Fentner

Markets · Opinion

Make every AI compute contract say whose money paid for the GPUs

Nvidia has put tens of billions into the customers who buy its chips. Anyone buying or renting AI compute should make suppliers disclose who funded whom before they sign.

In February 2026 Nvidia's $100 billion letter of intent to OpenAI, which never became a binding contract, was replaced by a $30 billion equity stake in OpenAI's roughly $110 billion round. In the same announcement OpenAI committed to 5 GW of compute on Nvidia systems. Some of that $30 billion therefore comes back to Nvidia as orders. If you buy or rent AI compute, you are bidding against customers whose cheques the seller helped write.

TechSpot, citing a Wall Street Journal profile of Colette Kress, says Nvidia pushed more of these deals in mid-to-late 2026 to soak up an oversupply of TSMC wafers and keep GPU prices high. I cannot confirm that. I found no primary source for Nvidia intending to prop up prices, and the report leans on people familiar with the matter. My argument works without intent, because the financing structure is documented well enough to act on.

A buyer financed by its supplier does not have to be sensible about price. Its willingness to pay is set by how much the supplier is prepared to lend, guarantee or invest, and the supplier can see all of those numbers. Everyone else bids against it blind. In August 2026 Nvidia gave residual-value guarantees to SB Energy for about 4.25 GW at OpenAI's Pike County campus in Ohio, with Nvidia's obligation capped at $105 billion. One tally puts total exposure at $108.5 billion in guarantees, $56 billion in further commitments to AI clouds, $29 billion in cloud service agreements where Nvidia rents compute back, and $25 billion in promised equity. Another analyst counts 24 deals worth $48.6 billion in a single year. Add memoranda with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to line up more than $500 billion of outside data-centre money, and a large share of what looks like independent demand has the seller's fingerprints on it.

The best objection is that those totals overlap and shouldn't be added together, and I agree. One analysis argues the real problems are double-counted headlines, a $100 billion pledge that was never finalised, and risk piled onto one company, rather than fake demand. It treats OpenAI's 10 GW Broadcom deal, with no equity disclosed, as an ordinary supply contract. Fine. But if the figures cannot be summed, no outsider can say which are real, and the one party that can is the one charging you. Concentration risk is also a buyer's problem. The report says Nvidia slowed, downsized and paused some financing deals, including the OpenAI backstop, when it worried it had overreached. A cloud provider that built its capacity plan on a chipmaker's guarantee can find that guarantee shrinking, and the repricing lands on its customers.

Huang has repeatedly said these deals are not circular. Perhaps. A customer reading a contract should not have to take either side's word for it, and the fix costs a few lines of drafting. Any compute or GPU-cloud agreement should require the provider to list every equity stake, guarantee, lease backstop, purchase commitment or compute-rental arrangement between it and its chip supplier, and to update that list when anything changes. It should also give you the right to renegotiate price or leave if such a backstop is withdrawn. Providers will say the terms are confidential, so procurement teams should ask for the list under NDA.

This matters most to mid-sized buyers, who have the least leverage and the most to lose if capacity prices swing. Put the clause into every compute renewal you sign from now on, and price any provider that refuses to answer as though the answer is yes.

Prompted by Nvidia deliberately kept GPU prices high through circular financing deals, report claims, TechSpot.