Bloomberg has reported that Anthropic signed a deal to rent $45 billion worth of cloud computing capacity over six years from Nscale, a UK-based AI infrastructure company. Reuters, citing people familiar with the matter, has confirmed a key term: roughly 460MW secured at a data center in West Virginia, USA. The arrangement reportedly involves NVIDIA's next-generation "Vera Rubin" systems, with supply expected to begin in late 2027. However, Anthropic declined to comment to Reuters, and there has been no joint announcement or public contract from either company.
The $45 billion figure is eye-catching. But the substance of this deal differs from simply buying a large volume of currently available GPUs. It appears Anthropic has reserved capacity more than a year in advance, contingent on power infrastructure, a data center, and GPU systems—slated for deployment starting in late 2027—all coming together as planned.
460MW represents about 34% of the previously announced 1.35GW plan
The reported counterparty is Nscale's "Monarch Compute Campus," under development in Mason County, West Virginia. In March 2026, Nscale announced it had signed a letter of intent to supply up to 1.35GW to Microsoft at this site, which spans up to 2,250 acres. The company said the campus could eventually expand beyond 8GW.
Dividing the reported 460MW by 1,350MW yields approximately 34%—roughly a third of the maximum capacity disclosed for Microsoft. However, it remains unclear whether the 460MW figure refers to IT-equipment load such as GPUs, total facility power, or a reservation ceiling. Nor can the number of GPUs be inferred from power capacity alone.
The timeline aligns with Nscale's public plans. In March 2026, the company said it would deploy Vera Rubin NVL72 systems in multiple phases starting in late 2027. The latest report similarly indicates that Anthropic's capacity would come online during the same window. That said, there is no indication that all 460MW would be operational from day one.
Power is planned to come from on-site natural gas generation, operating independently of the public grid—a design that avoids grid-interconnection delays. But this only becomes usable computing capacity once the power infrastructure, data center, and Vera Rubin deliveries all materialize as scheduled. The current figures are not evidence of operational capacity.
Even once GPUs arrive, large-scale clusters cannot run without power, cooling, and networking infrastructure in place. Nscale plans to build both on-site power generation and AI computing facilities at Monarch. The reported deal itself is not a GPU purchase but a lease of cloud computing capacity.
The strategy behind mixing Trainium, TPUs, and Nvidia GPUs
Anthropic has not concentrated its computing infrastructure with a single provider. In its May 2026 funding announcement, the company said it had secured up to 5GW from Amazon and up to 5GW of next-generation TPU capacity from Google and Broadcom, while also using SpaceX's Colossus 1 and Colossus 2. While treating AWS as its primary cloud and training partner, Anthropic distributes workloads across Amazon's Trainium, Google's TPUs, and NVIDIA's GPUs.
Simply adding up these maximum capacity figures does not mean Anthropic can immediately draw on more than 10GW. Each contract has a different timeline for coming online, and allocation between training and inference is not fixed. The figures also include regional supply variations and future expansion options. Published numbers represent ceilings on securable capacity, not simultaneous operational capacity.
The Nscale deal strengthens Anthropic's NVIDIA-based pathway in particular. At Monarch, Nscale plans to deploy Vera Rubin NVL72 systems in multiple phases. If Anthropic locks in 460MW long-term, it likely gains access to large NVIDIA-based clusters not only for model training but also to meet growing inference demand for Claude.
Anthropic has stated it secured up to 5GW each from Amazon and from Google/Broadcom. Nscale's 460MW equals about 9.2% of that 5GW figure, though this is not a number that can be directly compared to actual operational scale, given differences in the underlying chips, supply timelines, and how capacity and use cases are defined. What is clear is that Anthropic is securing future capacity across multiple chip types and suppliers.
$45 billion is roughly 3.1 times Nscale's valuation
In March 2026, Nscale raised a $2 billion Series C round, valuing the company at $14.6 billion. Dividing the reported $45 billion by $14.6 billion yields a ratio of about 3.1. For Nscale, this represents a customer deal whose total contract value substantially exceeds the company's own valuation.
However, the $45 billion figure will not simply translate into revenue or profit. Averaged evenly over six years, it comes to $7.5 billion per year—though it's unclear whether payments are actually structured evenly. Terms such as minimum usage commitments, upfront payments, and cancellation or reduction provisions have not been disclosed. Nscale, for its part, will incur costs to build facilities, procure GPUs, and provide services once operations begin.
Questions also remain around capacity allocation. In March, Nscale announced a letter of intent to supply Microsoft with up to 1.35GW. A letter of intent is not the same as a final, fully executed supply agreement. Public materials do not clarify whether the reported 460MW for Anthropic replaces part of the Microsoft allocation, adds to it as a separate allotment, or reflects some other reallocation.
This distinction matters for evaluating Nscale's business. Without knowing the relationship between the Microsoft letter of intent for up to 1.35GW and the reported 460MW figure for Anthropic, customer-specific capacity numbers cannot simply be added together. If allocations are finalized and customers commit to long-term payment obligations, Nscale would find it easier to raise capital for its massive infrastructure buildout. What matters is not the number of deals, but to whom—and under what terms—the same power capacity is being provided.
The IPO raises questions about how future capacity is being purchased
On June 1, 2026, Anthropic confidentially submitted a draft Form S-1 to the U.S. Securities and Exchange Commission ahead of a planned IPO. The company has not set a listing date, share count, or price, stating these will depend on market conditions and other factors. A listing "this fall" is not confirmed.
Just before that, in May, Anthropic raised $65 billion, pushing its post-money valuation to $965 billion. The company reportedly said its annualized revenue run rate exceeded $47 billion that same month. However, an annualized run-rate figure is a projection assuming the same pace continues for a full year—it is not confirmed annual revenue. Similarly, since the $45 billion figure represents total contract value over six years, it would be premature to place the two numbers side by side to judge payment capacity.
Looking purely at the numbers, $45 billion equals about 96% of the $47 billion annualized revenue run rate. Meanwhile, if the total contract value is divided evenly across six years, the resulting $7.5 billion per year equals about 16%. Neither figure reflects the actual burden ratio, since the revenue figure is a snapshot annualized estimate, and details about payment start dates, upfront costs, and usage-based components remain unknown. For investors, what matters more than this ratio is the minimum future payment obligations and cancellation terms.
Once a public prospectus is released, investors will be able to assess not only revenue growth but also how much fixed payment commitment Anthropic has made toward computing capacity. Will Monarch actually deliver capacity starting in late 2027, and can Anthropic convert that into growth for Claude? The value behind the $45 billion headline can only be properly assessed once contract terms and actual operational performance are disclosed.