Anthropic Rents Megawatts at 8x Its Own Silicon • The Weighted Average

Anthropic has agreed to pay roughly $45 billion over six years for 460 megawatts of compute at an Nscale campus in West Virginia, and the number nobody printed is the one that matters: $16.3 million per megawatt-year. That is what renting frontier capacity costs when you need it in late 2027. It is also about eight times the per-megawatt-year price of the custom-silicon capacity Anthropic secured from Amazon in April, and the gap is now the central economic fact of the model business.

The arithmetic is simple enough to check. CNBC confirmed the roughly $45 billion value and the 460 megawatts at Nscale’s West Virginia development, with Nvidia Vera Rubin systems expected online at the end of 2027; TechCrunch reported the six-year term citing Bloomberg’s original account. Divide $45 billion by 460 megawatts and again by six years: $16.3 million. Against Anthropic’s own disclosure that it is committing more than $100 billion over ten years to AWS for up to 5 gigawatts, the same division yields $2.0 million per megawatt-year. Same buyer, same year, an eight-fold spread.

Anthropic pays 8x more per megawatt-year to rent than to build on custom silicon

Contract value divided by contracted megawatts and contract years, $ millions per MW-year

Anthropic newsroom; CNBC; TechCrunch · Aug 2026

The spread is the whole strategy, not an accounting quirk

Two caveats belong in front of the comparison, because the honest version of this figure is narrower than the headline. The AWS commitment is a ceiling — “up to” 5 gigawatts across a decade — so the realized denominator could be smaller and the realized price higher. The Nscale and Volta figures are contract values against contracted capacity, which is a tighter binding. And the deals buy different things: Nvidia Vera Rubin racks at Nscale, Amazon’s Trainium2 through Trainium4 generations at AWS, hydro-cooled Vera Rubin capacity in Norway from Volta. A megawatt is not a unit of intelligence.

But that is precisely the point an operator should extract. The middle data point proves the spread is not an AWS discount fiction: Anthropic’s $10 billion, six-year Volta agreement covers a 133-megawatt Norwegian facility, per TechCrunch’s report on the Bitdeer-developed site — $12.5 million per megawatt-year, within a quarter of the Nscale number. Two independent neocloud contracts land in the same band. The custom-silicon commitment sits an order of magnitude below both.

What separates them is who absorbs the capital risk. The neocloud rents a finished megawatt with the GPU depreciation, the power contract, and the construction overrun priced in and marked up. Nscale acquired the Monarch Compute Campus with what it calls America’s first state-certified AI microgrid, scalable beyond eight gigawatts, and is deploying Caterpillar G3500 gas generator sets to reach two gigawatts of on-site generation by the first half of 2028. Somebody pays for that turbine fleet, and the lease rate is where it appears.

The custom-silicon side of the ledger looks different because Anthropic is a design partner rather than a tenant. Its AWS commitment spans Graviton and Trainium2 through Trainium4, with an option on future generations, and the company already runs more than a million Trainium2 chips in production. Amazon simultaneously put $5 billion more into Anthropic with up to $20 billion to follow. That is not a lease; it is a decade-long co-investment in which the buyer’s volume underwrites the seller’s fab commitments and the seller’s equity underwrites the buyer’s balance sheet. Nobody prices schedule risk into a relationship like that, because both parties own it.

The pattern repeats across Anthropic’s supplier list. It signed a $5 billion compute-related agreement with AMD that included an equity investment, and it committed $50 billion to custom-built American data centers with Fluidstack in Texas and New York, facilities the company says are engineered specifically for its workloads. Every arrangement where Anthropic shapes the hardware costs less per watt than every arrangement where it simply takes delivery.

Follow the balance sheet, find the constraint

Anthropic’s compute spree only makes sense against its revenue curve. When it expanded the Amazon agreement, the company disclosed run-rate revenue had surpassed $30 billion, up from roughly $9 billion at the end of 2025, and admitted that growth “places an inevitable strain on our infrastructure,” degrading reliability at peak hours. By late July that run rate had reached $65 billion, a figure this paper worked through when Anthropic’s IPO math showed a $42 billion 2025 net loss against it. A company growing that fast does not optimize the unit price of capacity; it optimizes the date the capacity arrives.

That is what the premium buys. Trainium capacity scales on Amazon’s roadmap and Amazon’s fabs. Nscale’s West Virginia tranches begin in late 2027 regardless of what Annapurna Labs ships. Paying 8x for schedule certainty is defensible when the alternative is telling enterprise customers that Claude is slow between nine and five.

The timing pressure is not abstract. Anthropic confidentially filed its IPO prospectus with the SEC in June and has begun preliminary investor meetings, CNBC reports alongside the Nscale deal, against a $965 billion private valuation set in its Series H round. A lab heading into public markets cannot present a capacity gap as a growth story. Every megawatt with a contracted delivery date is a line in the prospectus; every megawatt still in negotiation is a risk factor. That asymmetry alone justifies a premium no engineering team would otherwise approve.

