Nscale has agreed to sell Anthropic roughly $45 billion of AI computing capacity over six years, according to reports published on 26 August 2026 by Bloomberg and CNBC, both citing people familiar with the arrangement. The capacity sits at a data centre development in West Virginia, comes to about 460 megawatts, and will be served by Nvidia’s Vera Rubin generation of accelerators once they begin arriving late next year.
That is a large number attached to a company most people outside the infrastructure trade had not heard of eighteen months ago. It is also, on the arithmetic, about $7.5 billion a year — roughly one ninth of the $65 billion annualised revenue run rate Anthropic reported to investors at the end of July. The deal is neither reckless nor routine. It is the clearest signal yet that frontier labs have stopped shopping for cloud instances and started buying electricity, land and silicon directly.
We have written before about the economics of data center operations and about what the Nvidia Vera Rubin platform was designed to do. This piece is about the transaction itself: what was actually agreed, who the seller is, what 460 megawatts buys, how the price compares, and which parts of the story a careful reader should still treat as unconfirmed.
Table of contents
- What Nscale and Anthropic Actually Signed
- Who Nscale Is, and Why $45 Billion Changes Its IPO
- Inside the Nscale Campus That Will Host Anthropic’s Racks
- Why Nscale Chose Nvidia Vera Rubin
- Where Nscale Fits in Anthropic’s Compute Portfolio
- What the Nscale Deal Actually Prices Per Megawatt
- The Risks the Nscale Deal Does Not Price
- What the Nscale Deal Means If You Buy AI Rather Than Build It
- Nscale and Anthropic: Frequently Asked Questions
- References
What Nscale and Anthropic Actually Signed
Neither company has published a press release. Everything below comes from the 26 August reporting, and the figures are described by both outlets as approximate.
The headline terms
The agreement covers approximately $45 billion in payments from Anthropic to Nscale across six years. In exchange, Anthropic gets contracted access to around 460 megawatts of compute capacity at the Nscale development in West Virginia. Nscale will fit that capacity out with Nvidia Vera Rubin systems rather than the current Blackwell generation, which is why the capacity is described as coming online late in 2027 rather than immediately.
What 460 megawatts means in plain terms
Bloomberg’s framing is the useful one: 460 megawatts drawn continuously is roughly the electricity consumption of 345,000 American homes at any given moment. That is a mid-sized city’s worth of power, dedicated to one customer’s model training and inference, on one campus, under one contract.
Why Anthropic is buying this way
Anthropic said earlier this year that demand for Claude had put “inevitable strain” on its infrastructure. The company has responded with a run of very large capacity agreements rather than incremental cloud spend — Google and Broadcom, Amazon, Microsoft, Advanced Micro Devices, SpaceX and Fluidstack have all signed on in some form. Nscale is the newest name on that list and, notably, the smallest company on it.
The deal at a glance
| Item | Detail |
|---|---|
| Buyer | Anthropic |
| Seller | Nscale |
| Reported value | ~$45 billion |
| Term | Six years |
| Capacity | ~460 megawatts |
| Location | West Virginia |
| Silicon | Nvidia Vera Rubin |
| First capacity | Late 2027 |
| Implied annual spend | ~$7.5 billion |
| Announced by | Neither party; reported 26 Aug 2026 |
The one detail worth flagging
Neither Anthropic nor Nscale has confirmed the number publicly. Both Bloomberg and CNBC attributed it to sources. Nscale is in the middle of preparing a US listing, which is exactly the period in which contract values tend to be described generously and confirmed reluctantly. Treat $45 billion as a well-sourced report, not a filed figure.
Who Nscale Is, and Why $45 Billion Changes Its IPO
The seller matters more than usual here, because Nscale is not a hyperscaler with a balance sheet to absorb a bad year. It is a fast-growing infrastructure company that has been converting contracts into credibility at remarkable speed.
From one Norwegian site to 831 megawatts
Nscale is London-based and opened its first data centre in Norway in 2025. As of August 2026 it operates roughly 831 megawatts across more than a dozen locations worldwide, and it has stated a target of expanding toward the 10 to 11 gigawatt range. Founder and chief executive Josh Payne has pitched the company as a full-stack operator — energy through to compute — rather than a landlord.
