daily tokens by model, Jan 1, 2025–Oct 18, 2026; spend is our estimate at catalog prices (see the AI-demand pack) · OpenRouter rankings. Licensed under CC BY 4.0
• OpenRouter — List prices, including Opus 5.5 at $4 and $20 per million input and output tokens and Opus 5 at $5 and $25 · OpenRouter models API, Oct 19

Method

Figures. We collected 145 reported figures on the two companies’ revenue, run-rates, projections and users, with a link for each. Of these, 80 give a run-rate or revenue for a period; we fit 40, dropping repeats of the same figure, later restatements and figures on a different basis. Third-party estimates are shown but not fitted. OpenAI’s 2024 and 2025 full-year revenue includes items its run-rate figures leave out, so we don’t fit it. Each figure is placed on the date it describes; the back-test uses the date it was first reported.

Fitted paths. Each company’s log run-rate is a piecewise-linear curve with knots every half month and a penalty on curvature. Run-rates are fitted as points; quarterly and annual revenue as the integral of run-rate ÷ 12 over the period, weighted more heavily because they are booked numbers. Monthly growth in the second chart is month-end to month-end on these paths.

Back-test. At each month-end from March 2024 (January 2025 for Anthropic) to September 2026, a rule sees only figures reported by then. Its level is the latest figure (quarterly revenue counts as four times itself at mid-quarter) and its growth is the average over the previous 12 months, held constant or halving over n months. It is scored against every fitted run-rate and quarterly revenue figure dated up to 18 months later, as forecast ÷ outcome − 1.

Plans compare projections with what was later reported. Two were made partway through the period they cover: September 2024 for 2024 revenue and May 2026 for the second quarter. 2026 outcomes are our estimates.

OpenRouter. Spend on each lab’s models is estimated as in the AI-demand pack, at catalog prices. The share is Anthropic’s spend over all spend. The scenarios scale the AI-demand pack’s base path for total spend by Anthropic’s share, moving geometrically from October’s level to the end value, and convert spend to run-rate at October’s ratio.

Forecast. End-2026: OpenAI from its own target; Anthropic from an October run-rate of about $83B, between a launch-month reading of OpenRouter and its quarterly average ratio, then about 10% a month. 2027: growth halves every 18 months, g(t) = g₀ × 2^−t/18. For OpenAI g₀ is its December 2026 growth; for Anthropic it is set so December 2027 lands midway between that rule applied to its late-2026 growth and the OpenRouter lens with a 30% share. The 80% ranges simulate the year-end level (±7% for OpenAI and ±10% for Anthropic at one standard deviation), the starting growth (±25%) and a half-life between 6 and 54 months. Run-rates are each company’s own: Anthropic books sales through cloud marketplaces in full and pays the partners’ cut as an expense, about 16% of those sales, which were 47% of its 2025 revenue; we put Anthropic about 7.5% lower on OpenAI’s net method.

Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at 06.ai/trackers.

Not investment advice. Estimates are labeled as estimates.

OpenRouter routed 23.8T tokens a day in October, 33× as many as a year earlier. We estimate its customers are spending about $4.2B a year on them, 11× January’s pace. Tokens are growing faster than dollars, Chinese models supply most of the tokens while US labs collect most of the dollars, and most tokens go to code and agent work.

• 23.8T — Tokens a day, October. 33× October 2025. 2.2 quadrillion so far this year, against OpenRouter’s May forecast of more than 1 quadrillion for all of 2026
• $4.2B — Annualized spend, October. Our estimate of what customers spend through OpenRouter, 11× January. About $230M a year of fees for OpenRouter at 5.5%
• $0.48 — Average price, million tokens. October, against $0.96 in the first half, as cheaper models took share
• 73K — Tokens per request, top 100 apps. 5.4× January. Agents re-send their working context at every step

Tokens: 33× in a year

What OpenRouter can see

OpenRouter sells access to more than 400 models through one API and publishes the daily token count of the 50 biggest, plus a single row for everything else. It sees developers, startups and agent tools. It doesn’t see ChatGPT’s subscribers or companies that call OpenAI and Anthropic directly, and its public counts leave out private and zero-data-retention traffic. So it is a slice: about $4.2B a year of spend, against roughly $120–140B of annualized revenue at OpenAI and Anthropic combined, depending on how OpenAI’s revenue is counted. But it is one of the few public daily counts of AI demand, and it has been growing faster than Google, which gives a figure from time to time. Google said in May it was processing 3.2 quadrillion tokens a month, 7× a year earlier. OpenRouter grew 12× over the same twelve months.

Tokens up 23× this year, dollars 11×

Since January, daily tokens are up 23× and our estimate of spend 11×. The difference is price. The average million tokens cost $0.96 in the first half of the year and $0.48 in October. We price every model at today’s list price, so the whole fall is mix: cheap models, above all from DeepSeek and Z.ai, took share. At the first half’s average price, October’s tokens would be worth $8.3B a year, not $4.2B.

Dollars are also where growth has started to slow. Spend grew 30% a month from April to July and 21% a month from July to October. Tokens went the other way, from 35% to 48%.

Dollars are growing more slowly than tokens

Chinese models supply the tokens; US labs collect the dollars

China’s tokens, America’s dollars

Models from Chinese labs first passed US models in weekly tokens in the week of March 23 and have led every week since April 27. Over the past four weeks they were 57% of tokens (64% last week) and 26% of estimated spend. DeepSeek is the biggest lab by tokens, with 26%, and takes 4% of the dollars. Anthropic has 5% of the tokens and 35% of the dollars; OpenAI 9% and 26%. Between them, the two take 61% of the dollars from 14% of the tokens.

Much of what is spent on Chinese models doesn’t reach Chinese labs either. Their models are open-weight, so anyone can host them: 28 providers serve the September version of DeepSeek V4.1-Flash, its most-used model, and DeepSeek isn’t one of them. On a snapshot of where requests go, DeepSeek serves 9% of the spend on its own models, Z.ai 6% and Moonshot 10%. The rest goes to hosts such as Together, DeepInfra and Baseten. Tencent, Xiaomi and MiniMax keep most of theirs.

Those hosts often charge more than the single price OpenRouter lists for a model, which is the one we use. Weighting each host’s price by its share of requests in the same snapshot would add about 17% to October’s spend, most of it on Chinese models, and lift their share of the dollars to about 37%. We leave that out of our figures because a 30-minute snapshot can’t be applied to earlier months.

Agents are the demand

The average request from OpenRouter’s 100 biggest apps carried 73,000 tokens in October, against 13,000 in January. Coding agents re-send their working context at every step, so one task can run to millions of tokens. Hermes Agent, Claude Code, Kilo Code and Cline are the four biggest apps, and OpenRouter classifies 63% of last week’s tokens as code or agent work. Roleplay, about a third of tokens in mid-2025, is down to about 5%.

Agents also make dollars harder to measure. Much of what they send is context they have sent before, which providers bill at cached rates, often a tenth of the input price or less. The cached share is the biggest unknown in our spend estimate; matching Sacra’s July revenue figure implies about 51% of input.

Free tokens are a launch strategy

Labs buy demand on OpenRouter with free models. Free versions and stealth models, released under code names before launch, were 15% of tokens over the past four weeks and peaked at 31% in April 2025. Z.ai ran GLM-5.3-Flash as “Ox Alpha” for a week in August; on August 24, OpenRouter said it was on track for nearly 6 trillion tokens that day. A model called Space Bunny Alpha has done 57 trillion tokens in the past 30 days without its maker being named. We count all of it at $0, which is why tokens and dollars can diverge sharply within a month.

Requests got five times bigger

Spend on Anthropic’s models through OpenRouter has tracked its revenue

A read on Anthropic

Spend on Anthropic’s models through OpenRouter was 1.8% of its reported revenue in the first quarter, 2.1% in the second and 1.7% in the third. In October so far it is running 36% above the third quarter’s pace. If OpenRouter’s share stays in that range, Anthropic is on course for $19–23B of revenue in the fourth quarter, against $16.8B in the third. That is the quarter in which it plans to list; bankers are reported to be pricing the IPO on $190–200B of revenue in 2028. Anthropic released Opus 5.5 and Sonnet 5.5 in late September and Haiku 5.5 in October, and launch weeks can pull spend forward.

OpenAI is different. Spend on its models through OpenRouter grew 4× from the first quarter to the second and 13× by the third, while its reported revenue grew 18% from the first quarter to the second. OpenRouter is not a read on OpenAI.

Three paths for 2027

At the default setting (Growth halves every 6 months), from the interactive version:

• Spend in December 2027, a year: $18B
• Multiple of October: 4.2×
• OpenRouter fees at 5.5%: $1.0B
• Monthly growth by then: 4.7%

The forecast

Spend growth on OpenRouter slowed from 30% a month to 21% between the last two quarters. That is what a growth rate halving about every 6 months looks like, and it is our base case. On it, spend reaches about $18B a year by December 2027, 4.2× today’s, and is still growing about 5% a month. If growth halves every three months, spend levels off near $10B; if it halves only once a year, it reaches about $29B. Even the low case is 2.5× today’s. The high case needs another wave like 2026’s, with agents spreading from coding into other work.

The token paths are steeper because tokens have been growing faster than dollars. In the base case OpenRouter handles about 513T tokens a day by the end of 2027 at about $0.09 per million. That only happens if cheap models keep taking share.

The three paths barely differ over the next quarter: spend grows 60% to 72% from October to January on all of them. They separate during 2027, so the early quarters won’t settle which one we’re on.

What it means for the earlier packs

The funding-scale pack adds up to about $1.54T of debt and equity that the buildout needs through 2030. On OpenRouter, demand is not the constraint: spend is growing about 21% a month and tokens about 48%.

What changes is what a dollar buys. A dollar spent on OpenRouter buys 2.0× as many tokens as it did in the first half, because the mix has moved to cheaper models. This data can’t say how much of that is cheaper serving and how much is thinner margins; the models taking share are smaller, so some of it is efficiency. The token-price pack found the same pattern in list prices: the price of a top-tier token has held while everything below it falls. It is why the labs’ share of the dollars matters more than their share of the tokens: Anthropic and OpenAI take 61% of the dollars from 14% of the tokens.

