Nvidia Books $36B in GPU Backstop Obligations, Prices GB300 at Half the Market Rate
Nvidia’s Q2 FY1/27 earnings, released August 26, contained a new line item that caught SemiAnalysis’s attention: $36 billion in aggregate obligations tied to its AI Cloud Partner (AICP) program, spread across six years. The figure represents Nvidia’s financial guarantee that neoclouds participating in the program can rent out GPU capacity at a floor price — effectively making those clusters attractive enough for institutional investors to finance.
How the Backstop Works
Nvidia prices GB300 GPUs at approximately $2.35 per GPU-hour on average under the AICP program. The spot and contract market for the same hardware runs $4.50 to $4.60 per GPU-hour for five-year contracts — roughly double. The below-market rate is intentional. By guaranteeing a floor, Nvidia lets participating neoclouds present their GPU rental revenue as a more predictable stream, which is the condition institutional lenders and private equity funds need before committing capital.
Earlier in August, Jensen Huang announced a $500 billion PE mobilization target tied to this framework. His X post framed Nvidia’s backstop contribution as limited — a “residual value guarantee” capped at 25% of each deal. The $36B figure on the balance sheet reflects what that limited stake looks like at scale after the first cohort of deals.
What It Means for the GPU Market
The AICP program effectively introduces a two-tier GPU rental economy. Market-rate tenants pay $4.50+. AICP-backed clouds have a guaranteed floor at $2.35, with Nvidia absorbing the gap. That gap is the $36B obligation: the maximum Nvidia would be on the hook for if every AICP cluster sat empty for six years.
In practice, the bet is that demand for inference compute will be high enough that neoclouds won’t need the backstop. But if a wave of AI infrastructure overbuilding materialises — the scenario critics have flagged since 2024 — the obligation moves from a balance-sheet footnote to an actual liability.
Why This Structure
The previous model for financing GPU clusters was simple: raise equity, buy hardware, rent it out, service debt from revenue. Institutional investors found it hard to underwrite because GPU rental demand was uncertain. The AICP backstop converts that demand uncertainty into Nvidia’s problem for the bottom 25% of the risk distribution. Institutional capital can now treat the top 75% of a project’s value as underwritten by something other than optimistic demand projections.
The model is closest to a government-backed infrastructure guarantee — except the backer is a hardware manufacturer with an obvious interest in keeping demand for its own chips as high as possible.