The capital stack behind it
The company closed a $2 billion Series C in March 2026 at a $14.6 billion valuation, added a $900 million revolving credit facility in July, and agreed to acquire the distributed-computing software firm Anyscale for roughly $1.65 billion, a transaction expected to close in the second half of 2026. That last purchase brings the Ray scheduling stack in-house, which is how Nscale intends to sell software margin on top of leased power.
What full-stack actually means here
The phrase is worth unpacking, because it is the whole valuation argument. A conventional colocation operator sells space, power and cooling, then leaves the customer to buy, install and babysit the hardware. A neocloud goes further and rents out the accelerators by the hour. The claim being made in West Virginia is broader still: own the land, generate the electricity, buy the racks, run the scheduler, and sell the result as capacity.
Each layer added is another margin pool and another way to fail. Owning generation removes the interconnection wait but adds fuel procurement, turbine maintenance and emissions permitting to the job description. Owning the scheduling software removes a dependency but adds a software business to integrate. The strategy is coherent; it is simply a great deal of company to build at once.
The revenue ramp
The growth curve is genuinely steep. Nscale reported roughly $33 million of revenue for the whole of 2025, about $37 million in the first quarter of 2026, and more than $100 million in the second quarter. Annualising that second-quarter figure puts the run rate somewhere in the $400 to $500 million band.
Quarterly revenue is the number that tells you how young this company still is, relative to what it has promised to deliver.
The backlog, and the listing it is meant to justify
Before this week, Nscale was telling prospective investors it held roughly $51 billion in total contracted revenue, built largely on Microsoft agreements. Goldman Sachs and JPMorgan are steering a US listing targeted at a valuation of up to $25 billion, with a September window and reported proceeds of about $3 billion. Adding $45 billion takes the contracted book to roughly $96 billion — an increase of about 88 per cent, disclosed days before a roadshow.
The number that should give a buyer pause
A contracted book near $96 billion against an annualised revenue run rate of $400 to $500 million is a ratio of roughly 240 to 1. That is not fraud and it is not unusual for this asset class; multi-year compute contracts genuinely are booked as future revenue when signed. But it does mean Nscale is being valued on delivery it has not yet had to perform, at a scale it has not yet operated.
Inside the Nscale Campus That Will Host Anthropic's Racks
The West Virginia development is the Monarch Compute Campus, and it is the most interesting physical asset in this transaction.
2,250 acres in Mason County
Nscale acquired the site through its purchase of American Intelligence & Power Corporation, gaining roughly 2,250 acres in Mason County, West Virginia. The company has described a theoretical ceiling above 8 gigawatts for the campus, with an initial deployment of 1.35 gigawatts committed to Microsoft. Anthropic’s 460 megawatts is therefore about a third of the Microsoft tranche and under 6 per cent of the site’s stated long-run potential.
Power first, grid second
The distinguishing feature is that the campus does not wait for grid interconnection. Nscale is building on-site generation using Caterpillar G3500-series natural gas units, targeting 2 gigawatts of generation capacity by the first half of 2028, and has described the result as America’s first state-certified AI microgrid. Whatever one thinks of the emissions profile, it removes the single longest lead time in American data centre construction: the interconnection queue.
Why West Virginia, of all places
The state offers the three things this kind of build needs simultaneously: cheap land in industrial quantities, an existing gas supply, and a political environment that has actively courted the sector rather than resisting it. Mason County sits near the Ohio River, which matters for both cooling water and heavy freight during construction.
The trade-off is honest and worth stating. Off-grid gas generation at this scale is a substantial new source of combustion emissions in a state whose economy has been reorganised around energy extraction before, with mixed results for the communities involved. Supporters point to construction employment, tax receipts and a permanent operations payroll. Critics point out that a fully automated campus employs comparatively few people once built, and that the emissions are permanent while the tax abatements are not.
