Nvidia Paused The Guarantee Funding The AI Cloud Boom, Two Months After Offering It

In this storyNVIDIA

Nvidia has paused some deals under the financing programme it launched two months ago to help smaller AI cloud operators fund data centre construction, the Wall Street Journal reported on 27 August. The company stepped back the previous week and has left open the possibility of restructuring the initiative or folding it into another effort.

The programme, announced on 1 July in a blog post co-authored by chief financial officer Colette Kress and Raj Mirpuri, was called the AI Compute Partnership. It addressed a specific bottleneck. Operators building GPU capacity must commit billions to chips and infrastructure long before they have signed enough customers to service the debt. Nvidia offered to close that gap by taking a take-or-pay commitment on part of a facility's capacity and guaranteeing a minimum revenue floor, which gave lenders something to underwrite. In exchange, Nvidia took a share of revenue above that floor. Per the Journal, the two sides would set a base hourly rate covering the operator's costs, with Nvidia receiving half of anything earned above it.

The first two partners were Sharon AI, deploying up to 40,000 Grace Blackwell GB300 units, and Firmus Technologies, building a 360-megawatt campus in Batam, Indonesia designed for up to 170,000 GPUs. Together the commitments covered roughly 210,000 GPUs. Nvidia named Baseten, Fireworks AI and Together AI as expected demand-side customers.

A quarterly filing this week disclosed the programme's scale for the first time: about $36 billion in total commitments, on agreements typically running six years.

Nvidia disputes the framing. A spokesperson said the model introduced in July remains in place and continues to evolve on high demand.

Reporting points to two problems, and they turn out to be the same problem.

The Journal reported that Nvidia employees raised antitrust concerns with current and prospective customers, and that there were sensitivities about how far the company could direct the way its customers ran their businesses. One account claims operators were told they could lease the guaranteed GPUs only to Nvidia-approved customers. Separately, in the programme's early weeks, the degree of control Nvidia sought irritated several prospective partners.

The bind is structural. A guarantee is only bankable if Nvidia can constrain what happens to the capacity it is underwriting. No lender accepts a revenue floor from a guarantor with no say over utilisation. But a dominant chip supplier dictating who its customers may sell compute to is precisely the conduct competition regulators look for. The feature and the liability are the same clause.

That is why "pause" may understate things. This is not a pricing disagreement that gets renegotiated over a quarter. Any version that survives legal review carries weaker control provisions, and weaker control provisions make a thinner guarantee, which is worth less to a lender. Nvidia can rebuild the programme. Rebuilding it at the same strength is harder.

Kress described the appeal plainly on the earnings call: the structure lets Nvidia earn on the hardware sale and then again on the rental revenue that same hardware produces. Earning twice requires a claim on both transactions.

For anyone building or funding AI infrastructure, the useful takeaway is not that Nvidia blinked. It is what the episode revealed about the collateral.

Through 2025 the binding constraint on AI compute was data centre space. By early 2026 it was chip supply. By the middle of this year, as SemiAnalysis argued, it had become financing. Lenders wanted an offtake contract or an investment-grade backstop before releasing debt, operators needed the debt to buy the chips, and neither leg would move first. Nvidia's guarantee broke the deadlock by substituting its own balance sheet for a customer contract that did not yet exist.

That worked. It also meant an entire category of companies was being underwritten by one supplier's discretionary willingness to keep offering it, and that willingness turned out to have a two-month half-life. Even if the programme returns in modified form, lenders now know this collateral carries a revocation risk a conventional offtake agreement does not. The next round of neocloud debt gets priced accordingly.

The geography is the part worth noticing from India. The first two partners were not American. Sharon AI is Australian, and Firmus is co-headquartered in Singapore and building in Indonesia. Emerging-market AI capacity was being financed on a US chipmaker's guarantee, in markets where domestic lenders would not underwrite the risk alone. India's own data centre buildout runs into the same wall, and the same solution had started to look available. It is now less clearly available.

There is a longer echo here. Vendor financing has a history. Lucent and Nortel extended credit to the carriers buying their switches, and the receivables looked excellent right until they did not. Nvidia's version is more conservative, because it is buying capacity rather than lending cash. The underlying pattern is the same, and it is worth noting that Nvidia stepped back from it while demand was still strong rather than after anything broke.

For daily, sharp analysis of the biggest moves in the Indian business and startup ecosystem, follow StartupFox.