Anthropic Nscale Deal Locks In 460MW of Vera Rubin Compute for $45B

What happened

Anthropic has agreed to pay about $45 billion to rent AI computing capacity from Nscale, a British infrastructure company. Bloomberg broke the story, and CNBC and TechCrunch later confirmed it through their own sources.125 According to people familiar with the agreement, the deal runs for six years and covers about 460 megawatts of capacity at Nscale's flagship data center development in West Virginia. Bloomberg puts that much electricity on par with the demand of roughly 345,000 US homes at any given moment.5 The hardware will be Nvidia's Vera Rubin platform, which links six separate chips into one system. The capacity is not expected to start serving Anthropic until late 2027.24

More detail has come out since the first reports. The 460 megawatts would come from the first building on Nscale's Monarch campus. That campus is planned as three buildings totaling 1.35 gigawatts, with a full build cost estimated at $71 billion, $47 billion of it for AI processors.3 Spread evenly, the contract works out to about $7.5 billion a year.6 Neither company has published the full terms. Every account so far rests on unnamed sources, so the figures should be read as well-sourced reporting, not disclosed contract terms.6

A streak, not a one-off

TechCrunch called Anthropic's buying spree "compute-gobbling," and the label fits.2 Nscale is only the latest deal in a run that has sped up through 2026. Earlier in August, Anthropic signed a six-year, $10 billion contract with Volta, a cloud startup founded in January, for capacity at a data center in Norway. In July it agreed a $5 billion compute-related deal with AMD.2 In May, Anthropic said it had taken over all of the compute at SpaceX's Colossus 1 data center, which gave it more than 300 megawatts and over 220,000 Nvidia GPUs within a month. It tied that capacity directly to higher usage limits for Claude Code and the Claude API.16

The spring deals were larger still. In April, Anthropic committed more than $100 billion over ten years to AWS, securing up to 5 gigawatts of new capacity, and Amazon put in $5 billion of new equity with up to $20 billion more possible later.12 That same month, Anthropic expanded its agreement with Google and Broadcom for several gigawatts of next-generation TPU capacity starting in 2027. The company called it its largest compute commitment to date.11 On top of these sit a $30 billion Azure arrangement with Microsoft and Nvidia and a $50 billion US infrastructure investment with Fluidstack.16 Other reports add a $9.1 billion, 20-year contract with Riot Platforms3 and a 245-megawatt lease with Hut 8 in Louisiana from December 2025.17

One tally puts Anthropic's contracts since October 2025 at up to $517 billion, covering at least 14.8 gigawatts of new power. On that count, the company's total planned capacity would be around 16 to 17 gigawatts.14 That figure comes from adding up announced deals, not from Anthropic itself, but its rough size is consistent with the individual disclosures.

Where the reporting diverges

The broad story is consistent across outlets, but a few details are not.

The size of the Google–Broadcom deal. Broadcom's securities filing and the coverage of it put the TPU expansion at about 3.5 gigawatts.1820 In its May SpaceX announcement, Anthropic itself described the arrangement as a 5-gigawatt agreement.16 Separate reporting on a possible Google equity investment also cited up to 5 gigawatts over five years.19 The likeliest explanation is that the figures measure different things, such as Broadcom's share versus the whole Google relationship, but none of the coverage settles it.

The SpaceX deal. TechCrunch and others describe it as about $1.25 billion of capacity per month.28 Bloomberg-derived coverage gives a total of $45 billion.9 These figures can fit together, since $1.25 billion a month over a multi-year term would land in that range. Still, readers should be careful about assuming the Nscale and SpaceX contracts are the same size.

Microsoft's role at the West Virginia site. Nscale had announced a letter of intent to supply Microsoft with up to 1.35 gigawatts at the same Monarch campus.6 Channel Insider, citing Bloomberg, reports that Microsoft dropped that letter over the summer and Anthropic took its place.3 IBTimes says the relationship between the two arrangements is still unclear.6 The more specific account, that Anthropic picked up capacity Microsoft walked away from, is the more plausible one. It also makes the deal more interesting: one hyperscaler's retreat became another lab's opportunity.

