Nvidia's revenue-sharing pause re-prices AI cloud risk

Nvidia has paused some of the cloud revenue-sharing deals it introduced in July, the program under which the chipmaker backs cloud operators building AI data centers in exchange for a share of operating revenue, the Wall Street Journal reported and Data Center Dynamics relayed on Friday. It has not scrapped the program entirely—Nvidia told DCD the new business model "is still in place and continues to evolve due to high demand"—while the Journal attributes the halt to antitrust concerns, a twist that chills the merchant GPU cloud sector just as those operators try to finance their next wave of capacity.

The mechanics made Nvidia the buyer of last resort: under the deals, Nvidia acts as financial backstop for companies building large-scale AI data centers, takes a cut of revenues, and rents back unused GPUs at a fixed rate, absorbing utilization risk if demand falls short of projections. The first adopters were Firmus, deploying 170,000 GPUs in Batam, Indonesia, and Sharon AI, deploying 40,000 GB300 GPUs—both merchant clouds betting on speculative demand, without contracted tenants.

Days before the Journal report, in fiscal second-quarter earnings, Nvidia disclosed $108.5 billion in guarantees and other agreements to enable AI cloud computing—$105 billion toward SB Energy Corp for OpenAI's Ohio campus, and $3.5 billion for land, power, and shell guarantees for AI clouds. The $105 billion is a project-finance backstop tied to an anchor tenant, the guarantee that makes a campus financeable; Nvidia's separate $3 billion investment in SB Energy behind the Ohio campus matters less.

The anchor tenant divides the market: hyperscaler-anchored assets clear infrastructure pricing while merchant operators carry merchant risk, and Nvidia, largest supplier of both the GPUs and the risk appetite, is drawing the line itself. The antitrust rationale may be the immediate trigger, but the balance-sheet logic is the same: a rent-back clause is a call option on idle GPUs, and the option just got more expensive. A pause fits Nvidia's statement that the model is evolving; which operators get to stay in the program remains open.

Around that line, Lambda's $3 billion round and Nebius's $5.75 billion debt raise both leave GPU-cloud paper priced on contract demand, a figure the hardware alone cannot supply. Starcloud, raising $250 million to scale on an unbuilt chip, is a bet on Nvidia's roadmap before any tenant arrives. The revenue-sharing pause is a reminder that vendor financing has limits, and Nvidia will extend it selectively.

Nvidia's $108.5 billion of guarantees keeps the balance sheet open for projects with an anchor tenant; the pause shuts it, for now, for those without one. The two tiers of the AI capital stack are now starker, with the same company supplying both the chips and the credit deciding where the boundary falls. Merchant clouds still leaning on vendor backstops are solving next quarter's capacity with next decade's liabilities; Nvidia just made them price that trade.