Chart packs on markets, AI infrastructure and the capital behind it.
016Who owns the most Blackwell per dollar?
77×as many Blackwell GPUs per $1B of market cap at CoreWeave as at Microsoft. Most of the gap is debt Nvidia’s Blackwell GPUs owned by each listed company per $1B of market cap, and who is lined up for the most of what comes next. The neoclouds lead by one to two orders of magnitude, mostly because they buy their GPUs with debt.
015Who makes money serving open models?
37%utilization covers a Blackwell rack’s rent serving DeepSeek V4.1-Flash at off-peak prices. An H200 would need 120% Open-weight models now carry most of OpenRouter’s US traffic, served by dozens of hosts competing on price. What a GPU earns serving them, chip by chip and model by model, who keeps the margin, and who wins if GPU rents invert.
014How many tokens can a GPU make?
6.3×the output per GPU DeepSeek got in 2025, at 11× the speed per user. vLLM matched its engine three days later DeepSeek open-sourced the engine behind its own API. Its benchmarks show what a Blackwell GPU now yields, what that is worth, and which chips get the gain. vLLM matched it three days later.
013Can the cloud pay for the AI capex?
12.8%a year: what Amazon’s AI capex through 2027 earns from AWS alone, its lab stakes at zero. Microsoft’s, from Azure: 3.0%. Meta and Alphabet from pack 011 Azure is growing 43% and AWS 37%, the fastest in years. How much of that is AI, what Microsoft’s and Amazon’s AI capex earns from the cloud alone with their OpenAI and Anthropic stakes at zero, and what happens if the labs can’t pay.
012How much memory can each AI chip get?
~270 GBof HBM per accelerator, on average, is what 2027 supply can support; ~430 GB by 2030, against ~690 GB on announced roadmaps Supportable by supplyAnnounced roadmapsAI accelerator shipments nearly double by 2030, and HBM supply can’t keep up with the memory per chip on today’s roadmaps. We balance HBM supply and demand to 2030, size the memory per accelerator the market can afford, and weigh KV cache, Marvell’s optical pooling, China and Terafab.
011Can ads alone pay for the AI capex?
35%a year: Meta’s return on its AI capex through 2026 from ads alone, with every moonshot at zero. AI is about 22% of its 2026 ad revenue Ads without AIFrom AIMeta’s revenue is up 26% on a year ago and Alphabet’s 24%, faster than before the AI buildout. How much of that is AI, and what Meta’s and Alphabet’s AI capex earns if superintelligence, Waymo and every other moonshot never return a cent.
010Will AI compute run short or long through 2030?
21%short of AI compute demand at the end of 2030 in our base case. Supply about catches up in 2027, then lab spending outgrows the power and money for new capacity More than demandShortPack 009 sized the compute AI demand needs to the end of 2027. Here we add the supply side, the chips, power and money to switch on new capacity, carry demand to 2030 and compare the two at the end of each year.
009How much compute does AI demand need by 2027?
+47 GWof new GB300-class inference capacity by end-2027 if the market keeps OpenRouter’s pace. Its compute grows 10× for 22× the tokens TokensComputeDollarsPack 007 projects OpenRouter’s tokens to the end of 2027. We turn them into GPUs and gigawatts with benchmarks for each chip and model, then scale the answer up to the whole market.
008How big do OpenAI and Anthropic get by 2027?
$265BAnthropic’s run-rate in December 2027 in our base case; OpenAI about $230B We gathered 145 reported figures on the two labs’ revenue, rebuilt their run-rates month by month, tested simple forecasting rules against what was later reported, and projected to the end of 2027.
007How fast is AI demand growing?
33×tokens a day through OpenRouter, from a year ago. Chinese models now supply most of them USChinaOtherOpenRouter publishes daily token counts for the models it routes. Tokens are up 33× in a year and spend about 11× since January. What’s driving it, who captures it, and three paths for 2027.
006Who sets the price of a token?
42× cheaperthan Claude Opus 5 per output token, and a higher DeepSWE score on DeepSeek’s own figures DeepSeek’s newest model claims top-tier scores on agentic coding at OpenAI’s budget-tier price. What earlier DeepSeek releases did to US token prices, and what this one might do.
005Can markets fund the AI buildout?
$1.54Tof new debt or equity by 2030, $435B of it in 2028 HyperscalersOracle, SpaceXNeocloudsLabsPacks 001–004 added up: about $1.54T of debt or equity through 2030, set against government borrowing, mortgages and the markets that would have to supply it.
004Who funds the labs?
4×OpenAI’s own 2026–30 cash-burn forecast, Sep 2025 to Jul 2026 OpenAI and Anthropic are the customers behind packs 001–003. What each still has to raise to pay for its compute, and how much of that money comes from the companies selling it to them.
003The leveraged middle
43%of Oracle’s and SpaceX’s capex comes from outside. The hyperscalers: 8% Cash flow & prepaysDebt or equityOracle and SpaceX build like hyperscalers but fund like neoclouds. What each needs from outside, and who’s lending.
002Who funds the neoclouds?
56%of $623B of capex has to be borrowed or raised PrepaysCash flowDebt or equityCoreWeave, Nebius, IREN, Lambda and Nscale need about $623B of capex through 2030. What prepayments and cash flow cover, what has to be borrowed, and what that money costs now.
001Who pays for the AI buildout?
$438Bof capex cash flow can’t cover, 2026–30 Cash flowDebt or equityCapex the four hyperscalers need through 2030, and how much of it operating cash flow can’t cover.