The biggest contract in the AI buildout is signed, not funded

14 to 20 September 2026

Nscale, which builds and rents out AI computing capacity, disclosed $103.4 billion of signed contracts in its listing paperwork and only $2.6 billion of them active, on financing that was not yet in place. In the same week, the evaluators checking the models were paid, mostly, by the companies selling them.

Three ways to pay for the same buildout

• Equity. The AI data centre builder Crusoe closed a $3.9 billion Series F at a $30.9 billion post-money valuation, roughly three times its mark last October, and the funds that bought it are paid last, after the lenders, and only if the company is worth more later.
• Secured debt. Banks were reported to have assembled a $22 billion facility plus a $1 billion revolver for the Blackstone and Alphabet chip venture, secured on the chips and the contracts rather than the business, so the lender’s recovery depends on what that hardware is worth to somebody else if the venture fails.
• Convertible debt. Nscale raised at least $3.1 billion in convertible notes, debt that converts into shares when the company lists, split between $2.1 billion of unsecured notes and $1 billion from NVIDIA, so the lender is repaid in shares rather than cash, at a price the listing is supposed to set.

The price of money moved while those deals were being done. On 16 September the FOMC raised the target range by a quarter point to 3.75 to 4% on a unanimous 12 to 0 vote, the first increase in 1,148 days, and its projections put inflation at 3.7% this year and the median policy rate at 4.1% by December. New borrowing costs more as the policy rate rises, and equity investors expect a higher return for the same risk, so all three kinds of money above are more expensive than they were a month ago.

The largest contract in the week’s paperwork is conditional, and the filing says so

Nscale filed to list in New York on 18 September under the ticker NSCL, with no price range and no share count set. Its prospectus reported $103.4 billion of contracted value at the end of August, the sum of every payment its signed contracts could produce, of which $2.6 billion was active. The contracts are take-or-pay, average about 5.7 years, and cover roughly 461,000 GPUs.

Two customers hold $88.4 billion of that book: Microsoft’s statements of work allow for up to approximately $43.8 billion through December 2033, and the Anthropic agreements signed on 25 August cover up to $44.6 billion of capacity at Nscale’s Monarch campus in West Virginia.

The financing for the Anthropic capacity is not committed. Nscale had not obtained binding commitments for the financings required under the Anthropic agreements, and because its funding forecast relied on uncommitted debt and equity, the prospectus raised substantial doubt about its ability to continue as a going concern, the wording auditors use when they doubt a business can keep operating. Management’s answer was a plan to defer or cancel capital spending if the money does not arrive, against first-half revenue of $140.6 million and a $1.02 billion net loss, and the agreements let Anthropic walk away without liability if delivery misses its deadlines.

The people checking the models are paid by the people making them

• Anthropic published its own numbers: Claude “leads” 26% of its AI research and engineering work across roughly 30,000 agents, and its monitor stopped 0.002% of their decisions, about 1 in 47,000.
• Anthropic and Accenture will each invest at least $1 billion over five years in a team of evaluators placed inside Anthropic, the company being evaluated.
• Canada and Germany committed up to CAD $300 million to LawZero for Scientist AI, an independent monitor.

What a verification is worth depends on who pays the verifier, and this week that was mostly the verified, plus two states buying a second opinion.

One question I am carrying into next week: 85% of the largest contract book in this buildout depends on financing that had not been committed when the paperwork was filed, and the monitors watching the models are paid by the companies that make them. If both of those hold, who is carrying the risk in the parts of this I actually depend on?