What to watch

The dollar growth rate in the spring. Average monthly growth from December to March of about 11% would put spend on the low path, 15% on the base path and 18% on the high one. Single months are noisy: spend fell 2% in August and rose 47% in September.

Anthropic’s share. If OpenRouter stays near 1.7–2.1% of Anthropic’s revenue, it is a weekly read on Anthropic’s growth ahead of its IPO.

Free and stealth tokens. When they spike, token counts overstate demand.

Who hosts the open models. DeepSeek serves 64% of requests for its new V4.1-Flash-1017 itself, against none of the September version. Chinese labs taking back hosting would move dollars back to them.

What this is anchored on

Source: OpenRouter (openrouter.ai/rankings), as of 2026-10-19. Licensed under CC BY 4.0. Spend, prices per token and forecasts are our estimates.

• Tokens — Daily tokens for the 50 biggest models plus one row for the rest, Jan 1, 2025–Oct 18, 2026; prompt plus completion, counted with each provider’s tokenizer; public traffic only · OpenRouter rankings-daily dataset, as of 2026-10-19 01:29 UTC. Licensed under CC BY 4.0
• Prices — List price of every model, and of every host serving it, with each host’s requests over the 30 minutes to 01:30 UTC, Oct 19 · OpenRouter models and endpoints APIs
• Use and apps — Weekly token estimates by category (sampled); tokens and requests of the 100 biggest apps each month; 7-day task classification · OpenRouter Data API. Licensed under CC BY 4.0
• OpenRouter revenue — About $19M annualized at the end of 2025 and $50M in March 2026 · Sacra, Jun 1
• OpenRouter revenue — $140M annualized in July 2026 and $160M in August (estimates; OpenRouter doesn’t disclose revenue) · Sacra
• OpenRouter fee — 5.5% on pay-as-you-go credit purchases · OpenRouter
• OpenRouter volume — 25T tokens a week at the May 26 Series B · Pulse 2.0, May 26
• OpenRouter volume — Forecast at the Series B of more than 1 quadrillion tokens for 2026 · Implicator, Jul 24
• Google — More than 3.2 quadrillion tokens a month across its products, 7× a year earlier · Google, May 19
• Anthropic — Revenue $4.73B in Q1 2026 and more than $11.5B in Q2 (preliminary) · Bloomberg, via The Next Web, Aug 15
• Anthropic — Revenue $16.8B in Q3 2026 · AFP, via Khaleej Times, Oct 15
• Anthropic — IPO targeted for mid-to-late November; bankers reported to be pricing it on ~$190–200B of 2028 revenue · Bloomberg, Oct 1; Reuters, Aug 14
• OpenAI — Revenue $5.7B in Q1 2026 and $6.7B in Q2 · The Information, via The Next Web; WSJ
• Ox Alpha — Stealth model revealed as Z.ai’s GLM-5.3-Flash on Aug 26; OpenRouter said it was on track for nearly 6T tokens on Aug 24 · Digital Applied; Pasquale Pillitteri

Method

Tokens come from OpenRouter’s rankings-daily dataset: the 50 biggest models each UTC day plus one row for the rest, prompt plus completion tokens counted with each provider’s tokenizer. It covers public traffic only. Weeks run Monday to Sunday. June 15 and July 15, 2025 are missing from the dataset and are filled with the average of the days either side. Region is the developer’s home country; Ox Alpha is counted as Z.ai’s, and developers we couldn’t place are in Other.

Spend prices each model’s tokens at its standard catalog price; variants such as batch are priced as themselves, and free versions and stealth models at $0. We assume 10% of tokens are output and 51% of input is billed at cached rates, the share that reproduces Sacra’s $140M estimate of OpenRouter’s July revenue at its 5.5% fee. The level of spend depends on that assumption; growth rates and shares much less. The catalog price is often the cheapest host’s: weighting hosts’ prices by their share of requests in a 30-minute snapshot on October 19 would add about 17% to October’s spend, mostly on Chinese models. We leave it out because it can’t be applied to earlier months. Before May 2026, 10–27% of tokens came from models no longer in the price list; they are priced at the median for their lab.

The gray line prices every week’s tokens at the first half’s average of $0.96 per million.

Anthropic and OpenAI compare estimated spend on each lab’s models through OpenRouter, by calendar quarter, with reported revenue. Q4 scales October 1–18 to a full quarter.

Apps divide total tokens by total requests across the 100 biggest apps in each calendar month. Categories are OpenRouter’s sampled weekly estimates, which can overlap.

The forecast starts from the October 1–18 level and the monthly growth between July’s average and October’s, 2.8 months apart: 20.9% for dollars and 47.7% for tokens. That rate then halves every n months: growth in month m is g × 2^−(m−1)/n. The half-life is the main assumption; the slowdown in dollar growth between the last two quarters implies 5.7 months. It inherits the spend method’s assumptions, and the token paths assume the mix keeps moving to cheaper models.

Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at 06.ai/trackers.

Not investment advice. Estimates are labeled as estimates.

DeepSeek’s V4.1-Flash-1017 costs $0.15 per million input tokens and $0.60 per million output tokens in US business hours, about what OpenAI charges for its cheapest model. On DeepSWE, DeepSeek’s headline coding benchmark, it scores above GPT-5.6 Sol and Claude Opus 5, which cost 33× to 42× more per output token.

• $0.60 — DeepSeek V4.1-Flash-1017, output. Per million tokens in US hours; input $0.15; $0.30 / $1.20 at peak
• 33× — Cheaper than GPT-5.6 Sol. On output; 27× on input and 133× on cached input, at Sol’s discounted rate
• 77.0 — DeepSWE v1.1, DeepSeek’s figures. Claude Opus 5 74.0, GPT-5.6 Sol 73.0
• 16B — Active parameters. Of 552B, when writing output; 8B when reading input

The price ladder

The gap narrowed after R1, then widened again

What DeepSeek says you get for the price

DeepSeek’s published table, 10 of the 11 benchmarks it reported. Bold marks the best score in each row. It beats GPT-5.6 Sol on 9 of 10 and Claude Opus 5 on 4 of 10; Opus 5 leads on the harder long-horizon tests. Prices are input / output per million tokens.

• DeepSWE v1.1 — DeepSeek V4.1-Flash-1017: 77.0 · DeepSeek V4.1-Flash (Sep): 74.2 · Kimi K3: 67.5 · GLM 5.3: 66.9 · GPT-5.6 Sol: 73.0 · Claude Opus 5: 74.0
• Terminal-Bench 4.0 — DeepSeek V4.1-Flash-1017: 42.3 · DeepSeek V4.1-Flash (Sep): 31.2 · Kimi K3: 12.6 · GLM 5.3: 37.9 · GPT-5.6 Sol: 39.9 · Claude Opus 5: 51.8
• ProgramBench — DeepSeek V4.1-Flash-1017: 29.1 · DeepSeek V4.1-Flash (Sep): 20.3 · Kimi K3: 17.5 · GLM 5.3: 19.0 · GPT-5.6 Sol: 23.0 · Claude Opus 5: 37.0
• NL2Repo-Bench — DeepSeek V4.1-Flash-1017: 68.2 · DeepSeek V4.1-Flash (Sep): 65.4 · Kimi K3: 58.0 · GLM 5.3: 58.0 · GPT-5.6 Sol: 56.8 · Claude Opus 5: 75.3
• Automation-Bench v1.0.6 — DeepSeek V4.1-Flash-1017: 60.5 · DeepSeek V4.1-Flash (Sep): 54.8 · Kimi K3: 46.7 · GLM 5.3: 48.8 · GPT-5.6 Sol: 45.8 · Claude Opus 5: 50.3
• Agents’ Last Exam — DeepSeek V4.1-Flash-1017: 35.4 · DeepSeek V4.1-Flash (Sep): 31.8 · Kimi K3: 27.6 · GLM 5.3: 28.5 · GPT-5.6 Sol: 26.7 · Claude Opus 5: 28.6
• Apex Bench v1.1 — DeepSeek V4.1-Flash-1017: 72.1 · DeepSeek V4.1-Flash (Sep): 64.6 · Kimi K3: 47.8 · GLM 5.3: 56.6 · GPT-5.6 Sol: 54.3 · Claude Opus 5: 65.8
• HLE, no tools — DeepSeek V4.1-Flash-1017: 42.6 · DeepSeek V4.1-Flash (Sep): 36.8 · Kimi K3: 43.5 · GLM 5.3: 42.0 · GPT-5.6 Sol: 44.5 · Claude Opus 5: 56.3
• Chartography, with tools — DeepSeek V4.1-Flash-1017: 83.0 · DeepSeek V4.1-Flash (Sep): 78.9 · Kimi K3: 68.1 · GPT-5.6 Sol: 79.9 · Claude Opus 5: 84.0
• BabyVision — DeepSeek V4.1-Flash-1017: 92.5 · DeepSeek V4.1-Flash (Sep): 89.6 · Kimi K3: 85.7 · GPT-5.6 Sol: 88.9 · Claude Opus 5: 94.1

The model

V4.1-Flash-1017 is a new version of the V4.1-Flash model DeepSeek released in September, with its weights posted under the MIT license, so anyone can host it. It is large, at 552B parameters plus 196B in a separate memory layer, but sparse: only 8B parameters are active when it reads and 16B when it writes, which is why it is cheap to serve. DeepSeek says it beats September’s version on all 11 benchmarks it reports and gives shorter answers. Its table shows it ahead of GPT-5.6 Sol on 9 of 10 benchmarks and of Claude Opus 5 on 4. Opus 5 still leads on the harder long-horizon tests, Terminal-Bench (51.8 against 42.3), NL2Repo (75.3 against 68.2) and ProgramBench (37.0 against 29.1), and on HLE (56.3 against 42.6). These are DeepSeek’s numbers, several from benchmarks few others report, run on its own agent harness. Independent evaluations aren’t out yet.

