Anthropic has agreed a six year deal worth roughly forty five billion dollars with Nscale, renting about four hundred and sixty megawatts at the British company's flagship data centre development in West Virginia. Reported on Wednesday, August twenty sixth, 2026, the arrangement runs on Nvidia Vera Rubin systems and is expected to start serving traffic in late 2027. It is now the largest single entry in an Nscale backlog the company puts at fifty one billion dollars, weeks before a planned United States listing. The site is one Microsoft had signed a letter of intent for in March and then walked away from.
The short answer
Anthropic has agreed a six year deal worth roughly forty five billion dollars with Nscale, renting about 460 megawatts at the London based company's flagship West Virginia campus. The capacity runs on Nvidia Vera Rubin systems and is expected to come online in late 2027. It is the largest single entry in an Nscale contracted backlog of about fifty one billion dollars, disclosed weeks before a planned United States listing. Microsoft had signed a letter of intent for the same site in March 2026 and withdrew over the summer. The full campus is designed for about 1.35 gigawatts.
The number that tells you most about this deal is not forty five billion dollars. It is 2024, the year Nscale was founded. A two year old company just booked the largest contract in its history from one of the few buyers on earth capable of signing it.
What was agreed
Anthropic will rent roughly 460 megawatts at Nscale's flagship development in West Virginia, over six years, for approximately forty five billion dollars. The compute arrives as Nvidia Vera Rubin systems, and is expected to begin serving Anthropic's traffic in late 2027.
Megawatts have quietly become the unit of account in these agreements, and for a good reason. A contract written in chips ages badly, because the chips change generation inside the contract term. A contract written in power does not, and power is the thing that is genuinely scarce. For scale, 460 megawatts is roughly what three hundred and forty five thousand American homes draw at any one time. The finished campus is designed to reach about 1.35 gigawatts across three buildings, so Anthropic is taking around a third of it.
Development of the campus and its on site generation is estimated at around seventy one billion dollars, with about forty seven billion of that going to semiconductor hardware. Those are estimates in reporting rather than published accounts, and worth treating as such.
The site Microsoft let go
There is an unusual detail in the history of this development. Microsoft signed a letter of intent for the West Virginia site in March 2026, then withdrew over the summer.
Neither company has explained why in any detail, and it would be easy to over read. Letters of intent on data centre sites are routinely signed and later abandoned when interconnection queues, permitting or construction schedules fail to line up. What is certain is the practical consequence: a very large, partly progressed development came free at a moment when the list of organisations that can absorb 460 megawatts on a six year commitment is extremely short. Anthropic was on it.
One deal in a chain of deals
Taken alone this looks enormous. Taken in sequence it looks like policy.
Over roughly eight months Anthropic has agreed a ten billion dollar arrangement with Volta for a Norwegian site, about five billion with AMD, monthly capacity worth around one and a quarter billion with SpaceX across two data centres, an additional five gigawatts of access through Amazon, and an expanded partnership with Google and Broadcom. It has also been reported in talks with Meta over a compute lease.
The common thread is not simply volume. It is that no two of these deals share the same supplier, the same silicon family, or the same grid. AMD and Nvidia. Amazon and Google. Norway and West Virginia. That is a deliberate hedge against the failure mode that actually bites in this market, which is not price but a single delayed interconnection or a single supplier missing a ramp.
The same instinct shows up elsewhere. AMD leasing 530 megawatts from Core Scientific on a fifteen year term and AWS committing to two million more Nvidia GPUs for 2027 and 2028 are the same behaviour from different seats: lock capacity early, worry about the details later.
Two companies heading for the public markets
The timing is not incidental on either side.
Nscale reports a contracted backlog of about fifty one billion dollars, and this agreement is now the largest single line in it. Reporting suggests a United States listing could come as soon as September 2026. A backlog is the central number in an infrastructure company's story, and adding forty five billion to it weeks before a roadshow changes that story considerably. Nscale also bought Anyscale, the company behind the Ray framework, in July, giving it a software layer rather than only racks to rent.
