Treat every AI vendor that rents cheap compute as a credit risk in your contracts

Four banks that were scrambling to underwrite AI data centre loans only months ago, JPMorgan, Morgan Stanley, SMBC and MUFG, are now hunting for outside investors to take the riskiest slices off their hands, the Financial Times reports. The reason is mundane. Each bank caps how much of its balance sheet can ride on one sector, and AI lending is pressing against that ceiling. If the people who wrote the paper want less of it, I think the procurement lead about to sign a three-year AI contract should want less of it too.

I started reading this expecting a banking story. I finished it looking at my own vendor list. The chain runs like this. Banks fund special purpose vehicles that build data centres. Those buildings are leased to Oracle, CoreWeave and the hyperscalers. They rent capacity to model companies, which sell tokens to the SaaS vendor whose shiny AI feature you added to your renewal last spring. Every price on that chain was set on the assumption that money for this stuff stays cheap and plentiful.

It is getting less cheap. A bank group including JPMorgan and MUFG spent more than six months trying to distribute $38 billion of construction debt for Oracle-leased sites in Texas and Wisconsin, Bloomberg reported, with some lenders willing to sell slices to non-bank buyers at a discount. Matthew Moniot, who co-heads credit risk sharing at Man Group, told the FT the deal sizes are "out of scale to anything we've thought about, ever." Oracle's five-year credit default swaps sat around 151 basis points in early October on Bloomberg data; by 15 October GuruFocus reported them at a record 288 after a 26-point jump in a day. S&P cut Oracle to BBB-, one notch above junk, on 9 July. Over four quarters Oracle burned through more than $20 billion after capital spending, and the OpenAI contract worth $300 billion, half of its $638 billion backlog, does not start paying until 2027.

The pricing shows who carries the risk. Bonds on CoreWeave-leased buildings price at 7.00% to 9.875%, against 5.70% to 7.875% for hyperscaler-leased paper. Order books for the big AI issuers shrank from about five times the amount on offer in February to under two times by July. Underneath all of it, the 10-year Treasury hit 5.041% on 15 September, its highest since July 2007, and passed 5.1% a week later. A neocloud paying close to 10% to finance its halls has to recover that from tenants at some point, and the tenants recover it from their customers, which includes any company buying an AI feature from a vendor that rents from a neocloud.

The sceptic has decent ammunition. The $38 billion package did get placed, across roughly two dozen lenders. Oracle's shares rose 2.7% on the day of the S&P downgrade because investors cared more about the backlog. In February its default swaps fell 17% after it laid out a plan to raise up to $50 billion. Credit scares around this trade have come and gone all year, and nobody serious is predicting Oracle defaults next month.

I agree with most of that, and it does not change what I would sign. Slow placement at a discount means the marginal buyer is demanding a higher return, and higher returns get billed to whoever rents the building at the next renewal. Your vendor then has two options: eat the increase in gross margin or push it through to you as a price rise, a usage cap or a quiet downgrade of the model behind the feature. The smaller the vendor and the weaker its compute supplier, the sooner that bill arrives, because a startup borrowing next quarter is asking the same wary lenders for money as Oracle, with a fraction of Oracle's assets and customers behind it.

So underwrite your AI vendors the way a bank underwrites a borrower. Ask which provider runs their inference and how long their capacity agreement has left. If the answer is a single neocloud, treat it as supplier concentration risk and negotiate accordingly: a cap on AI price pass-throughs, prepayment limited to twelve months, a termination right if the vendor changes compute provider or reprices usage mid-term, and a tested data export. Sellers should run the same exercise on themselves. If you priced an AI feature as a flat per-seat fee, your finance team ought to model what happens to margin if your compute bill rises by the gap between hyperscaler and CoreWeave bond yields.

Before your next AI renewal lands, email the vendor and ask it to name, in writing, the company that hosts its models and the end date of that contract. If it cannot answer within a week, pay quarterly and keep your exit clause.

Prompted by Banks are now trying to offload the AI data center debt they just underwrote - Startup Fortune, Startup Fortune.