Inside Nvidia’s $366B Commitment Stack: How Compute Became Wall Street’s Newest Credit Asset Class
Nvidia's latest 10-Q, filed alongside a Q2 in which revenue hit $96.2 billion and Data Center sales grew 117% year-over-year, buries a structural revelation inside its commitment disclosures. The company now carries $366 billion in ordinary future commitments—supply contracts, cloud purchases, data-center leases, equity-investment pledges, and capital expenditures—on top of its existing balance sheet. A separate layer of customer-support arrangements adds $36 billion in AI-cloud capacity backstops, $20 billion in third-party data-center leases, and a maximum guarantee exposure of $108.5 billion, headlined by a $105 billion residual-value guarantee backing OpenAI's tenancy at SB Energy's 4.25-gigawatt Ohio campus.
Nvidia is no longer a company that sells chips to whoever shows up with cash. It is becoming the credit architect of the AI infrastructure buildout.
The Mechanics of Credit Orchestration
The preliminary agreements Nvidia signed with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR aim to channel more than $500 billion of institutional money—pension funds, insurers, sovereign wealth—into dedicated AI-compute financing pools. The capital flow runs through a specific chain: institutional investors fund a financing vehicle, which lends to an AI infrastructure project, which buys Nvidia systems, which lease compute to an end customer whose long-duration payments service the debt.
Nvidia books a GPU sale upfront. The long-term credit risk sits with third-party lenders. Goldman Sachs has said explicitly it wants to create a market for credit backed by Nvidia compute—language that describes the birth of a fixed-income asset class, with NVL racks as collateral subject to loan-to-value ratios, debt-service coverage covenants, and residual-value assumptions.
The AI-cloud model adds another revenue layer. Cloud operators like Sharon AI and Firmus buy Nvidia infrastructure. Nvidia commits to purchase some of that cloud capacity. If external customers absorb the capacity instead, Nvidia's obligation shrinks and Nvidia can receive a share of the third-party revenue. The company gets paid on the hardware sale, then again on utilization—Jensen Huang's phrase "compute is revenue" made contractually literal.
The $105 Billion Guarantee and Its Circularity
The OpenAI/SB Energy structure is the most aggressive expression of this model. Nvidia guarantees up to $105 billion of residual value and lease support across nine project phases, activating progressively between fiscal 2028 and 2030. If OpenAI defaults, Nvidia can assume the lease, find another tenant, force a sale, or pay the shortfall. OpenAI has agreed to reimburse Nvidia for any amounts paid—an indemnification whose value depends on the solvency of the entity that just defaulted.
Nvidia's own blog estimates the campus could represent roughly $600 billion of Nvidia compute through 2030. The circularity is direct: Nvidia's guarantee helps SB Energy finance the facility; the facility hosts Nvidia hardware; OpenAI buys Nvidia systems to fill it. The guarantee creates the bankability that generates the hardware demand.
S&P affirmed Nvidia's AA rating with a stable outlook, modeling an adjusted-debt equivalent of approximately $4.2 billion in 2028 rising toward $37.7 billion by 2031 before declining. The agency is treating the guarantee as probability-weighted debt, a framework that implies neither panic nor complacency about the exposure.
The Duration Mismatch Nobody Can Ignore
Private credit needs to believe GPU useful life roughly equals financing life. Nvidia itself is shipping new architectures annually—Hopper, Blackwell, Blackwell Ultra, Rubin—and Vera Rubin is already in production. If Rubin compresses Blackwell economics faster than residual-value models assume, lenders discover their collateral is depreciating semiconductors bolted to a building, and recovery rates fall precisely when tenant credit is weakening. That correlated collapse—falling GPU residuals, deteriorating tenant solvency, surplus data-center capacity appearing simultaneously—is the scenario where off-balance-sheet exposures migrate onto Nvidia's books through guarantee activations, capacity backstop payments, and weakening hardware orders.
Nvidia is partially addressing this by structuring guarantees around the site rather than the chip. A 4.25-gigawatt energized campus can host Rubin, then Feynman, then whatever follows. As 20-year collateral, energized land with interconnection rights and cooling infrastructure is far more defensible than any single GPU generation. The $279 billion in supply commitments—$92 billion for the remainder of fiscal 2027, $87 billion for fiscal 2028, $88 billion for fiscal 2029—tells a parallel story: Nvidia has pre-underwritten years of AI demand onto its own cost structure. If that demand disappoints, the first shock may arrive as excess inventory and noncancelable supplier obligations, before any customer defaults surface.
Separating the Power Asset from the Chip
Nvidia is attempting to finance GPUs and data centers inside the same long-duration capital structures, a bundling that obscures a critical mismatch. The physical site—substations, transformers, transmission rights, cooling—can support multiple generations of compute over 15 to 25 years. The GPU hardware retains premium economics for perhaps two to four years before the next architecture arrives.
The highest-conviction position in this phase of the AI buildout belongs to capital that separates those two assets cleanly. Finance power and land long. Amortize GPU hardware over the contract period that economics can actually defend, and force refinancing or refresh decisions when that period ends. Require customer prepayments or investment-grade offtake to cover a material share of GPU capital expenditure. Build refresh reserves contractually. Nebius's recent secured financing at SOFR plus 250 basis points, backed by deployed GPU infrastructure and an investment-grade customer, shows what a defensible structure looks like. So does IREN's combination of customer prepayments and sub-6% GPU financing covering roughly 95% of GPU capital costs.
Whoever builds reliable pricing for GPU rental curves, residual values, and workload migration costs—the Bloomberg of compute finance—will own the infrastructure that a $500 billion credit asset class requires to function. The smart money is buying bottlenecks and senior contracted cash flows. The residual-value risk of rapidly refreshing semiconductors is someone else's trade.
not investment advice