024When does replacing old GPUs beat building new power?

The rule

An old fleet should keep its megawatt while its cash margin there beats what that megawatt is worth to the next use. If new power can arrive in time, that is the cost of a new powered megawatt. If it can’t, it is what the newest chips would clear in this one after paying for themselves. Size enters only through which of the two applies: a bigger build-out makes the next megawatt slower to get, which moves more sites from the first case to the second.

Divided by the kilowatt-hours a megawatt delivers in a year, each case becomes a price per kWh of IT capacity, and any GPU’s rent can be set against it. A GPU’s rent per kWh is its hourly rent over its power, counting its share of the server and network. There are three lines:

• Power available, $0.56 per kWh: a new powered megawatt plus power and upkeep. Below it, a GPU isn’t worth its slot anywhere.
• Power wall: the newest GPU’s rent per kWh less its capital charge, $1.33 for a B300. Below it, a site that can’t get new power in time earns more by swapping in the newest chips.
• New chips idle: the newest GPU’s whole rent per kWh. Below it, frontier chips already bought and waiting for power should take the slot.

What a megawatt earns

A megawatt of H100s clears $7.9M–$19.5M a year after power and upkeep; a new powered megawatt costs about $3.0M a year. At every point on the price curve, Hopper out-earns its slot.

The new megawatt is mostly building, not power. Epoch AI puts a 1 GW AI site at $38B upfront and $8.5B a year all-in, with servers about $5B and operations $0.9B of that. The rest annualizes to about $2.6M per MW, between hyperscale leases and wholesale colocation at $100–250 per kW-month.

Generation is the small part. BNEF put new combined-cycle gas plants at $2,157/kW in 2025, and GE Vernova’s CEO expects 2026 gas-turbine orders to price about 50% higher per kW than 2025’s. Even at $5,000/kW, more than twice the BNEF figure, on-site generation adds only about $0.6M per MW-year over 20 years.

Spread over the kilowatt-hours a megawatt delivers, the slot and upkeep ($3.0M + $1.2M) come to $0.56 per kWh of IT capacity. An H100 or H200 keeps its slot above $0.79 an hour, an A100 above $0.51, a B200 above $1.07. Every GPU in the index clears that by 1.8× or more, on every price basis.

A hard power wall

When new power arrives after the old chips’ earning life ends, a megawatt is worth what the newest chips would earn in it, net of their own capital cost. By that measure, at on-demand prices, swapping H100s for B300s has paid since mid-March 2026.

The test, per kWh: swap when the newest chip’s rent beats the incumbent’s by more than the newest chip’s capital charge. For a B300 that charge is $1.33 per kWh: about $63,000 per GPU all-in over 2.0 kW, recovered over 4 years at 10%. The server price is American Compute’s $400,000–450,000 per new 8-GPU B300 machine at volume, plus about 20% for network, storage and installation.

On Nov 5 a B300 rented for $9.59 an hour on demand, or $4.80 per kWh. That puts the power-wall line at $3.46 per kWh: $4.85 an hour for an H100, which rented for $3.90, or 0.80× of its line. A100s sat lower still at 0.71×; MI300Xs (1.06×), H200s (1.13×) and B200s (1.18×) cleared it.

The crossing came when B300 rents jumped in March. In GetDeploying’s weekly medians the B300 went from $5.74 to $7.13 an hour between Mar 2 and Mar 16, when only 4 providers listed it. H100s have rented below their line every week since.

The three-year curve disagrees. At implied three-year prices the B300’s premium over the H100 shrinks to $0.95 per kWh, below its capital charge, and the H100 clears its line by 1.46×. That curve rests on a single provider’s B300 term ratio, but it says the premium is priced to fade before a B300 pays for itself.

If the new chips are already bought and idle, there’s no capital charge left to clear. The line becomes the B300’s whole rent, $6.72 an hour for an H100. That was Microsoft’s position when Satya Nadella said on the BG2 podcast in October 2025 that it had chips in inventory and no powered shells to plug them into.

The wall is real for the next few years. GE Vernova’s gas-power backlog and slot reservations reached 122 GW at the end of the third quarter of 2026, and management says it has sold slots for 2032 delivery. American Compute’s conservative residual forecast puts used H100s at 11–19% of list in 2027 and 5–7% in 2028, so most of Hopper’s earning life ends before those turbines arrive. Evicted chips still have buyers: a used 8-GPU H100 node was worth about $195,000 in August.

If compute stops being scarce

Then Hopper survives as long as renters value its kilowatt at more than about 30% of the newest chip’s. Today they value it at about half.

In that world the newest chip rents at its full cost: its $1.33 capital charge plus the $0.56 slot and upkeep, or $1.90 per kWh. If renters keep paying the same ratio between generations, an old GPU rents at $1.90 divided by that ratio. It keeps its slot while that beats $0.56, so the break comes at a ratio of 3.4×.