Anthropic, valued at nine hundred and sixty five billion dollars in May 2026, has been holding investor meetings ahead of its own listing. Committing forty five billion dollars of future spending is, among other things, a statement about expected demand made in front of the people who will price that listing.
What this means if you are not building data centres
Most readers are not signing six year power leases. The useful signal is still there, in two places.
The first is timing. Capacity contracted today serves traffic in late 2027. That gap is why inference pricing has stayed tight despite the volume of announcements: none of the capacity being announced this month exists yet. If you are budgeting model spend for 2027, the relief arrives late in it.
The second is concentration. A very small number of buyers are locking up a very large share of the power that will exist in two years, on terms measured in gigawatts. If you depend on a frontier model API, your cost curve is being set right now by contracts you have no visibility into. That is an argument for keeping a smaller open weight model working somewhere in your stack, not because it matches frontier quality, but because it prices differently.
Sources and further reading
- Anthropic and Nscale strike $45 billion cloud deal, sources say, CNBC, August 26, 2026
- Anthropic continues compute-gobbling streak in $45B deal with Nscale, TechCrunch, August 26, 2026
- Anthropic signs $45B compute deal with Britain's Nscale, The Next Web, August 26, 2026
- Anthropic Escalates Cloud Spending With $45 Billion Nscale Agreement, PYMNTS, August 2026
Frequently asked questions
What does 460 megawatts actually mean in practical terms?
It is the electrical draw of the leased capacity, and it is the unit that matters because power, not floor space, is what constrains a modern AI site. Reporting put four hundred and sixty megawatts at roughly what three hundred and forty five thousand American homes draw at any one time. The wider campus is designed to reach about one point three five gigawatts across three buildings, so Anthropic is taking roughly a third of the finished development. Quoting a deal in megawatts rather than in racks or chips has become standard because it is generation agnostic: the contract survives whatever silicon actually gets installed.
Why is Anthropic signing yet another compute deal?
Because it has been signing one every few weeks and the pattern is deliberate. Over roughly eight months the company has agreed a ten billion dollar arrangement with Volta for a Norwegian site, five billion with AMD, monthly capacity worth about one and a quarter billion with SpaceX across two data centres, an additional five gigawatts of access through Amazon, and an expanded partnership with Google and Broadcom. Anthropic has said publicly that demand for Claude has put inevitable strain on its infrastructure. The strategy is diversification as much as volume: multiple suppliers, multiple silicon families, multiple grids, so no single delay stops the roadmap.
What is Nscale and why has it not come up much before?
Nscale is a London based AI infrastructure company founded in 2024, which is remarkably young for a counterparty on a forty five billion dollar contract. It also counts Microsoft among its customers, and reports a contracted revenue backlog of about fifty one billion dollars, of which this deal is now the single largest entry. It has been preparing a United States public offering that reporting suggests could come as soon as September 2026. The company also bought Anyscale, the firm behind the Ray distributed computing framework, in July, which gave it a software layer above the racks rather than only capacity to rent.
Why did Microsoft walk away from the same site?
Microsoft signed a letter of intent for the West Virginia development in March 2026 and withdrew during the summer. Neither company has published a detailed reason, so anything beyond the sequence of events is inference rather than fact. What is worth noting is that a withdrawal at that stage is not unusual in data centre development, where letters of intent are routinely signed on sites that later fail to clear power, permitting or schedule requirements. The practical effect is that a large, partly progressed development became available, and Anthropic took it at a moment when few other buyers could absorb that much capacity.
When does this capacity actually run anything?
Late 2027, according to the reporting, which is a long lead time and a normal one. A greenfield campus of this size involves grid interconnection, on site generation, construction across three buildings and then the installation of an entire Nvidia Vera Rubin fleet. Development of the campus and its power generation is estimated at around seventy one billion dollars, with about forty seven billion of that for semiconductor hardware. Those figures are estimates rather than disclosed accounts. The gap between signature and first token is the single most underappreciated feature of these announcements: they describe capacity arriving in two years, not now.