On July 1, Nvidia announced it would stand behind the smaller cloud companies buying its chips — guaranteeing a floor under their revenue in exchange for a cut of what they earned renting the hardware out. Seven weeks later, the Wall Street Journal reported the company has paused deals under the program.
The structure was called the AI Compute Partnership, and it addressed a real problem. A “neocloud” — one of the second-tier providers building AI capacity outside Amazon, Microsoft and Google — has to commit billions to GPUs and data centre shell before a single customer signs. Lenders want to know who pays if the customers never show. Under the partnership, the answer was Nvidia: a take-or-pay commitment covering a slice of each facility’s capacity, typically running six years, with Nvidia standing ready to buy compute at prices agreed in advance. In return, Nvidia took a share of the rental revenue — paid once on the chip, and again, for years, on the rent.
Two things went wrong, according to the reporting. Nvidia’s own employees raised antitrust exposure with current and prospective customers. And some of the partners the program was built to help balked at how much control Nvidia wanted: providers were told they could lease the GPUs only to customers Nvidia approved.
Nvidia disputes the framing. A spokesperson said the July model “is still in place and continues to evolve due to high demand.” The company has also left the door open to restructuring the initiative or folding it into a different effort later.
Our take: Strip the acronym off and the AI Compute Partnership was vendor financing — the seller helping the buyer pay for the thing being sold. That is not automatically sinister; aircraft, telecom gear and farm machinery have moved that way for a century. It does, however, make a demand number hard to read. When the chipmaker guarantees the revenue of the company buying its chips, some fraction of that order book is the chipmaker’s own balance sheet talking back to it. Whatever Nvidia calls the current status, a prop the second tier was counting on has stopped being load-bearing.
The obligations that don’t pause
Even without the partnership, Nvidia is already deep in the property side of the buildout. Its own quarterly CFO commentary discloses data centre leases of roughly fifteen years, expected to commence in fiscal 2028 and 2029, which the company expects to reassign to third parties. That is a chipmaker carrying long-dated real estate obligations on behalf of tenants it has not yet named.
The cost side is not helping either. Memory pricing has already forced Nvidia to push server prices up roughly 15%, and the company’s own $279 billion in supply commitments assumes the buyers on the other side can keep financing purchases. Cheap credit was the assumption underneath the assumption.
What to watch
- Whether it returns in a different wrapper. Nvidia has signalled it may restructure the program. A version with less control attached would suggest the antitrust concern, not partner resistance, was the binding constraint.
- The next neocloud debt raise. Terms are the tell. If pricing widens on the first deal priced without a backstop, the market is telling you what the guarantee was worth.
- Any formal antitrust interest. Employees flagging exposure internally is not the same as a regulator opening a file — but it is the kind of detail that attracts one.
- Nvidia’s next quarterly filing. Supply commitments and lease obligations are disclosed. Those lines describe the buildout more honestly than any keynote.