Nvidia pauses its AI cloud revenue-share program

Nvidia has paused the financing program it launched in July to backstop AI cloud providers buying its chips, according to Wall Street Journal reporting picked up by Reuters. The program lasted under two months. If you buy inference from a smaller cloud rather than a hyperscaler, that vendor's ability to finance the GPUs you rent just got harder — and you probably won't be told.

What actually happened

The structure, as reported: Nvidia provided credit backing so AI cloud firms could buy its accelerators, in exchange for roughly half of cloud revenue above a set threshold, plus the right to lease back capacity that went unsold. Reporting says Nvidia employees raised concerns internally about how much control that gave the company over customers' businesses — including telling some customers which buyers they could rent to, and steering capacity toward several smaller AI firms instead of one large one — and about the antitrust exposure that follows. Nvidia has reportedly left the door open to restructuring the effort later.

The pause lands two days after the earnings call. On August 26 Nvidia posted Q2 revenue of $96.2 billion, up 106% year over year, guided Q3 to $108 billion, and issued a first-ever year-ahead forecast of about 70% growth for fiscal 2028. On the same call, CFO Colette Kress said demand from the AI labs Nvidia backs will account for roughly a quarter of its business next year, and addressed the obvious objection directly: "we know some will call this circular financing. We see it differently."

A quarter of next year's revenue tied to customers Nvidia capitalizes. That is the context for quietly shelving a program that extended the same idea to the cloud layer.

Why your inference vendor's financing matters for your business

The cheap token has a balance sheet behind it. Neocloud pricing that undercuts the hyperscalers is usually funded by someone — vendor credit, debt against the hardware, or an equity story. When the credit source pauses, the pricing follows within a renewal cycle or two. We wrote the same warning when $500 billion of third-party capital showed up behind the compute.

Match your contract length to your visibility. You cannot audit your provider's financing. You can refuse to sign 36 months against pricing that depends on it. Short terms, no minimum commits you can't hit, and a stated price for what happens at renewal. Same rule we applied to the OpenAI backstop.

Keep the provider swappable at the code level. One interface, provider behind a config value, evals in your repo so you can prove a replacement is good enough before you cut over. Then a vendor's funding problem costs you an afternoon, not a quarter.

Sources: Reuters, Fortune.