Why that changes the timeline
Deployment at Monarch begins in late 2027, using Nvidia’s DSX AI Factory reference design and Vera Rubin NVL72 racks. Building generation alongside compute is why Nscale can credibly promise 2027 capacity while operators queuing for utility power quote 2029 and beyond.
How the site scales against Nscale’s existing estate
The comparison below puts Anthropic’s block in context against the campus commitments and the wider fleet.
Why Nscale Chose Nvidia Vera Rubin
The chip choice is not incidental. It is most of the reason the capacity is priced where it is, and most of the reason it arrives when it does.
What a Vera Rubin rack contains
A Vera Rubin NVL72 rack — briefly marketed as NVL144 — holds 72 Rubin GPU packages comprising 144 GPU dies, paired with 36 Vera CPUs and stitched together over NVLink. Nvidia quotes 20.7 terabytes of HBM4 per rack at a cumulative 1.6 petabytes per second of bandwidth, plus 54 terabytes of LPDDR5X on the CPU side.
The generational deltas Nvidia claims
Against the GB200 NVL72 it replaces, Nvidia claims roughly five times the inference throughput and three and a half times the training throughput, with 1.5 times the HBM capacity, 2.8 times the memory bandwidth and 2.5 times the LPDDR5X. Those are vendor figures measured on vendor workloads, and they should be read as direction rather than guarantee.
Why inference throughput is the number that matters
Five years ago a purchase like this would have been justified almost entirely by training. That has changed. Serving agentic workloads — long conversations, tool calls, code execution loops, documents held in context across many turns — consumes far more accelerator time per unit of revenue than a one-shot chat response ever did. Inference is now the recurring cost of goods sold, not a rounding error after the training bill.
That reframes the five-times inference claim. If it holds even approximately, it changes the gross margin on every token Anthropic serves from that campus for the life of the contract. It also explains why a six-year commitment on unshipped silicon is less reckless than it sounds: the alternative is committing the same money to hardware that will be two generations old by 2031.
Generation-on-generation comparison
| Attribute | GB200 NVL72 | Vera Rubin NVL72 |
|---|---|---|
| Inference throughput | Baseline | ~5x |
| Training throughput | Baseline | ~3.5x |
| High-bandwidth memory | Baseline | 20.7 TB HBM4, ~1.5x |
| Memory bandwidth | Baseline | 1.6 PB/s, ~2.8x |
| CPU-side memory | Baseline | 54 TB LPDDR5X, ~2.5x |
| GPU dies per rack | 72 | 144 |
| Availability | Shipping | From late 2027 at Monarch |
The long-context variant sitting alongside it
Nvidia also ships Rubin CPX, a distinct part aimed at very long context windows, packaged as an NVL144 CPX rack quoting 8 exaflops and 100 terabytes of fast memory. For a lab whose products increasingly involve agents holding large working sets in context, that split between context-heavy and generation-heavy silicon is strategically relevant, and it is part of why a 2027 fit-out is worth waiting for.
Why waiting is the point
Buying Blackwell capacity today would deliver compute sooner and leave Anthropic holding a depreciating fleet through the back half of the contract. Committing to a Vera Rubin fit-out shifts the delivery risk onto Nscale and Nvidia while keeping the silicon current for most of the six-year term. Models trained with reinforcement learning loops consume compute in a way that rewards that patience.
Where Nscale Fits in Anthropic's Compute Portfolio
This deal is large in isolation and modest in context, which is the most surprising thing about it.
The portfolio as reported
| Partner | Reported value | Capacity | Silicon |
|---|---|---|---|
| Google and Broadcom | ~$200bn / 5 years | Up to 5 GW | TPU |
| Fluidstack | ~$50bn | Not disclosed | Not disclosed |
| Nscale | ~$45bn / 6 years | 460 MW | Nvidia Vera Rubin |
| Microsoft and Nvidia | ~$30bn | Azure capacity | Nvidia |
| Amazon | Not disclosed | Up to 5 GW | Trainium |
| SpaceX | Lease | 300 MW | 220,000 Nvidia GPUs |
Headline values, compared
Ranking the disclosed dollar figures makes the shape of the portfolio obvious at a glance.