Today’s prices are far from that equilibrium. A B300 rents for $4.80 per kWh on demand, 2.5× its full cost; even its implied three-year price ($2.17) sits 14% above. Per kilowatt, though, a B300 rents for only 1.7–2.1× an H100, depending on the price basis. If scarcity ended at those ratios, an H100 would rent for about $1.25–1.54 an hour, still well above its $0.79 line.

Vendor claims point the other way. In Nvidia’s own tests, GB300 NVL72 delivers up to 15x the throughput per megawatt of an H200 NVL8 on DeepSeek V4 Pro, and Vera Rubin NVL72 up to 10x the tokens per megawatt of GB200 NVL72. Those are chosen benchmarks at favorable operating points, but even half the claimed multiples clears 3.4×.

The gap is about what’s being sold. GPU-hours rent across a mix of training, fine-tuning and smaller models, where newer chips win by far less than on frontier mixture-of-experts inference. If you rent out GPUs by the hour, Hopper survives the end of scarcity. If you sell frontier-model tokens, it doesn’t.

The tell is the per-kilowatt rent ratio between the newest chip and Hopper. When it climbs toward 3.4×, the market has started paying for tokens rather than hours, and Hopper is done even with power to spare.

The halls don’t fit

Rack power has more than tripled from Hopper to Blackwell and is headed for about 600 kW, so most Hopper halls can’t take the racks that would replace Hopper without new electrical gear and liquid cooling. The scarce asset is the grid connection, not the building.

Rack power runs from about 30–40 kW for Hopper to 120–140 kW for GB200 and GB300 NVL72 and an estimated 190–230 kW for Vera Rubin NVL72, with Rubin Ultra’s Kyber rack at about 600 kW in 2027. Halls built for earlier densities need their power and cooling rebuilt before they can host Rubin-class racks.

The retrofit is cheap next to what the slot earns. STL Partners puts a liquid-cooling retrofit at about $2M per MW, against $11M or more for new liquid-cooled capacity. The cost that bites is time: the hall earns nothing while it’s rebuilt.

The used market has priced this in. In 2025 H100 rents fell while the median used H100 price rose, because many data centers couldn’t yet host Blackwell; the old chips were worth more to operators whose halls fit nothing newer.

So swaps start in liquid-ready halls, and air-cooled Hopper halls keep air-cooled gear until the building itself is redeveloped. Over a few generations the building becomes a consumable, rebuilt around a durable interconnect. Resale value and retrofit cost shift the line at the margin: a liquid secondary market lowers the bar to evict, and a retrofit raises it.

Where scale matters

Scale enters through the growth rate of powered capacity, not its size. If powered capacity grows by g a year and GPUs leave after L years, this year’s replacements are the deployments of L years ago, so replacement’s share of this year’s new megawatts is (1 + g)^−L. It reaches half once growth slows to about 15–26% a year.

Goldman Sachs sees US data-center power demand rising from 31 GW in 2025 to about 66 GW in 2027, an average of about 46% a year. At that pace replacement is 15–32% of deployments, depending on chip life. Power scarcity shortens that life, which pulls the crossover forward.

Site size cuts the other way. Frontier training needs contiguous multi-gigawatt campuses that no legacy site can provide, so the frontier stays greenfield. In-place refresh is an inference-fleet story.

And decommissioning rarely means scrap. The efficient outcome is a cascade: hyperscalers keep dense, power-scarce slots for the newest silicon and sell older GPUs to operators with air-cooled halls or stranded power. Aggregate power demand keeps rising; the market just tiers.

What it means

The decommission wave starts at walled-off sites with liquid-ready halls, and two spreads set its timing, not the size of the build-out.

• Rent migrates to megawatts. Where power binds, a slot is worth the newest chip’s profit, so whoever owns the interconnect captures the scarcity rent. The power-wall line has risen with every B300 price increase this year.
• Nvidia’s demand grows a replacement leg. At 46% annual growth in powered capacity, replacement is 15–32% of deployments; at 20%, it’s 40–58%.
• Hopper’s book value is the exposure. Amazon cut the useful life of a subset of its servers and networking gear from six years to five in January 2025, citing AI’s pace. Microsoft’s CFO says about two-thirds of its capex goes to short-lived assets, mainly GPUs and CPUs, depreciated over six years; most of that capital is contracted for most of its useful life. A walled-site swap writes Hopper off early, so useful-life changes and impairments are the accounting tell.
• Used-GPU prices carry the cascade. It works only while buyers with power pay around $195,000 per H100 node, and residual forecasts narrow that window through 2028.

What to watch

The spread between the newest chip and Hopper, per kWh, against the newest chip’s $1.33 capital charge. Above it, walled sites should be swapping.

The ratio between them, per kW, against 3.4×. Above it, Hopper loses even with power to spare.

Hopper’s rent per kWh against $0.56. Below it, Hopper loses its slot anywhere.

Turbine lead times, which decide how many sites sit behind a wall.

Rubin’s first rents. When enough providers list it, the newest chip moves from the B300 to Rubin, and the power-wall and new-chips-idle lines reset with it. The rents behind all of these update every day in the GPU Compute Price Index.