Six names signed the memorandum: Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR. The figure on the table, more than $500 billion of third-party capital, exceeds the total global venture investment in most years. And the asset being financed is compute.
The move reframes graphics processing units (GPUs) from depreciating hardware into a long-lived, revenue-generating asset class, financeable on terms usually reserved for physical infrastructure.
Execution risk is real: these are memoranda, not closed deals, and the credit market for compute has no track record yet.
What a compute financing platform actually is
The mechanics matter more than the headline. Nvidia is not lending the money itself. Under the partnerships announced on August 10, each financial institution creates a dedicated pool of capital, its own platform, that its customers can draw on to buy or lease compute. Nvidia's role is to certify the asset's value and bring the offtakers.
This is the part worth sitting with. A chip vendor is organising the financing of its own product. Historically, that product was sold and the buyer carried the risk of it becoming obsolete in 18 months. Now the risk is being spread across balance sheets built for exactly that job.
The pitch is that compute behaves like infrastructure, not like consumer electronics. In AI, compute is revenue: the lowest-token-cost, highest-revenue, longest-lived asset in the stack, supported by a customer and offtaker base built on CUDA, its parallel-computing software layer.
We began by building chips; today, we are helping create a new class of productive, investable infrastructure: AI factories.— Jensen Huang, founder and CEO, NVIDIA
Why the asset-class framing matters now
The framing is the story. Chips depreciate; asset classes get financed and traded. Huang is asking the world's largest infrastructure investors to treat a rack of GPUs the way they treat a toll road: a long-lived productive asset with cash flows and residual value.
That reframe changes what capital is available. Infrastructure money is patient. Decades, not quarters. If compute qualifies as infrastructure, it stops competing for the same venture dollars as every other software company and starts drawing from pension-fund and insurance pools that never touch a Series A.
Target pool across six financing platforms
Mobilized over time for AI infrastructure buildout, per the August 10 memorandum · NVIDIA press release, 2026
Asset managers signing memoranda
Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, each building an independent platform · NVIDIA press release, 2026
The mechanics carry a specific design choice: usage-linked revenue. A customer pays for compute as it is used, which gives the financing platform a cash-flow profile closer to a utility than to a one-off hardware sale. That is the property institutional capital is built to underwrite.
The six signatories and what each brings
Three of the six disclosed their books. Apollo reports roughly $1.05 trillion in assets under management. Blackstone runs over $1.3 trillion. Brookfield manages more than $1 trillion. Goldman Sachs and KKR describe their roles in investment and distribution; KKR's co-CEOs note that Nvidia is a founding investor in its Helix Digital Infrastructure platform, and BlackRock's Larry Fink ties the move to the AI Infrastructure Partnership Nvidia already participates in.
| Firm | Reported AUM | Role named in the memorandum |
|---|---|---|
| Apollo | ≈ $1.05 trillion | flexible, long-term capital |
| Blackstone | > $1.3 trillion | infrastructure investing |
| Brookfield | > $1 trillion | core infrastructure pillar |
| BlackRock | n/d | connecting long-term capital to infrastructure |
| Goldman Sachs | n/d | credit market backed by compute |
| KKR | n/d | long-duration capital, Helix partnership |
Read the quotes and a shared thesis appears. Apollo's Jim Zelter calls modern compute "a scarce, mission-critical asset class." David Solomon describes the ambition as creating "a market for credit backed by NVIDIA compute." That is a novelty: a credit market for a technology asset, constructed before any such market existed.
Where the credit market for compute stands
Financial Times describes the effort as one of Wall Street's most ambitious lending programs. That cuts both ways. The scale is unprecedented, and so is the absence of precedent.
The first open question is the memorandum itself. These are agreements to negotiate, not agreements signed. The release itself carries the standard disclaimer that the partnerships remain subject to execution of final agreements.
Second is the residual-value problem. A financing platform that lends against compute needs confidence in what a GPU is worth after four or five years of service. The GPU generations turn over quickly; the collateral does not stay static. Infrastructure lenders have a century of valuation data for physical assets. For silicon, the data is a few years old.
Third is concentration. The asset being financed is, at the core, one vendor's platform. If a competitor's chip wins a generation, the collateral base of these platforms shifts. That is a risk the lenders are consciously taking; the CUDA moat is exactly what makes the asset "fungible and transferable" in the company's own language.
Why lenders are comfortable underwriting compute
The counterpoint: adoption breadth is a claim, not a balance-sheet fact, and it has not survived a full depreciation cycle yet.
The investor takeaway is simpler than the mechanics. When the world's largest asset managers sign up to treat compute as infrastructure, the distinction between technology capex and infrastructure capex starts to blur. For a high-net-worth reader evaluating private opportunities in AI factories, GPU clouds and their financiers, the meaningful shift is that a whole new pool of long-duration capital is now structurally committed to this asset class, assuming the platforms close.
And that is the whole question in one sentence.
How much of the $500 billion becomes real capacity by 2028?
Probability: 55%. The signatories already operate comparable infrastructure platforms, but the memoranda phase and residual-value uncertainty argue against a faster conversion.
✅ Arguments for
The six institutions already run multi-trillion-dollar infrastructure platforms with long-duration capital to deploy.
Usage-linked revenue gives platforms a cash-flow profile lenders can underwrite today.
Confirmation criteria: first platform reaches financial close with a named anchor customer within 12 months.
❌ Arguments against
No residual-value track record exists for GPU collateral across a full depreciation cycle.
A single competitive chip generation could shift the collateral base of all six platforms.
Disconfirmation criteria: no platform closed within 18 months, or a major signatory publicly backs away.
First financial close of any platform, and the identity of its anchor customer.
Pricing of the first compute-backed credit issuance: the spread tells you how the market values residual risk.
Whether a non-Nvidia chip platform (AMD, custom ASICs) attracts comparable financing.
The depreciation assumptions lenders accept in final agreements: the real measure of the asset-class claim.
Development scenarios
🟢 Optimistic scenario (35%)
Implications: private compute assets become a standard institutional allocation, and the gap between AI-lab balance sheets and their compute needs narrows.
🟡 Base-case scenario (45%)
Implications: compute financing becomes a boutique infrastructure niche rather than a new market, with pricing still set by GPU supply.
🔴 Pessimistic scenario (20%)
Implications: the asset-class claim recedes, and compute stays a capex line on AI-lab balance sheets rather than a financed infrastructure pool.
As we wrote in August, Nvidia's grid-tied battery storage line was self-qualifying as standard AI factory infrastructure. The financing platforms run the same play one level up, organizing the capital structure around compute itself. Whether the market accepts compute as collateral is now the test, and the six signatures on the memorandum are the market's first answer.