Priced like a budget model

US business hours fall in DeepSeek’s off-peak window, when it charges $0.15 per million input tokens and $0.60 per million output tokens. That is 27× less than GPT-5.6 Sol on input and 33× less on output, even at Sol’s discounted $4/$20, and 133× less on cached input. Cached input matters most for agents, which re-read the same context at every step. Against OpenAI’s cheapest model, GPT-5.6 Luna at $0.20/$1.20, it costs about the same. DeepSeek is selling a model it benchmarks against Sol and Opus at Luna’s price.

What earlier DeepSeek releases did

V2, in May 2024, started a price war in China: Alibaba cut a Qwen model 97%, Baidu made two Ernie models free and Zhipu cut 80%. US prices didn’t move. R1, in January 2025, roughly matched OpenAI’s o1 at $0.55/$2.19 against $15/$60, and Nvidia lost about $589B of market value on January 27. The biggest US cuts came over the following year. OpenAI released the cheaper o3-mini four days after the selloff and cut o3 by 80%, to $2/$8, in June. GPT-5 launched in August at $1.25/$10, and Anthropic took Opus from $15/$75 to $5/$25 in November. By mid-2025 the gap between DeepSeek and OpenAI’s flagship was the narrowest it has been.

2026 went the other way. GPT-5.5 launched the same week DeepSeek previewed V4, at $5/$30, double GPT-5.4. DeepSeek made a 75% cut to V4-Pro permanent in May, and US flagships didn’t follow. US labs cut later and selectively: OpenAI cut its cheapest model, Luna, by 80% in August and put Sol on a three-month discount, and Anthropic priced Opus 5.5 20% below Opus 5. None came near DeepSeek’s prices. DeepSeek itself raised prices in August and added peak-hour rates, a sign it couldn’t serve all the demand at its old prices.

The pattern

Each DeepSeek release resets the price of “good enough”. US labs answer by repricing the tier that competes and by temporary discounts, while holding flagship list prices around $4–5 for input and $20–30 for output. The price of a given level of capability falls very fast (Epoch AI measured 9× to 900× a year, depending on the benchmark), but the price of a top-tier token hasn’t. Two things limit DeepSeek’s direct pull: its own serving capacity, and the US companies and government users that won’t send data to a Chinese API. The pressure arrives through US companies hosting the open weights.

What this one might do

Mid-tier prices go first. If independent tests confirm Sol-level agentic scores at Luna-level prices, GPT-5.6 Terra ($2/$12) and Anthropic’s Sonnet-class models are the most exposed. Expect cuts and discounts there, and more bundling into subscriptions such as Codex and Claude Code, where per-token prices are harder to compare.

Cached input gets cheaper. It costs the labs little to serve, and most of what an agent reads is cached.

Flagship prices are the tell. Sol’s discount runs at least to November 21. Extending it or making it permanent would be the DeepSeek effect; going back to $5/$30 would mean OpenAI still has pricing power. Google’s plan to double Gemini Flash prices on January 1 is the same test.

DeepSeek is also going after agent developers. It now sells web search and page fetching to agents at $0.75 per 1,000 calls, free until November 1 when the calls come through Claude Code, Codex or DeepSeek’s own agent tool.

What it means for the lab-funding and funding-scale packs

OpenAI’s plan needs revenue to grow from $36B this year to $350B in 2030, and Anthropic’s IPO is being priced on $190–200B of revenue in 2028. Both assume usage grows much faster than prices fall. A model that caps mid-tier prices pushes more of that revenue onto the top tier, where DeepSeek still trails. Each 10% that OpenAI’s revenue falls short of plan adds about $84B to the $116B it still has to raise; try it below. For compute, cheaper tokens have so far meant more demand, not less: DeepSeek is rationing capacity with peak pricing, and V4’s launch barely moved US markets. The risk sits with the labs’ margins and the backlogs that depend on them, such as Oracle’s, Microsoft’s and CoreWeave’s, more than with the chipmakers.

If token prices take OpenAI’s revenue below plan

OpenAI’s July 2026 plan from the lab-funding pack: $840B of revenue and $278B of cash burn over 2026–30, with $162B in hand. Assumes costs don’t fall with revenue, since much of its compute is contracted.

At the default setting (Revenue below plan 0%), from the interactive version:

• Cash burn, 2026–30: $278B
• Still to raise: $116B
• Total need in the funding-scale pack: $1.54T

What this is anchored on

List prices per million tokens as of October 17, 2026, standard rates for short prompts.

• V4.1-Flash-1017 — Released Oct 17 as a new version of September’s V4.1-Flash; 552B parameters plus 196B ‘Engram’ parameters, 8B active on input and 16B on output; 1M-token context; $0.30/$1.20 at peak and half off-peak; beats the September version on all 11 benchmarks reported · MarkTechPost, Oct 17
• V4.1-Flash-1017 — Weights on Hugging Face under the MIT license; Search and Fetch APIs at $0.75 per 1,000 requests, free via DeepSeek Harness, Claude Code or Codex until Nov 1 · AI Weekly, Oct 17
• V4.1-Flash-1017 — $0.15 input, $0.60 output, $0.003 cached input per million tokens · OpenRouter
• DeepSeek rates — Peak hours 01:00–04:00 and 06:00–10:00 UTC on weekdays; V4.1-Flash $0.30/$1.20 peak and $0.15/$0.60 off-peak; V4-Pro $1.32/$3.96 at peak · Yotta Labs
• GPT-5.6 Sol — $4 input, $0.40 cached input, $20 output; discounted pricing available at least through Nov 21, 2026 · OpenAI
• GPT-5.6 Sol — Cut from $5/$30 to $4/$20 for three months on Aug 21; Claude Opus 5 at $5/$25 and Claude Fable 5 at $10/$50 · Let’s Data Science, Aug 22
• GPT-5.6 family — Terra $2/$12; Luna $0.20/$1.20; general availability Jul 9, 2026 · CostBench
• Claude Opus 5.5 — Released Sep 22, 2026 at $4/$20, against $5/$25 for Opus 5 · Metric Nexus
• August 2026 — Luna cut 80% to $0.20/$1.20; Gemini 3.7 Flash launched at $0.75/$3.75, set to double on Jan 1; DeepSeek V4-Flash moved from $0.14/$0.28 to $0.22/$0.66 off-peak and $0.44/$1.32 at peak · DevTk.AI
• DeepSeek V4-Pro — Output price raised from $0.87 to $3.96 at peak and $1.98 off-peak on Aug 16; Nvidia lost about $589B of market value on Jan 27, 2025 · Regards of Wall Street
• DeepSeek V4-Pro — Launched about Apr 24, 2026 (1.6T parameters, 49B active, open weights) at $3.48 output; 75% discount · The Next Web, Apr 27
• DeepSeek V4-Pro — 75% cut made permanent May 23; GPT-5.5 at $5/$30 and Claude Opus 4.7 at $5/$25; no US price response · TheStreet, May 25
• DeepSeek V4 — Muted initial market reaction to the Apr 24 preview · Let’s Data Science
• GPT-5.5 — Launched at $5/$30, double GPT-5.4’s $2.50/$15 · AI Pricing Guru, Apr 24
• Opus 4.5 — Released Nov 24, 2025 at $5/$25, against $15/$75 for Opus 4 and 4.1 · ClaudeFast
• GPT-5 — Launched at $1.25/$10, against $15/$75 for Claude Opus 4.1 · TechCrunch, Aug 8, 2025
• DeepSeek V3.2-Exp — API prices cut by more than half on Sep 29, 2025, to $0.28 input and $0.42 output · GIGAZINE, Sep 30, 2025
• o3 — Cut 80% on Jun 10, 2025, from $10/$40 to $2/$8 · VentureBeat
• R1 vs o1 — DeepSeek-R1 at $0.55/$2.19 (Jan 20, 2025) against o1 at $15/$60 · AnotherWrapper, citing both companies’ price pages
• o3-mini — OpenAI released o3-mini on Jan 31, 2025, in the wake of DeepSeek · BNN Bloomberg
• China, 2024 — DeepSeek-V2 at ¥1–2 per million tokens (May 6, 2024); Alibaba cut Qwen-Long 97%, Baidu made Ernie Speed and Lite free, Zhipu cut GLM-3-Turbo 80% · KrASIA, May 30, 2024
• Price trends — The price of reaching a fixed level of performance fell 9× to 900× a year, depending on the benchmark · Epoch AI, Mar 2025

Method

Prices are list prices per million tokens on October 17, 2026, at standard rates for prompts under each provider’s long-context threshold. DeepSeek charges its peak rate, double the off-peak rate, from 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays; US business hours fall outside both windows. GPT-5.6 Sol is shown at its discounted rate, available at least through November 21, with its list price as the hollow dot. Gemini 3.7 Flash is shown at its launch price, which is set to double on January 1.

Benchmarks are DeepSeek’s own table. Kimi K3 and GLM 5.3 have no prices here because their pay-as-you-go API prices weren’t published in the sources used.

The price history uses launch list prices of each lab’s top model: OpenAI’s o1, o3, GPT-5, GPT-5.5 and GPT-5.6 Sol; Anthropic’s Opus line; DeepSeek’s R1, V3.2-Exp, V4-Pro and V4.1-Flash-1017, at peak rates where DeepSeek has them. It is a picture of list prices, not of what any customer paid after discounts.

The funding sensitivity scales the lab-funding pack’s plan: a revenue shortfall adds to cash burn one for one, with no cost savings. The funding-scale pack’s total adds the same amount to its five-year figure.

Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at 06.ai/trackers.

Not investment advice. Estimates are labeled as estimates.

Thirteen companies need about $1.54T of debt or equity over 2026–30, $435B of it in 2028. Next to Treasury borrowing and mortgages, that’s small. Next to the new money US businesses borrow each year, and the markets the weakest of these borrowers use, it isn’t.

