The talks cover a commitment in the range of 45 billion dollars, according to accounts from Bloomberg and Reuters published within hours of each other on the evening of October 19. Neither company would discuss the terms. The buyer is Anthropic, the model developer preparing its own listing. The seller is Nscale, the British cloud company that signed its first contract with Anthropic barely a month ago.
That first agreement runs six years and covers about 460 megawatts at Nscale’s campus in Loughton, in Essex, northeast of London. The value was not disclosed. The second would be roughly the size of the first and would arrive before a single server in Essex has been installed.
The timing is awkward for a reason that has nothing to do with Anthropic. Two days before the talks surfaced, Microsoft ended its agreement for capacity at the same Loughton site, as the Telegraph reported on October 18. Grid officials had told Nscale they did not expect to deliver power to the location until the 2030s.
Microsoft moved the workload elsewhere. Nscale kept the land, the permits and the customer list.
So the campus that lost a hyperscaler this week is the campus at the center of a 45 billion dollar negotiation the same week. Both facts can be true. A site can be the right place to build and the wrong place to connect, and the gap between those two conditions is where the industry’s scheduling problems now live.
Nscale is not a small company in the way that phrase usually means. It has filed for an initial public offering in the United States and told prospective investors before the filing that it holds about 103 billion dollars in contracted revenue, up from 100 million dollars roughly two and a half years ago. Nvidia is an investor. The prospectus names Microsoft and Anthropic as the customers expected to account for most of the revenue.
That concentration is the part of the story a lender reads first. A neocloud is a business that borrows against contracts and spends the money on equipment that depreciates on a schedule set by the next generation of chips. Two customers accounting for most of the revenue makes the borrowing cheaper while both are committed and much harder to roll when one is not.
The economics of a contract like this are unusual in ways that are easy to miss. Anthropic is not buying electricity or floor space. It is buying the right to place a defined number of accelerators in a building and run them for years, at a price agreed before the building exists.
The seller takes the construction risk, the power risk and the financing risk. The buyer takes the risk that it has ordered too much or too little.
For Anthropic, ordering too little has been the more expensive mistake. Its annualized revenue has passed 65 billion dollars, and demand for inference capacity has grown faster than any forecast the company published. Compute commitments are how a model developer converts revenue into the ability to serve the next customer, and the lead time on a large campus is measured in years.
For Nscale, the second contract is worth more than the revenue. It is evidence, in the month before shares are priced, that the loss of one customer at one site did not cost it the account that matters most. A prospectus listing two anchor customers reads differently when one of them has just doubled down.
The financing behind these agreements is where the risk sits. Contracts of this kind are typically assigned to a project entity that borrows against the future payments. Lenders underwrite the credit of the customer, the residual value of the hardware, and the assumption that the capacity will be delivered on the date in the contract. Two of those three are now regularly wrong.
CoreWeave’s bonds illustrate what the market does when it starts to doubt the third. Its 9.75 percent note due in 2031 has traded this month at a yield near 16 percent, a level usually reserved for borrowers already talking to their lenders. The company’s order book did not change. What changed was the willingness of investors to fund a business model they had been happy to fund a quarter earlier.
The difference between Nscale and CoreWeave is mostly a difference of stage. Nscale is asking public investors to price a company whose two largest contracts have not yet produced a megawatt of delivered capacity, at a site whose interconnection date is a decade away in the worst case. The bull case is that the contracts are with the two most creditworthy buyers in the sector. The bear case is that both buyers can afford to wait, and the seller cannot.
There is a version of this that ends well. X-energy’s reactors, grid operators, and the utilities that own the queues all move slowly, and the industry has spent two years learning to build around them rather than through them. On-site generation, purchased capacity and temporary bridges have kept schedules roughly intact at several sites in Texas and Virginia.
What Loughton shows is that the contract and the connection have come apart. A customer can commit 45 billion dollars to a campus and still not know when the lights go on. The seller can book the revenue and still have to explain, to the people pricing its shares, why the first tenant left.