The silicon diversification story
Read the table by chip family rather than by vendor and a strategy appears. Anthropic has committed to Google TPUs, to Amazon Trainium, and now — through Nscale and through Microsoft — to Nvidia at scale, with AMD in the mix as well. No single supplier can hold the roadmap hostage. The Nscale agreement is the piece that keeps the Nvidia option live outside of Azure.
What a neocloud is, and why labs keep signing with them
Nscale belongs to a category that barely existed three years ago: specialist operators who build AI-specific capacity at speed and sell it wholesale to a small number of very large customers. They win business on two axes the incumbent clouds struggle with — time to power, and willingness to hand a single tenant an entire hall rather than a slice of a shared fleet.
The corollary is concentration on both sides. A specialist operator with four customers is exposed to any one of them changing plans, and a lab renting from a young company is exposed to that company’s ability to finance the build. Both parties know this, which is why these contracts are long, large and structured to make walking away expensive.
The revenue side of the ledger
Anthropic’s annualised run rate reached about $65 billion by the end of July 2026, roughly seven times where it stood a year earlier, with Claude Code the largest single driver. Against that, $7.5 billion a year to Nscale is a serious commitment but not an existential one. The company closed a Series H in May 2026 at a reported $965 billion valuation and has been linked to a public listing as early as October.
What the Nscale Deal Actually Prices Per Megawatt
Strip out the headline and the unit economics are the interesting part.
The arithmetic
Forty-five billion dollars across 460 megawatts is roughly $97.8 million per megawatt for the whole six-year term, or about $16.3 million per megawatt-year. Divided by time instead, the contract is about $7.5 billion annually. Those are the only three numbers you need to sanity-check anything else written about this deal.
What that price actually covers
It is not a power bill. The figure bundles the buildings, the on-site generation, the Nvidia hardware, the networking, the operations staffing and Nscale’s margin. Comparing it to a wholesale electricity rate is meaningless; comparing it to a fully loaded GPU-hour price from a public cloud is closer, and on that basis committed capacity of this size normally clears well below list.
Cost per megawatt-year, illustrated
The one-line takeaway: at $16.3 million per megawatt-year, Anthropic is paying for the silicon and the site, not for the electrons.
The comparison nobody can make yet
Because Nscale has not filed, and because Anthropic reports nothing, there is no public price per GPU-hour to benchmark against. Anyone quoting one is inferring it. What can be said is that the ratio of dollars to megawatts here is consistent with a contract that includes current-generation silicon rather than bare shell space.
The Risks the Nscale Deal Does Not Price
A story this large deserves its objections stated plainly rather than buried in a final paragraph.
Delivery risk sits with a young operator
Nscale has never operated at the scale this contract requires. Going from 831 megawatts live to a multi-gigawatt estate while commissioning on-site generation, integrating an acquired software company and completing a listing is four hard programmes running concurrently. Execution risk here is real and it is concentrated.
Circular accounting across the sector
Nvidia sells to Nscale, Nscale sells capacity to Anthropic, Nvidia has investment exposure across the AI supply chain, and each announcement inflates the others’ contracted revenue. None of this is improper, but a $96 billion backlog resting on a handful of buyers is not the same as a diversified order book.
Power, permits and politics
Off-grid gas generation is fast, and it is also the part most likely to attract local objection, environmental challenge or a change in state policy. The microgrid is what makes 2027 credible; it is also the single point on which the timeline most plausibly slips.
Silicon timing
Vera Rubin is a new platform on a new memory generation. Slippage in HBM4 supply, rack-scale power delivery or liquid cooling would push the fit-out. Anthropic’s exposure is a delay; Nscale’s exposure is a contract it cannot begin billing against.