• $1.54T — Total need, 2026–30. Debt or equity · $1.26T for infrastructure, $278B for the labs
• $435B — Peak year. 2028 · 1.3% of GDP
• $438B — Hyperscalers. 28% of the total · peak $155B in 2027
• $476B — Builders. 31% · Oracle $65B, SpaceX $411B
• $347B — Neoclouds. 23% · plus $75B of GPU loans to roll
• $278B — Labs. 18% · all OpenAI; Anthropic is cash-flow positive

Debt or equity needed, by year

Oracle is converted from fiscal years ending May. Neocloud figures are new money only: another $75B of GPU-backed loans comes due over the period and has to be refinanced. The labs’ figure is OpenAI’s own plan; on Anthropic’s own figures it needs no outside money.

Against government, mortgages and credit markets

Each group against the market it borrows in

The total hides where the money has to come from. Each group’s peak-year need against a year of issuance in the market it mostly draws on. For the neoclouds, the peak includes GPU-backed loans that have to be refinanced that year.

Hyperscalers · High grade

6.8% · Peak-year need as a share of a year of US investment-grade bonds

• Need, 2026–30: $438B
• Peak year: $155B in 2027
• That market: $2.27T a year
• Raises through: Investment-grade bonds, equity

Builders · BBB / BBB−

4.9% · Peak-year need as a share of a year of US investment-grade bonds

• Need, 2026–30: $476B
• Peak year: $110B in 2028
• That market: $2.27T a year
• Raises through: Investment-grade bonds, chip financing, equity

Neoclouds · High yield or unrated

12% · Peak-year need as a share of a year of high-yield bonds, convertibles and direct-lending funds

• Need, 2026–30: $347B + $75B to roll
• Peak year: $98B in 2027 incl. $8B to roll
• That market: $799B a year
• Raises through: GPU-backed loans, high-yield bonds, converts, equity

Labs · Private, unrated

22% · Peak-year need as a share of a year of US IPOs and share sales

• Need, 2026–30: $278B
• Peak year: $111B in 2028
• That market: $502B a year
• Raises through: Private funding rounds, strategic investors, an IPO

Small next to Treasuries and mortgages

At the 2028 peak the need is $435B, about 1.3% of GDP. That is 22% of the $1.99T deficit the federal government ran in FY2026, and 21% of the $2.1T of mortgages lenders will write this year. Debt across every US borrower outside finance (households, businesses and governments) is growing about $4.4T a year, and the AI peak would be 10% of that. Even on stated targets, where SpaceX builds to 10 GW and 2027 reaches $755B, the peak is 38% of one year’s deficit.

Large next to what’s actually new

Those comparisons flatter it. Most mortgage lending and bond issuance replaces debt that already exists; the AI need is new money. With mortgage rates near 7%, US mortgage debt grew only $182B over the past year, so the 2028 peak is 2.4× the net growth in all US mortgage debt. If it were all borrowed, it would also be 39% of the $1.1T a year that all US nonfinancial businesses are adding in debt, and 68% on stated targets.

Hyperscalers are the easy part

The four need $438B over five years, peaking at $155B in 2027: 6.8% of a year of US investment-grade issuance at the 2026 pace. Hyperscalers as a group have gone from about 2% of that market in 2023–24 to 11% this year, by Vanguard’s count, including Amazon’s $54B sale in March, the largest corporate bond deal on record, and investment-grade spreads have barely moved: 84bp on October 15, 4bp wider than at the end of August. The need is counted before dividends and buybacks. If the four keep paying out, they raise more; if markets tighten, payouts are the first thing they can cut.

The strain is in three places

SpaceX on its stated targets. Building to about 10 GW by end-2027 needs about $370B in 2027 from one BBB-rated issuer. That is roughly what the whole US high-yield market raised in 2025, 6.9× the largest corporate bond deal ever, and 4.3× what SpaceX raised in its IPO, itself a record. On Street numbers it needs $80B that year. Either the build is slower than Musk has described, or chip vendors and partners finance most of it.

The neoclouds. They need $347B of new money plus $75B of GPU loans to roll, mostly through high-yield bonds, convertibles and private credit. At the 2027 peak that is about 12% of what those markets raise in a year, and convertibles alone carry up to $32B a year, 17% of US convertible issuance. Those markets are already straining. High-yield spreads have widened 60bp since the end of August, to 323bp, after the heaviest month of junk supply this year, while investment grade moved 4bp. CoreWeave’s bonds yield over 15%.

OpenAI. Its own plan burns $278B over 2026–30, $116B more than the cash it started the year with plus the $122B it raised in March. That has to come from private rounds and, eventually, an IPO that Altman says won’t happen in 2026. At the 2028 peak the need is 22% of what US IPOs and share sales raised at this year’s record pace, and two deals, SpaceX’s IPO and Alphabet’s raise, were about two-thirds of that. Anthropic is the opposite case: on its own figures it already funds itself, and its risk is $518B of mostly non-cancellable compute commitments that need its revenue to keep growing.

Issuance is already running ahead of the gap

Goldman counts nearly $600B of AI-linked debt priced this year, against a $259B gap for 2026 here. Part of the difference is issuers raising ahead of need: by Vanguard’s count, hyperscalers had sold about $132B of bonds by the end of July, against a 2026 gap here of $83B for the four, and Alphabet alone raised $85B in June. The rest is borrowers beyond these companies: data-center developers and the lease-backed vehicles that build sites and buy chips for the labs.

Some of the money goes in a circle

Since the start of 2025, at least $110B of the labs’ new equity has come from companies they buy compute from. Amazon alone put $39B into the two labs in the first half of 2026, close to its $45B capex gap for the whole year. That doesn’t change the total here. It moves part of the labs’ need onto their suppliers, whose gaps in the hyperscaler pack don’t include these investments.

Read the total as a floor

It excludes about $830B of data-center leases the four hyperscalers have signed but not started, which developers finance with their own debt; the debt partners and lessors raise to build and equip the labs’ capacity (almost $100B against OpenAI’s sites by late 2025, and more than $100B of packages this year for chips and sites serving Anthropic, among others); power and grid; and payouts. Broader estimates run higher: Morgan Stanley sees a $1.5T data-center financing gap for 2025–28, and JPMorgan expects more than $2.1T of data-center financing from investment-grade bonds alone over five years.

The telecom comparison

The last debt-funded network buildout is the obvious reference. Telecom companies raised about $1.3T of debt in the five years to 2001, by Cambridge Associates’ count, in an economy a third the size of today’s, and by 2001 telecom bonds were nearly a quarter of the US high-yield market. This one needs $1.54T over five years: more in dollars, far less relative to the economy. Where it rhymes is in who is stretching: the weakest borrowers, funded in high yield, selling to a few customers.

So the total is manageable. The risk is in its shape: a few issuers, the weakest credit markets, and revenue that traces back to two AI labs, one of which still has to raise another $116B.

What the comparisons are anchored on

Latest data as of October 17, 2026. Company figures come from the hyperscaler, neocloud, leveraged-middle and lab-funding packs, which cite their own sources.

• Government — Figure: FY2026 deficit $1,993B (preliminary), against $1,775B in FY2025 · CBO, Monthly Budget Review: September 2026
• All borrowers — Figure: Q2 2026: domestic nonfinancial debt $84.1T, growing at a 5.2% annual rate; nonfinancial business $23.97T (4.6%); federal government $34.92T; home mortgages $14.01T · Federal Reserve, Financial Accounts of the US (Z.1), Q2 2026
• Mortgages — Figure: Mortgage balances $13.117T at Q2 2026, up $182B on the year; Q2 originations $505B · New York Fed, Household Debt and Credit Report, Q2 2026
• Mortgages — Figure: Originations about $2.1T in 2026 and $1.99T in 2027, with rates near 7% · MBA forecast, via HousingWire, Oct 11
• Investment grade — Figure: $1.7T issued Jan–Sep 2026, up 31%; $195B in September · IR+M, Monthly Fixed Income Market Update, September 2026
• Investment grade — Figure: Spread 80bp at Aug 31, 84bp at Oct 15 (ICE BofA US Corporate OAS) · FRED, BAMLC0A0CM
• Hyperscalers — Figure: Bond sales $93B in 2025 and about $132B in 2026 to July 31; about 11% of IG issuance in 2026, against about 2% in 2023–24 · Vanguard, The AI buildout comes to the bond market (Aug 19); share as charted by CIO
• Investment grade — Figure: Amazon’s ~$53.8B two-currency sale in March 2026, the largest corporate bond sale on record (Verizon’s $49B in 2013 was the previous one) · IFR
• Corporate bonds — Figure: $12.1T outstanding at Q2 2026 · SIFMA, US Corporate Bonds Statistics
• High yield — Figure: Record $219B issued in H1 2026; $364B in 2025. Other providers count less for 2025 ($298–328B) · Aranca, US High-Yield Newsletter 1H’26 (Bloomberg data)
• High yield — Figure: $38.5B issued in September, the heaviest month of 2026; Goldman counts nearly $600B of AI-linked debt priced this year · Bloomberg, via Briefs, Sep 28
• High yield — Figure: Spread 263bp at Aug 31, 312bp at Sep 30, 323bp at Oct 15 (ICE BofA US High Yield OAS) · FRED, BAMLH0A0HYM2
• High yield — Figure: $1.53T par outstanding at Sep 30, 2026 (ICE index) · Morgan Stanley IM, High Yield Market Monitor Q3 2026
• Convertibles — Figure: $131B raised by US-listed companies this year to early September, a record; 44% AI-related · Bloomberg, via Advisor Perspectives, Sep 11
• Private credit — Figure: Direct-lending funds raised $129.15B in Jan–Sep 2026, up 35% · Reuters, via KSL, Oct 15
• Private credit — Figure: US direct lending about $1.7T (March 2026) · BlackRock, 1H2026 Credit Currents chartbook
• Leveraged loans — Figure: Morningstar LSTA index outstandings a record $1.55T at end-2025 · LSTA
• Equity — Figure: Record $251B of US IPOs and share sales in H1 2026, including SpaceX’s $86.2B IPO (the largest ever) and Alphabet’s $85B raise · Bloomberg, via Briefs, Jun 29
• Labs — Figure: Amazon paid $15.0B and $13.7B into OpenAI in Q1 and Q2 2026, and $5B into each of Anthropic’s Series G and H · Amazon 10-Q, quarter to Jun 30, 2026
• Economy — Figure: Nominal GDP $32.56T at an annual rate, Q2 2026 · BEA, via FRED
• Street estimates — Figure: ~$2.9T of data-center capex 2025–28 and a ~$1.5T financing gap: ~$800B private credit, ~$200B corporate bonds, ~$150B ABS/CMBS, ~$350B other · Morgan Stanley Research, July 2025
• Street estimates — Figure: More than $2.1T of data-center financing from high-grade bonds over five years · JPMorgan, via Fortune, Jun 25
• Telecom, 1996–2001 — Figure: About $1.3T of debt raised in the five years to 2001 (unsourced in the report) · Cambridge Associates, Telecom Investing 2002
• Telecom, 1996–2001 — Figure: Telecom bonds nearly a quarter of US high-yield bonds outstanding in 2001 · Federal Reserve Board staff paper, for the BIS CGFS, April 2001