Anthropic's revenue. Anthropic said in April that its run-rate revenue had passed $30 billion, up from $9 billion at the end of 2025.18 Reuters, as cited by IBTimes, later reported a run rate of $47 billion and projected 2028 revenue of $190 billion to $200 billion.6 Another analysis says annualized revenue is now above $65 billion.14 Rapid growth explains some of the gap, but these are numbers from different points in time and should not be treated as one consistent series.

Why it matters: compute as the bottleneck

Anthropic has been open about why it is buying. Earlier this year it said demand for Claude had put "inevitable strain" on its infrastructure, hurting reliability and performance at peak hours.1 The SpaceX announcement made the connection directly: more capacity led straight to higher usage limits for paying subscribers.16 Put plainly, Anthropic's research roadmap and its products now compete for the same chips, and every new gigawatt eases that conflict.

The company also says compute is needed to train models, not only to serve them. When announcing the Amazon deal, CEO Dario Amodei said the expansion would let Anthropic keep "advancing AI research while delivering Claude to our customers."12 CFO Krishna Rao has described the approach as a "disciplined" one meant both to serve customers and to keep Claude at "the frontier of AI development."11

The Nscale deal also continues Anthropic's habit of spreading its bets across hardware. The company trains and runs Claude on AWS Trainium, Google TPUs, and Nvidia GPUs.16 One supply-chain analysis argues that Anthropic buys compute "like a procurement organization," securing capacity across several vendors years ahead because fabs, EUV lithography tools, and high-bandwidth memory cannot be scaled up quickly.15 Vera Rubin capacity from Nscale adds another top-end Nvidia source to a portfolio that has leaned heavily on Amazon's and Google's custom chips.

The IPO backdrop

The timing points to public markets. Anthropic filed a confidential IPO prospectus with the SEC in June and has been holding early meetings with prospective investors.1 CNBC says the company has to justify a $965 billion valuation and expects its infrastructure strategy to be central to its pitch to Wall Street.1 Other reports describe an IPO target valuation near $2 trillion, a raise of as much as $100 billion, and a claimed revenue opportunity above $30 trillion.97

Nscale has its own reasons to want the deal. The company was founded in London in May 2024 and has raised about $4.9 billion, reaching a $14.6 billion valuation.6 It has told prospective investors it holds $51 billion in contracted revenue and has been preparing a US listing that could raise up to $3 billion.9 The Anthropic contract is its largest to date, and it gives Nscale an anchor tenant just as its order backlog faces questions.76 For a two-year-old company, this is a large step. It now has to deliver gigawatt-scale capacity on schedule, and few companies of its age have tried that.6

The risk on both sides

The concern is the same one that runs through the entire AI infrastructure boom. Anthropic is committing tens of billions of dollars to chips that have not shipped and to buildings that are not finished.6 If models become more efficient, chip prices shift, or revenue grows more slowly, those commitments could end up costing more than expected.3 One analysis estimates that even if Anthropic spent its full annualized revenue on compute, it would take about eight years to cover $517 billion in contracts. It also points out the tension between this buildout and Amodei's warning earlier this year that rivals "don't really understand the risks they're taking."14 Data-center analysts have compared the arrangements in which vendors finance the customers who buy their equipment to earlier telecom cycles.19

Our view is that the Nscale deal is less a bet on one data center than a clear sign of how Anthropic now thinks about frontier AI research: capacity is the binding constraint, and it has to be secured years before anyone knows exactly how much will be needed. Anthropic appears willing to accept the financial risk in exchange for avoiding a shortage of compute during its most important growth period. Whether that turns out to be sound planning or overreach will depend on three things over the next few years: whether revenue keeps growing, whether the capacity is actually used, and whether Nscale delivers the West Virginia site on time.