Risk summary
| Risk | Who carries it | Early warning sign |
|---|---|---|
| Build slips past 2027 | Both | Generation milestones missed |
| Rubin supply constrained | Nscale | Nvidia guidance revisions |
| Backlog concentration | Nscale investors | Few named buyers |
| Gas permitting challenge | Nscale | State or county objections |
| Demand falls short | Anthropic | Run-rate growth flattening |
| Terms never confirmed | Readers | No filing references the figure |
What the Nscale Deal Means If You Buy AI Rather Than Build It
Most organisations reading this will never sign a megawatt contract. The deal still tells you three useful things about the market you are buying in.
Capacity through 2027 is being pre-sold
The frontier labs are locking down power and silicon years ahead. Whatever capacity your vendor resells in 2027 has largely been spoken for already, which is why long-term committed pricing is becoming available and why spot GPU capacity keeps getting squeezed at peak.
Model prices are underwritten by contracts like this
When a lab commits $7.5 billion a year to one supplier, that cost has to be recovered through token pricing, seat pricing and enterprise agreements. Falling per-token prices are real, but they are being funded by scale, not by cheap infrastructure. Build your business case on the value of the workload, not on an assumption that inference gets free.
Diversification is now a procurement question
Anthropic is hedging across TPU, Trainium and Nvidia. A sensible buyer hedges too — at the model layer rather than the chip layer. Keeping an abstraction between your application and any single provider is the same instinct that drives a lab to sign with Nscale as well as Google.
Plan for capacity, not just for price
The practical implication for a mid-sized organisation is unglamorous. If your 2027 roadmap assumes you can simply buy more inference when the workload grows, write down what happens if you cannot get it at the price you modelled. That usually means having a smaller model qualified for the same task, a queueing strategy for non-urgent work, and a clear view of which processes genuinely need frontier capability rather than merely benefiting from it.
None of that is exotic engineering. It is ordinary capacity planning, applied to a resource most teams have been treating as infinite because, so far, it mostly has been.
Questions worth asking your AI vendor
| Question | What a good answer looks like |
|---|---|
| Where does your capacity come from? | Named providers, more than one |
| What happens at peak? | Documented throughput commitments |
| Can we move models? | Portable prompts, no lock-in clauses |
| How is pricing reviewed? | Fixed window with notice periods |
| Where is data processed? | Named regions and residency terms |
If you are early in this journey, our notes on cloud adoption and our AI models and tools hub cover the practical version of those questions.
Nscale and Anthropic: Frequently Asked Questions
Has the deal been officially announced?
No. As of 27 August 2026 neither company has issued a statement. The terms come from Bloomberg and CNBC reporting on 26 August, both citing people familiar with the matter.
When does Anthropic actually get the compute?
Not immediately. The capacity depends on Nvidia Vera Rubin systems, which begin arriving late in 2027 as the West Virginia campus is fitted out.
Is 460 megawatts a lot?
For a single customer at a single campus, yes. It is roughly the continuous draw of 345,000 US homes, and more than half of everything Nscale currently operates worldwide.
Does this replace Anthropic’s other providers?
No. It sits alongside agreements with Google and Broadcom, Amazon, Microsoft, AMD, SpaceX and Fluidstack. On reported value it is the third largest of them.
What does Nscale get out of it beyond revenue?
Validation. A second marquee customer alongside Microsoft, disclosed in the run-up to a listing, materially changes how a $25 billion valuation reads.
Should this change what I pay for AI today?
Not directly, but it should inform your planning horizon. Committed capacity at this scale signals that supply through 2027 is being allocated now rather than sold on demand later.
References
Anthropic to Pay Nscale $45 Billion for AI Computing Power — Bloomberg
Anthropic and Nscale strike $45 billion cloud deal, sources say — CNBC
Nscale: 1.35GW AI factory campus in West Virginia
Nscale to Deploy NVIDIA Vera Rubin Platform in 2027
NVIDIA Vera Rubin Opens Agentic AI Frontier
NVIDIA Unveils Rubin CPX for Massive-Context Inference
Anthropic expands partnership with Google and Broadcom for compute
AI data center builder Nscale reportedly seeking $3B IPO — SiliconANGLE
Nvidia’s Vera Rubin platform in depth — Tom’s Hardware
Anthropic says it hit a $30 billion revenue run rate — VentureBeat
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.