Method

The AI need is each pack’s debt-or-equity line summed by calendar year: capex minus customer prepays minus operating cash flow, with no surplus carried between years. Hyperscalers come from the hyperscaler pack and the builders from the leveraged-middle pack. Oracle’s fiscal years end in May, so calendar year t is 5/12 of FY t plus 7/12 of FY t+1 (FY31 is taken as equal to FY30). That is why Oracle shows $65B here against $79B in the leveraged-middle pack: seven months of FY26 fell in 2025. Neoclouds are the neocloud pack’s base case, new money only. The labs are the lab-funding pack: OpenAI’s July 2026 plan ($278B over 2026–30, split by year as described there), and zero for Anthropic, which reports adjusted free cash flow above zero. Adding the labs doesn’t double count: what a lab pays a cloud is a cost in its burn and revenue in the cloud’s operating cash flow, so the two cancel out. Supplier investments in the labs move need between groups without changing the total.

Scenarios. “Street consensus” uses Street capex for the hyperscalers and SpaceX. “Stated targets” uses Meta’s $600B-through-2028 pledge, Amazon’s doubled power by end-2027 and SpaceX building toward ~10 GW by end-2027. Oracle has no separate target, and the neocloud and lab figures are already each company’s own plan, so none of those change.

Benchmarks. Flows are annual, and partial-year 2026 figures are annualized as labeled. Net flows (the deficit, the change in mortgage debt, net business borrowing) are the like-for-like comparison for new money; gross issuance includes refinancing. The AI need includes equity, so comparing it with borrowing assumes all of it is borrowed. High-yield totals vary by provider: Bloomberg-based counts put 2025 at $364B, others at $298–328B. Mortgage debt growth uses the New York Fed’s household data; debt outstanding uses the Fed’s Financial Accounts, which measure mortgages slightly differently. Several neoclouds are based outside the US but raise mostly in US markets.

Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at 06.ai/trackers.

Not investment advice. Estimates are labeled as estimates.

OpenAI and Anthropic, the customers behind the hyperscaler, neocloud and leveraged-middle packs, have opposite problems. OpenAI’s own plan burns $278B of cash through 2030, $116B more than it has in hand. Anthropic reports it is already cash-flow positive, but it has signed $518B of compute it mostly can’t cancel.

OpenAI

$278B · Cash burn over 2026–30 in its own July 2026 plan

• In hand: $162B: about $40B of cash at the end of 2025 plus the $122B round that closed in March
• Still to raise: $116B, with the money in hand running out in 2028
• Revenue: About $50B a year at the end of September (about $70B counted Anthropic’s way); the plan has $350B in 2030
• Compute: $856B planned on compute and infrastructure through 2030, about $171B a year, 3.4× current revenue

Anthropic

$518B · Compute commitments over about a decade, roughly 80% non-cancellable

• Cash flow: Positive since Q3 2026 on an adjusted basis, by its own figures
• Raised: $121.5B since 2025, $105B of it this year, plus an IPO of up to $100B planned for November
• Revenue: $16.8B in Q3 2026, about $67B annualized, against $4.6B in all of 2025
• Compute: About $52B a year of commitments on average, 0.8× current revenue

OpenAI keeps raising its own burn forecast

On its own plan, OpenAI runs through what it has raised in 2028

Every 10% that revenue falls short of plan over the period adds about $84B to the burn if costs don’t fall with it. Much of the compute is contracted.

Who writes the checks

Where the debt sits

Little of it is in the labs’ own names. Most is raised by the partners and lessors who build sites and buy chips for them, and repaid from the labs’ rent.

• Credit lines in its own name — OpenAI: About $4.7B, undrawn; a $520M Bank of America line · Anthropic: $2.5B, being expanded to $15B
• Loans from suppliers — OpenAI: None reported · Anthropic: Broadcom, up to $42B, convertible into shares, toward $125B of TPU leases
• Debt raised by partners and lessors — OpenAI: Almost $100B lined up against Stargate sites by late 2025 (Abilene ~$10B, New Mexico $18B, Vantage $38B); $16B for Oracle’s Michigan campus · Anthropic: $35B from Apollo and Blackstone; a $60B Broadcom chip package that also serves others; ~$15B for a campus in Hubbard, Texas
• Supplier money not yet paid in — OpenAI: $21.3B of Amazon’s $50B, due after June 30 · Anthropic: Google up to $30B; Amazon up to $15B; AMD up to $5B; Nvidia up to $10B in the IPO, in talks

Where the labs’ spending lands

The backlogs behind the operating cash flow in the hyperscaler, neocloud and leveraged-middle packs, and how much of each is these two labs.

• Oracle — Backlog: $664B (Aug 2026) · From the labs: About half OpenAI · Pack: leveraged-middle
• Microsoft — Backlog: $625B (Dec 2025) · From the labs: About 45% OpenAI · Pack: hyperscaler
• Google Cloud — Backlog: $514B (Jun 2026) · From the labs: Anthropic’s $200B commitment was over 40% of the Q1 backlog · Pack: hyperscaler
• Amazon (AWS) — Backlog: $496B (Jun 2026) · From the labs: About half OpenAI and Anthropic · Pack: hyperscaler
• CoreWeave — Backlog: $99B (Mar 2026) · From the labs: Under 30% from non-investment-grade AI labs · Pack: neocloud
• SpaceX · From the labs: Anthropic pays $1.25B a month, cancellable on 90 days’ notice · Pack: leveraged-middle

OpenAI: the plan needs another $116B, and the plan keeps growing

OpenAI’s July plan has it burning $278B of cash over 2026–30 while revenue grows from $36B to $350B. It started the year with about $40B and closed a $122B round in March, so on its own plan it runs through that money in 2028 and needs about $116B more. The plan has moved the wrong way: the same five years were $69B of burn in its September 2025 plan and $179B in February’s. And it assumes the revenue arrives. With costs unchanged, every 10% revenue shortfall over the period adds about $84B.

Half of its new money comes from its suppliers

Of the $164B OpenAI has raised since the start of 2025, at least $80B came from companies it buys compute from: $50B from Amazon and $30B from Nvidia. Another $60B came from SoftBank, its partner in Stargate. The March round also arrives in installments: Amazon had paid $28.7B of its $50B by the end of June. Supplier money doesn’t reduce what has to be raised. It moves part of it onto the suppliers, and the capex gaps in the hyperscaler pack don’t include these investments.

Anthropic: revenue caught up, commitments are fixed

Anthropic’s revenue went from $4.6B in all of 2025 to $16.8B in the third quarter of 2026 alone, and it reported positive adjusted free cash flow for the quarter. On those figures it no longer needs outside money to run the business, but it is raising it anyway: $105B of equity this year and an IPO of up to $100B planned for November. The reason is on the other side of the ledger. It has signed at least $518B of compute over about a decade, about 80% of it non-cancellable, and bankers are pricing the IPO on $190–200B of revenue in 2028. Two customers are about a quarter of its revenue. Revenue comparisons between the two also flatter it: counted Anthropic’s way, OpenAI’s ~$50B run-rate would be ~$70B.

The debt sits one step away

Neither lab has borrowed much in its own name. OpenAI’s ~$4.7B credit line is undrawn, and Anthropic’s is growing to $15B. The borrowing happens at the companies that build and equip their capacity. Partners had lined up almost $100B of debt against OpenAI’s Stargate sites by late 2025. Lessors buying chips and building sites for Anthropic are raising $35B from Apollo and Blackstone and about $15B in Texas; Broadcom is arranging a $60B chip package that serves Anthropic among others, and will lend Anthropic up to $42B toward leasing its TPUs. That debt is repaid from the labs’ rent. It carries their credit risk without sitting on their balance sheets, or in the hyperscaler, neocloud and leveraged-middle packs.

Why the hyperscaler, neocloud and leveraged-middle packs depend on this

About half of Oracle’s $664B backlog is OpenAI. OpenAI was about 45% of Microsoft’s commercial backlog at the end of 2025. Anthropic’s $200B Google commitment was more than 40% of Google Cloud’s backlog, and the two labs are about half of Amazon’s. That backlog is what turns into the operating cash flow in the hyperscaler, neocloud and leveraged-middle packs. For OpenAI it depends on raising another $116B or more; for Anthropic, on revenue that keeps growing into commitments it can’t cancel. Added to the hyperscaler, neocloud and leveraged-middle packs, the labs’ own need lifts the five-year total to $1.54T: see the funding-scale pack.

What this is anchored on

Public reporting as of October 17, 2026. Anthropic’s figures come from press accounts of its draft IPO prospectus and investor updates, and are unaudited. OpenAI’s plans are internal projections reported by the press.

