In August 2026 NVIDIA said it had signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish compute financing platforms intended to mobilize more than 500 billion dollars of third-party capital. The agreements are, in NVIDIA's own words, subject to execution of the final agreements. Read past the headline number and there is a structural change worth an AI vendor durability assessment on your side of the table.

Compute Stopped Being a Purchase and Became a Financing

The mechanics are ordinary once stated plainly. Six of the largest pools of institutional capital in the world set up platforms that independently underwrite AI infrastructure, so that data centres and the hardware inside them can be funded by investors rather than paid for out of an operator's cash. NVIDIA describes the intent as dedicated pools of capital at significant scale at attractive rates for its customers, and describes what it gets back as long-duration usage-linked revenue.

That last phrase is the one to sit with. Long-duration means the capital is repaid over years. Usage-linked means repayment depends on the compute actually being used and paid for. Between those two words sits an assumption about demand holding up over the life of the financing.

None of this is unusual as a category. Telecommunications networks, power generation and shipping fleets are all financed rather than bought outright, precisely because they are capital-heavy assets with long lives. What is new is that the thing being financed sits underneath software you are buying on an annual contract.

Why This Changes What Vendor Risk Means

Most vendor due diligence asks whether a supplier can build the thing and keep it running. Financed capacity adds a question that procurement checklists rarely contain: what happens to your service if the economics underneath it move.

The exposure is indirect and that is exactly why it goes unexamined. You do not contract with the financing platform. You contract with a software vendor, which rents capacity from an operator, which may be servicing capital raised against the expectation of sustained utilisation. A change three steps up that chain reaches you as a price change, a capacity constraint, or a quiet deprioritisation of smaller accounts.

This is not a prediction that anything goes wrong. Long-lived infrastructure has been financed this way for a century. It is an argument that the question belongs in your assessment, in the same way you already ask where a supplier's data sits without expecting a breach.

Four Questions for an AI Vendor Durability Assessment

These are answerable in a normal vendor conversation. A supplier who cannot answer any of them is telling you where their own visibility ends.

  1. Whose capacity are you actually selling me? Own hardware, a reserved commitment with a cloud, or spot access resold. The three behave completely differently when supply tightens, and only the first is under the vendor's control.
  2. What is your price protection, and for how long? A one-year rate on a multi-year dependency is a repricing event with a date on it. Ask what happens at renewal if their input cost has moved, and get the answer before you build a workflow that assumes today's number.
  3. Where do I sit if capacity is rationed? Every provider has an implicit priority order and very few publish it. The useful version of this question is concrete: if the vendor is short next quarter, which customers get served first, and what tier am I in.
  4. What is my exit, mechanically? Not whether a contract permits leaving, but how long a migration actually takes, whether your prompts, evaluations and fine-tunes are portable, and what you would lose. A dependency you cannot leave in a quarter is a dependency you have to underwrite yourself.

The pattern behind all four is the same. Capability questions dominate AI vendor selection right now because capability is what visibly differs. Continuity questions are the ones that decide whether the thing you selected is still there in two years at a price you planned for.

What Not to Read Into It

Two guards, because this subject attracts more heat than it deserves.

These are memorandums of understanding, not closed transactions, and NVIDIA says so directly. Treating an announced intention to mobilize capital as capital already deployed would be the wrong reading, and any analysis you build on the larger number should carry that caveat.

Financing infrastructure is also not evidence of fragility. The presence of Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR is a sign that these assets are considered underwritable, which is closer to the opposite conclusion. The reason to pay attention is not that something is wrong. It is that a cost structure two or three steps removed from your contract now has a shape, and shapes can be asked about.

Sources

  1. NVIDIA, "NVIDIA Partners With Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Establish AI Compute Infrastructure Financing Platforms," 2026. Link.
  2. CNBC, "Nvidia lines up 500 billion dollars in financing as CEO Jensen Huang tells CNBC his chips are an investable asset," 2026. Link.

Next Steps

If your AI roadmap depends on capacity you have never traced past your direct supplier, the four questions above are a short conversation that closes a long gap. Stable Solutions runs vendor assessments that cover continuity and exit alongside capability, and builds the portability in while the system is being designed. Explore our Digital Growth Strategies or contact our team to review your AI vendor exposure.