The competitive backdrop makes it more defensible still. AWS and Nvidia announced they will deploy 2 million additional GPUs across AWS infrastructure in 2027–2028, on top of the more than 1 million planned from 2026, because “demand has exceeded those expectations.” When the largest cloud on earth says its own forecast was too small, a lab with a 460-megawatt gap in 2027 does not wait for the queue.

Nvidia’s hardware economics argue the same way from the other end. The company rates Vera Rubin NVL72 at one-tenth the cost per million tokens and up to 10x more tokens per megawatt than GB200 NVL72 for interactive agentic reasoning — vendor projections on a specific model and context length, but the yardstick the lease will be judged against. If Rubin delivers even half that, the $16.3 million megawatt-year is cheap per token even while it is expensive per watt. This paper made the same conversion when India’s AM Intelligence order implied $40 million per megawatt of Rubin capex — a build cost, against today’s rent, on comparable silicon.

The ways this spread stops being clever

The first risk is that the schedule premium buys nothing. Anthropic’s Nscale capacity is expected at the end of 2027; Volta’s Norwegian phases target late 2026 and early 2027. If Trainium3 and Trainium4 land on time at a fifth the unit cost, the lab will be carrying six-year leases signed in a scarcity panic while its cheapest supply comes online underneath them. Long-dated contracts are only smart when the shortage outlasts the term.

The second is counterparty youth. Nscale was founded in 2024; Volta was founded in January of this year. Both are building on financing structures that assume AI demand holds, and Nscale is heading for a New York listing. A megawatt promised by a two-year-old company with a gas-turbine supply chain is a different asset from a megawatt inside an Amazon region, and no lease rate fully prices that. The Monarch campus itself changed hands only this year, when Nscale acquired American Intelligence & Power Corporation and its 2,250-acre Mason County site — an asset that was a power project before it was a compute vendor.

The third is the denominator itself. “Up to 5 gigawatts” is a planning figure, not a purchase order. If Anthropic ultimately draws 2 gigawatts against the $100 billion AWS commitment, the custom-silicon price rises to $5 million per megawatt-year and the spread compresses from 8x to roughly 3x. That still favors owned silicon, but it changes the strategic conclusion from “renting is a luxury” to “renting is a manageable premium.”

The fourth is the one that would break the whole frame: power. Nscale’s campus runs on an on-site microgrid with Caterpillar engines rather than the grid, which is how it can promise 2027 delivery at all. Gas-fired capacity tied to data centers is being proposed faster than it can be built — Global Energy Monitor counts 189 gigawatts of US gas capacity in development and turbine lead times stretching to multiple years, with developers turning to reciprocating engines to skip the queue. A slipped turbine delivery turns a schedule premium into a schedule penalty.

What a buyer should do with an 8x spread

None of these risks argues against signing. They argue for reading the spread as a dated measurement rather than a permanent law. The 8x gap describes August 2026, when Rubin-class racks are scarce, gas turbines are backordered, and a frontier lab’s reliability is a headline risk. Each of those conditions has a plausible path to easing inside the six-year term the contract locks in.

The number generalizes past Anthropic. Any team negotiating capacity — a Series B with a training run, an enterprise reserving inference, a neocloud pricing a lease — now has a public band for what rented frontier megawatts cost in 2026 dollars: $12.5 million to $16.3 million per megawatt-year on Rubin-class hardware, six-year terms, delivery 12 to 18 months out. Anything materially above that band is a seller exploiting your deadline. Anything materially below it should prompt a question about who is bearing the power risk.

The second implication is about vertical integration. The 8x spread is the clearest public estimate yet of what co-designed silicon saves at scale, and it explains why every lab with the capital is building or funding chips. It also explains why Anthropic keeps tying its hardware roadmap to partners rather than owning fabs — the savings come from the commitment, not the ownership.

For operators without gigawatt budgets, the checklist is short:

If you are buying reserved capacity, price it per megawatt-year and compare against $16.3 million. Vendors quote total contract value because the total obscures the rate. Ask for capacity, term, and price, then divide.
If your workload can wait, wait. Two million additional AWS GPUs land in 2027–2028, and Nvidia says Vera Rubin is ramping into full production with Taiwanese server makers shipping at scale. Scarcity pricing peaks before supply arrives, not after. Our reading of Lambda’s 8x revenue multiple ahead of its IPO says the same thing from the seller’s side.
If you depend on a frontier API, model the lease behind it. Anthropic’s token prices must eventually cover $16.3 million megawatt-years. Today’s premium-tier token economics are the visible surface of that math, and the pressure runs upward.
Watch the delivery dates, not the deal sizes. The evidence that would overturn this analysis is a Trainium slip or a Monarch turbine delay. Either one moves the spread, and the spread is the story.
Treat equity in the supplier as a price signal. Amazon’s $5 billion and AMD’s investment both came bundled with capacity, and both landed at the cheap end of the band. When a vendor will not take a stake, expect neocloud pricing.

Two of today’s other reports run on the same current: the open-weight side of the market is pricing intelligence at a fortieth of frontier rates, and the memory shortage created by these buildouts is now rewriting the rules for phone software. Compute scarcity does not stay in the data center.

The spread is the whole strategy, not an accounting quirk

Follow the balance sheet, find the constraint

The ways this spread stops being clever

What a buyer should do with an 8x spread

Sources