• OpenAI — July 2026 plan: negative free cash flow of $278B over 2026–30; revenue $36B in 2026 rising to $350B in 2030 ($840B in total); ~$856B on compute and infrastructure; the March 2026 raise used up by 2028 · FT, via Reuters, Sep 18
• OpenAI — May 2026 projection: ~$305B of burn over 2026–30 · The Next Web, citing the FT
• OpenAI — February 2026 plan: cash flow −$25B (2026), −$57B, −$85B, −$51B, +$39B (2030); 2025 revenue $13.1B, burn ~$8B, ~$40B of cash at year-end · The Information, via The Decoder, Feb 21
• OpenAI — September 2025 plan: cash flow −$17B (2026), −$35B, −$47B, −$8B, +$38B (2030); 2030 revenue ~$200B · The Information, via The Decoder, Sep 6, 2025
• OpenAI — 2025 audited: revenue $13.07B, costs $34B, operating loss $20.92B · Where’s Your Ed At (verified by the FT), Jun 15
• OpenAI — Q1 2026 revenue $5.7B, burn ~$3.7B, >$73B of cash and securities at quarter-end; Q2 revenue $6.7B · The Information, via The Next Web; WSJ, via Outlook Business
• OpenAI — Annualized revenue ~$50B at end-September (~$70B counted the way Anthropic counts cloud-partner sales); expects at least $70B by year-end · Bloomberg, via Bloomberg Línea, Oct 9
• OpenAI — About $1.4T of data-center commitments over the next eight years · TechCrunch, Nov 6, 2025
• OpenAI — $41.0B round fully funded by Dec 2025, $30B from SoftBank and $11B from co-investors · SoftBank Group, Dec 31, 2025
• OpenAI — $110B committed in February by Amazon ($50B), Nvidia ($30B) and SoftBank ($30B); round closed Mar 31 at $122B and an $852B valuation, ~$3B from individuals · OpenAI; TechCrunch, Mar 31
• OpenAI — Amazon paid $15.0B in Q1 2026 and $13.7B in Q2 toward its $50B, with $21.3B to come after June 30 · Amazon 10-Q, quarter to Jun 30
• OpenAI — Talks to raise at least $30B at ~$1.4T; IPO ‘not 2026’ · Bloomberg, via ProPakistani, Sep 30; Reuters, Sep 13
• OpenAI — Partners’ borrowing tied to OpenAI sites ‘almost $100 billion’ (Abilene ~$10B, New Mexico $18B, $38B for Vantage in talks); OpenAI’s own credit line undrawn · FT, via Irish Times, Nov 28, 2025
• OpenAI — Credit line expanded to ~$4.7B, undrawn; Bank of America $520M line in July · TechCrunch, Mar 31; Reuters, via Kitco, Jul 8
• Anthropic — 2025 revenue $4.59B, operating loss $8.06B, $7.33B of compute and infrastructure costs; $20.28B of cash and short-term investments at Dec 31; two customers about a quarter of revenue · Reuters, on the draft IPO prospectus, Sep 28
• Anthropic — At least $518B of compute commitments over about a decade with six partners, ~80% non-cancellable: Broadcom ~$161.2B of equipment leases, Google ≥$111.1B, Amazon $110B, xAI/SpaceX up to $84.5B (cancellable on 90 days), Microsoft $31.4B, AMD >$20B · Reuters, on the draft IPO prospectus, Sep 29
• Anthropic — Q3 2026 revenue $16.8B, adjusted operating margin 11%, adjusted free cash flow positive · AFP, via Khaleej Times, citing US media reports and the FT and Bloomberg, Oct 15
• Anthropic — Q1 2026 revenue $4.73B; Q2 more than $11.5B (preliminary) · Bloomberg, via The Next Web, Aug 15
• Anthropic — Bankers price the IPO on ~$190–200B of 2028 revenue · Reuters, Aug 14
• Anthropic — Series E $3.5B (Mar 2025), Series F $13B (Sep 2025), Series G $30B at $380B (Feb 2026) · Orrick; Anthropic; DatacenterDynamics
• Anthropic — Series H $65B at $965B (May 2026), including $15B previously committed by hyperscalers ($5B from Amazon) · Anthropic, May 28
• Anthropic — Google: $10B in cash in April, up to $30B more tied to performance targets · Bloomberg, Apr 24
• Anthropic — Amazon: $5B in Series G and $5B in Series H; milestone facility for further investment reduced to $15B · Amazon 10-Q, quarter to Jun 30
• Anthropic — Broadcom to lend up to $42B, convertible into shares, toward a $125.2B five-year TPU lease commitment · Reuters, on the draft prospectus, Oct 1
• Anthropic — Lessor financing for its capacity: $35B initial tranche from Apollo and Blackstone; a $60B Broadcom chip package ($42B senior, $18B junior); ~$15B for a Hubbard, Texas campus · Apollo; Bloomberg, via Advisor Perspectives, Oct 2; Reuters, Jul 30
• Anthropic — Credit line expanding from $2.5B to $15B · Bloomberg, via Advisor Perspectives, Sep 5
• Anthropic — IPO targeted for mid-to-late November on Nasdaq, raising up to $100B at about $2T; Nvidia in talks to anchor with up to $10B · Bloomberg, Oct 1; WSJ, via PYMNTS; Reuters, Sep 11
• OpenAI — Related Digital finalizing $16B of financing for Oracle’s Saline Township, Michigan campus serving OpenAI · Bloomberg, via Bloomberg Law, Apr 1
• Oracle — RPO $664B in the quarter to August, roughly half tied to OpenAI · ERP Today, Sep 11
• Microsoft — About 45% of the $625B commercial RPO at Dec 31, 2025 came from OpenAI commitments · Cloud Wars, on Microsoft’s FY26 Q2 slides
• Google — Google Cloud backlog $513.9B at Jun 30, 2026; Anthropic’s $200B commitment was more than 40% of the backlog disclosed for Q1 · Alphabet 10-Q; Reuters, citing The Information, May 5
• Amazon — AWS backlog ~$496B at Jun 30, 2026; OpenAI and Anthropic about half of the announced backlog · Amazon 10-Q; Michael Parekh, citing The Information
• CoreWeave — Backlog $99.4B at Mar 31, 2026, of which non-investment-grade AI labs were under 30% · CoreWeave Q2 2026 results
• SpaceX — Anthropic pays $1.25B a month through May 2029; either side can end it on 90 days’ notice · SpaceX S-1, via Simon Willison, May 20

Method

OpenAI’s need is the $278B of negative free cash flow over 2026–30 in its July 2026 plan, as reported by the FT. That plan’s year-by-year split hasn’t been reported. We keep the February 2026 plan’s 2026 figure ($25B) and spread the rest over 2027–29 in February’s proportions, with 2030 at zero. If 2030 turns cash-positive, as it did in February’s plan, the money needed before then is larger.

Money in hand is about $40B of cash at the end of 2025 plus the $122B round that closed in March. Part of that round is paid in installments: Amazon had paid $28.7B of its $50B by June 30.

Burn forecasts by vintage. The September 2025 and February 2026 figures sum each plan’s reported cash flows for 2026–30, including a positive 2030. May and July 2026 are reported totals.

Anthropic is shown with no outside need on its own figures: it reports positive adjusted free cash flow since Q3 2026. Its cash flow for the first half of 2026 hasn’t been disclosed.

Raised counts new equity only, not tender offers or buybacks. Supplier money is investment from companies the lab buys compute from. Compute a year is OpenAI’s planned compute and infrastructure spending over 2026–30 divided by five, and Anthropic’s signed commitments divided by their roughly ten-year span. One is a spending plan and the other signed contracts, so they are shown for scale only.

Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at 06.ai/trackers.

Not investment advice. Estimates are labeled as estimates.

Over five years Oracle needs about $78.7B of debt or equity even after $67.1B of customer prepays — and has already raised about $67.9B. SpaceX needs $411B on Street numbers, against $93.5B of cash.

• $1.08T — Capex, next 5 years. Both · Street consensus
• $526B — Covered by operating cash flow. Capped at each year's capex
• $67.1B — Covered by customer prepays. Cash received ahead of service
• $490B — Needs debt or equity. 45% of capex

Capex by year, and how it gets paid for

Cumulative next 5 years by company

$B. Dividends and buybacks aren't deducted, so the true need is larger where payouts continue.

How credit markets price them

Ratings, default-swap levels and lender exposure. CoreWeave included as the neocloud reference point.

• ORCL — Rating: BBB- (S&P, cut Jul 2026) · 5-year CDS: 288bp 5y — record, Oct 15 2026 · Debt: $125.3B (Aug 31 2026) · Customer concentration: OpenAI ≈ half of $664B RPO
• SPCX — Rating: BBB · 5-year CDS: No liquid CDS quoted · Debt: $38.4B principal (Jun 30 2026); $40B chip financing in talks · Customer concentration: Anthropic $1.25B/mo + Google $920M/mo, cancelable on 90 days
• CRWV — Rating: High yield · 5-year CDS: >1,000bp 5y, Oct 2026 · Debt: 2032 bonds ≈77¢, >15% yield · Customer concentration: Reference point: neocloud

Oracle: the gap is funded; the counterparty isn’t diversified

Oracle’s FY27 guide is $90–95B of capex but “not more than $70B” of net cash capex. Customers prepay the difference: $11.4B came in during the first quarter alone, and most new contracts are prepay or bring-your-own-hardware. Add $67.9B of notes, convertible preferred and at-the-market equity raised since FY26, and most of the five-year gap is already financed.

The risk sits elsewhere. Oracle books these prepays with a “significant financing component” — economically, a loan from the customer that it repays in compute. About half of its $664B backlog is one customer, OpenAI. That’s why its CDS hit a record 288bp in October with the funding gap largely closed: the market is pricing who pays, not how much is needed. If you treat the prepays as debt, Oracle’s external need rises to $146B.

SpaceX: cash now, a cliff if it builds to plan

On Street numbers SpaceX needs about $411B over five years. The $93.5B of IPO cash covers 2026 and roughly half of 2027. Building to the ~10 GW Musk has described for end-2027 raises the five-year need to $651B, and $426B of that lands in 2026–27 alone, while the build is happening. Against that, the $40B chip financing now in talks is a down payment. There are no prepays: Anthropic and Google pay monthly, and either can walk on 90 days’ notice. The revenue backing this buildout can be cancelled faster than the debt can be repaid.

What this is anchored on

Company filings, calls and press reports as of October 17, 2026.

• ORCL — FY26: capex $55.7B, OCF $32.0B (incl. $4.6B customer prepays), FCF −$23.7B · Oracle FY26 results; Investing.com, Aug 2026
• ORCL — FQ1'27: capex $28.5B, OCF $23.1B incl. $11.4B prepays booked as deferred revenue with a financing component; FCF −$5B · Oracle Q1 FY27 release & 10-Q, Sep 2026
• ORCL — FY27 capex guide $90–95B with net cash capex 'not more than $70B'; most new contracts via prepay or bring-your-own-hardware · Oracle Q1 FY27 call (CFO Hilary Maxson)
• ORCL — Raised $43B senior notes + $5B mandatory convertible (FY26) and $19.9B ATM equity (FQ1'27) · Oracle 10-Q, Aug 31 2026
• ORCL — New RPO won't hit capex or revenue until FY28+; ~half of $664B RPO converts within 36 months · Oracle Q1 FY27 call
• SPCX — H1'26 capex $28.5B (Q2 $18.4B, of which AI $15.8B); H1 operating cash flow $3.5B · SpaceX Q2 2026 10-Q
• SPCX — $14.3B deferred revenue relates to Space and Connectivity contracts — no AI-compute prepayment disclosed · SpaceX Q2 2026 10-Q
• SPCX — Cash $93.5B after ~$86B IPO; $25B senior notes (5.86% avg coupon); seeking $40B ($30B IG + $10B loans) for Nvidia chips · SpaceX 10-Q; FT/Reuters, Oct 6 2026
• SPCX — >2 GW compute end-2026, 'closer to 10 GW' end-2027 · SpaceX Q2 2026 call (Musk)

Method

Oracle uses fiscal years ending May. FY26 is actual (capex $55.7B; OCF $32.0B split into $27.4B ex-prepay and $4.6B prepays, which reproduces the reported −$23.7B FCF). FY27 uses the $92.5B capex midpoint, with prepays set so net cash capex lands at the $70B ceiling. FY28–30 are our estimates: capex peaks in FY28 as the RPO Oracle says “won’t impact capex until fiscal 28” is built, then falls; prepays fade as fewer new contracts are signed; operating cash flow ramps as about half the backlog converts within three years.

SpaceX is calendar-year. Capex: Street blends ~$65B for 2026 (H1 $28.5B, H2 at the Q2 run rate) with analyst paths above $100B for 2027; the stated-target path builds toward ~10 GW by end-2027 at about $50B per GW ($400B of capex in 2027), then spends about $150B a year in 2028–30. Musk hasn’t given a target beyond end-2027, so the path doesn’t keep building at the 2027 pace. Operating cash flow is our estimate from H1 2026’s $3.5B and the ~$26B annual run rate of the Anthropic and Google compute contracts; it is the softest number here.

Prepays appear only where disclosed. SpaceX’s $14.3B of deferred revenue relates to space and connectivity contracts, not AI compute, so it isn’t counted.

Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at 06.ai/trackers.

Not investment advice. Estimates are labeled as estimates.

CoreWeave, Nebius, IREN, Lambda and Nscale need about $623B of capex through 2030 to hit their own targets. Prepayments and operating cash flow cover $278B. The other $347B has to come from debt or equity, plus $75B of GPU loans to roll, at a time when CoreWeave’s bonds yield 15%.

• $623B — Capex, 2026–30. All five · base case
• $136B — Covered by operating cash flow. 22% of capex; $54B in 2030
• $142B — Covered by customer prepays. 23% of capex
• $347B — Needs debt or equity. 56% of capex · peak 2027
• $422B — Gross issuance. Incl. $75B of GPU loans to roll

Capex by year, and how it gets paid for

Sliders at their defaults: Capex per GW 100% · Prepay coverage 100% · EBITDA vs base 100%.

By company

Company by company

The gap peaks in 2027, then cash flow takes over

In the base case, net new money peaks at $90B in 2027. Operating cash flow is only $3B in 2026, with Nscale still losing money and IREN about break-even, but it reaches $54B by 2030 and covers 22% of capex over the five years. CoreWeave is $148B of the $347B the group needs.

Most of that cash is already spoken for

The operating cash flow here is after $61B of interest and excludes $69B of revenue customers prepaid in earlier years. Much of it comes from capacity financed with amortizing GPU loans, which get paid down first: $75B of scheduled amortization has to be raised again, which takes gross issuance to $422B. The 2026 need is roughly covered already: the five raised or committed about $76B this year against a $68B net need. The open question is 2027–30.

What the money costs

Rate at issue on each 2026 raise, with earlier deals for reference. Rings show where CoreWeave’s bonds trade now and what Nebius’ accreting converts yield if held to maturity. Floating loans are shown at SOFR of 3.88% plus their spread.

5.9–6.9% Secured GPU debt rated or priced investment grade · 8.4–9.4% Secured GPU debt below investment grade · 15.1% CoreWeave 9.75% 2031s on Oct 16, issued at par in April · 0.5–4.5% Convert cash coupons; 3.2–7.1% for Nebius with accretion

Convertible coupons understate the cost: the rest is paid in conversion optionality and dilution. Nscale’s $3.36B pre-IPO converts (September, including $1B from Nvidia) are left off because no coupon has been disclosed; they convert at a double-digit discount to the IPO price, capped at a $30B valuation.

What they raised in 2026

Secured debt got cheaper. Unsecured debt didn’t.

GPU loans backed by investment-grade customers now price at 5.9–6.9%, against about 15% on CoreWeave’s first GPU loan and 13–15% on Nscale’s earlier private credit. Unsecured paper went the other way. CoreWeave’s 9.75% 2031s, issued at par in April, traded at 81.8 on October 16 for a 15.1% yield, and its five-year CDS hit a record 1,194.5 bp the day before. Its September convert needed a 2.875% coupon and a 22.5% premium, against 1.75% and 30% in April.

A third of the need leans on converts

If each company keeps its 2026 financing mix, $139B of the $347B comes as secured GPU debt and $123B as convertibles. Nebius drives that: 93% of what it raised this year was convertible. The cash coupons look cheap at 0.5–4.5%, but the notes accrete to 110–125% at maturity, a 3.2–7.1% yield if they never convert, plus the dilution if they do. That channel depends on the share prices, and CoreWeave and Nebius each lost about a fifth of their value this week.

The numbers

$B, calendar years, for the selection above. Reflects the sliders.

Cumulative 2026–30 by company

What this is anchored on

Guidance, filings and market data as of October 16, 2026.

• CRWV — 2026 capex $35–39B; revenue $12.4–13.2B; ≥1.85 GW active power by year-end; ≥8 GW by 2030; 4.2 GW contracted · CoreWeave Q2 2026 call, Aug 11 2026
• CRWV — Q2 adj. EBITDA $1.51B (59%); net interest $640M; deferred revenue $9.7B at Jun 30; H1 operating cash flow $3.66B · CoreWeave Q2 2026 results
• CRWV — DDTL 4.0 $8.5B (A3, ~5.9% fixed / SOFR+225); DDTL 5.0 $3.1B (Ba2, SOFR+450); DDTL 5.5 $2.6B (SOFR+550) · Company releases Mar 31 and May 18 2026; Stocktitan debt summary
• CRWV — 9.75% 2031 notes $2.75B (Apr); 9.625% 2032 $1.25B and 8.50% € 2032 €2B (Jun); converts 1.75% 2032 $4.0B (Apr), 2.875% 2033 $4.2B (Sep) · Company releases, Apr–Sep 2026
• CRWV — 9.75% 2031s at 81.8 / 15.08% yield (Oct 16); 5-year CDS record 1,194.5 bp (Oct 15); Nvidia $2B equity (Jan) · ICE via Bloomberg/Benzinga, Oct 16 2026; Bloomberg, Jan 26 2026
• NBIS — 2026 capex $20–25B; revenue $3.0–3.4B; ~40% adj. EBITDA margin; 0.8–1 GW connected and 5 GW contracted by year-end; >$9B prepayments in 2026 covering 50–60% of associated capex · Nebius Q1/Q2 2026 results and calls
• NBIS — Converts $4.34B 1.25% 2031 / 2.625% 2033 accreting to 120% (Mar); $5.75B 0.50% 2030 (110%) / 4.50% 2034 (125%) (Aug); first secured loan ~$775M at SOFR+250 (Jul) · Nebius 6-K filings, Mar and Aug 2026
• IREN — FY27 capex $25–30B; 0.3 GW IT in 2026, 0.8 GW cumulative in 2027; $4B contracted run rate for 2026 capacity; prepays 45–55% of GPU capex · IREN FY26 results and call, Aug 27 2026
• IREN — $3.65B investment-grade GPU financing at 6.0% (Jun); $2.4B Blue Owl/PIMCO at 9.0% (Aug); $3.0B 1.00% 2033 converts (May) · IREN releases, May–Aug 2026
• Lambda — 3 GW by 2030 target; backlog $15B (Jun) to $50B (Sep); 2026 revenue above $1.5B; raising up to $4B pre-IPO · WSJ via TechCrunch, Oct 6 2026; Bloomberg
• Lambda — $926M term loan B (Baa2, SOFR+300, Aug); $1.008B DDTL at 6.78% fixed (Baa1, Oct); $1B secured facility (May) · Lambda releases, May–Oct 2026
• Nscale — $103.4B contracted (Microsoft ≤$43.8B, Anthropic ≤$44.6B for 460 MW); $24B equipment + $3.5B construction committed, mainly 2026–27; H1 capex $3.2B; deferred revenue $2.0B to $6.5B · Nscale S-1, Sep 18 2026
• Nscale — $3.36B pre-IPO converts incl. $1B Nvidia (Sep); newest facilities SOFR+237.5 (IG); legacy Sandton loans 13–15% · Nscale release Sep 25 2026; S-1
• Rates — SOFR 3.88% and fed funds 3.75–4.00% after a September hike; 10-year UST 5.12%; US high-yield index effective yield ~8.1% · Fed/market data, Oct 13–16 2026

Method

Capex. 2026 is company guidance at midpoints, with IREN converted from its June fiscal year and Nscale built from the commitments in its S-1. Later years turn each company’s stated power target into capacity added per year, priced at a company-specific capex per GW. Lambda’s whole path is our estimate.

Prepayments are gross cash received, from guidance (Nebius, IREN), filings (CoreWeave, Nscale) or an assumed share of capex.

Operating cash flow is revenue × adjusted EBITDA margin, less cash interest on the debt the model raises (solved year by year), less revenue a customer prepaid in an earlier year. No cash taxes (bonus depreciation, losses) and no working-capital swings. It is the biggest swing factor in 2028–30; the EBITDA slider flexes it.

Debt or equity needed is what’s left, split by each company’s 2026 financing mix. Scheduled amortization of secured GPU loans (about five years) is shown as refinancing on top. Bullet maturities inside the window, such as CoreWeave’s 2030 notes and Nebius’ February 2030 converts if they don’t convert, are not included.

Not included: GPU refresh as 2024–25 fleets come off their first contracts; leased powered shells, whose rent is already inside EBITDA; and Crusoe, whose Abilene campus sits in a separately financed joint venture. Power units differ: CoreWeave counts active power, Nebius connected power, IREN critical IT load.

CoreWeave CRWV

At least 8 GW active by 2030 (1.85 GW guided for year-end 2026; 4.2 GW contracted in August).

• Capex: 2026 at the $35–39B guidance midpoint. Then +6.15 GW to reach 8 GW (1.35 GW in 2027, then 1.5 / 1.6 / 1.7 GW) at about $32B per GW; most data centers are leased.
• Prepayments counted gross: about $2B arrived in H1 2026, so 2026 is set at $4.0B, then 10% of capex.
• Cash flow: $12.8B of revenue in 2026, then about $9.5B per GW-year of average active power (consistent with the ~$30B exit run rate pointed to for 2027). 60% adjusted EBITDA margin (Q2: 59%). 8% cash interest on all debt. The $8.2B deferred-revenue balance comes out over four years.
• 2026 raised: $14.2B secured, $6.3B unsecured notes, $8.2B converts, $3B equity.

Nebius NBIS

5 GW contracted by year-end 2026, nearly all online within 2 to 3.5 years; 0.8–1 GW connected at year-end 2026; more than 1 GW of new capacity in 2027.

• Capex: 2026 at the $20–25B guidance midpoint. Then about 4.1 GW to reach roughly 5 GW, phased 1.1 / 1.4 / 1.1 / 0.5 GW so it lands by early 2030, at about $30B per GW (more than 75% owned sites).
• Prepayments: more than $9B guided for 2026, then 38% of capex (about 70% of Q2 deals carried prepays covering 50–60% of capex).
• Cash flow: $3.2B of revenue in 2026, then about $8B per GW-year of average connected power. EBITDA margin 40% (guided), 45% in 2027, 50% after. 3% cash interest; convert accretion is paid at maturity, not counted here.
• 2026 raised: $10.1B of converts and a first $0.78B secured loan.

IREN IREN

0.3 GW IT delivered in 2026 and 0.8 GW IT cumulative in 2027; more than 5 GW of secured power; no stated pace after 2027.

• Capex: FY27 (Jul 2026–Jun 2027) guided at $25–30B. Calendar 2026 ≈ $3.3B H1 actual plus about half of FY27 ≈ $16B; calendar 2027 ≈ $25.5B. 2028–30 assumes 0.5 GW IT a year at about $43B per GW IT.
• Prepayments at 33% of capex: recent prepays fund 45–55% of GPU capex, and GPUs are about two-thirds of the total.
• Cash flow: about $1.4B of revenue in calendar 2026, then about $13B per GW-year of IT load ($4B of contracted run rate on 0.3 GW). EBITDA margin 35% in 2026, 60–65% after (owned sites, cheap power). 5.5% cash interest.
• 2026 raised: $6.45B secured GPU financing, $3.0B converts, about $2.5B equity.

Lambda private

3 GW of compute under management by 2030 (private; no capex disclosure).

• Capex: backlog went from $15B in June to $50B in September, largely one AI lab, so the path is front-loaded: $6B / $13B / $17B / $19B / $18.5B at about $25B per GW. Prepayments assumed at 10% of capex.
• Cash flow: $1.6B of revenue in 2026 (reported above $1.5B), then about $9B per GW-year. EBITDA margin 30% rising to 45%, lower than owners of their own sites because colocation rent is an operating cost. 7% cash interest.
• 2026 raised: $2.9B secured (August and October facilities rated investment grade), plus up to $4B of equity in progress.

Nscale private

$103.4B contracted (Microsoft up to $43.8B; Anthropic up to $44.6B for 460 MW at Monarch, WV); 1.37 GW active and contracted; Monarch to 2 GW by H1 2028.

• Capex, from the S-1: $3.2B in H1 2026 plus $24B of equipment and $3.5B of construction committed, payable mainly in 2026–27, so about $12B in 2026 and $20B in 2027. 2028 covers the Anthropic and Monarch build (~$15B); 2029–30 assumed at $8–9B a year.
• Prepayments average about 23% of contract value at signing; deferred revenue went from $2.0B to $6.5B in H1, so 2026 is set at $6B and later years at 30% of capex.
• Cash flow: revenue from the $103.4B of contracts as they deliver, about $0.75B / $5B / $11B / $16B / $18B. EBITDA margin −50% in 2026 (H1: −142%), 35% in 2027, 50–55% after. 6.5% cash interest. Early losses add to the external need.
• The S-1 says no binding financing is in place yet for the Anthropic contracts.

Source links

• CoreWeave Q2 2026 call
• CoreWeave Q2 2026 results
• CoreWeave 8 GW target
• CoreWeave DDTL 4.0
• CoreWeave DDTL 5.0
• CoreWeave 9.75% 2031 notes
• CoreWeave June notes
• CoreWeave April converts
• CoreWeave September converts
• CoreWeave and Nebius debt stack
• CoreWeave CDS and bond prices, Oct 16
• CoreWeave bonds and the HY index
• Nebius Q2 2026 call
• Nebius March 2026 converts
• Nebius August 2026 converts
• IREN FY26 results
• IREN FY27 capex guidance
• IREN investment-grade GPU financing
• IREN May 2026 converts
• Lambda term loan B
• Lambda fixed-rate financing
• Lambda pre-IPO raise
• Nscale S-1 contracts and commitments
• Nscale S-1 breakdown
• Nscale pre-IPO converts
• Fed funds and SOFR, Oct 15

Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at 06.ai/trackers.

Not investment advice. Estimates are labeled as estimates.

On Street consensus numbers, Microsoft, Alphabet, Amazon and Meta need about $438B of debt or equity over 2026–30 — 8% of capex. The rest comes out of operating cash flow.

• $5.29T — Capex, 2026–30. All four · Street consensus
• $4.85T — Covered by operating cash flow. Capped at each year's capex
• $438B — Needs debt or equity. 8% of capex

Capex by year, and how it gets paid for

By company

Cumulative 2026–30 by company

$B. Dividends and buybacks aren't deducted, so the true need is larger where payouts continue.

The gap is real, but small against the balance sheets

Microsoft funds everything from cash flow on every path. Alphabet ($165B), Meta ($163B) and Amazon ($110B) run gaps concentrated in 2027–28, and some of that is already pre-funded — Alphabet raised $84.75B of equity in June, and all three have issued bonds. Moving to management’s stated targets adds only about $36B: their own targets already roughly match Street numbers.

Why there’s no prepay line

These four are the funding source. They invest in the labs, and the labs commit that money back as cloud spend. None of them reports capex-linked customer prepayments. That changes if a cash-rich lab — after an IPO, say — starts paying hyperscalers up front; that would be worth watching.

Where Oracle and SpaceX went

Both build at hyperscaler scale but fund like something in between: capex far above cash flow, customer prepays or chip financing, concentrated lab customers, BBB-area credit. They have their own pack: The leveraged middle.

What this is anchored on

Guidance, stated targets and analyst estimates as of October 16, 2026.

• MSFT — CY26 capex incl. finance leases ~$175B (was ~$190B before lease-accounting change); FY27 to grow YoY; FY26 OCF $183B · Microsoft FY26 Q4 call, Jul 29 2026
• MSFT — $329B of data-center leases signed but not yet commenced (operating leases, outside capex) · Microsoft FY26 10-K
• GOOGL — 2026 capex $195–205B; 2027 to rise 'significantly'; Q2 FCF −$5.9B (first negative since IPO) · Alphabet Q2 2026 results, Jul 2026
• GOOGL — $84.75B equity raise (Jun 2026) incl. $40B ATM from Q3; LT debt $98.2B at Jun 30 · Alphabet release Jun 3 2026; Q2 balance sheet
• AMZN — 2026 cash capex ~$220B; goal to double power capacity by end-2027 vs 2025; TTM FCF −$7.6B · Amazon Q2 2026 call, Jul 30 2026
• META — 2026 capex $130–145B; 'at least $600B' US infra spend through 2028; ~14 GW compute target · Meta Q2 2026; Zuckerberg (Sep 2025); Citi
• ALL — 2027 Street capex: Citi GOOGL $308B / META $205B / AMZN $288B; Morgan Stanley META $225B→$250B (2028), AMZN $308B→$318B · Citi Jul 2026; Morgan Stanley Sep 2026

Method

Capex is calendar-year. 2026 is latest company guidance at midpoints (Microsoft on its post-lease-change basis). 2027–28 blend Citi, Morgan Stanley and LSEG consensus; 2029–30 assume growth flattening toward 2030, as consensus implies. “Stated targets” uses paths that hit what management has said: Meta’s $600B-through-2028 pledge and Amazon’s doubled power by end-2027. Microsoft and Alphabet have no quantified target beyond “grow”, so they match Street.

Operating cash flow is our estimate, triangulated from FY26 actuals and published 2027 FCF estimates (S&P, MarketScreener, BofA, TIKR); 2029–30 are extrapolated. Each year, cash flow covers capex up to its level; a surplus isn’t carried into the next year.

Not in capex: operating leases signed but not commenced — $329B Microsoft, ~$279B Meta, $137B Amazon, $85B Alphabet — a second, off-balance-sheet funding channel.

Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at 06.ai/trackers.

Not investment advice. Estimates are labeled as estimates.