What Was Actually Signed
On August 10 Nvidia announced memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, setting up compute financing platforms to mobilise over $500 billion of third-party capital.
Read the structure, not the headline. There are six independent platforms, one per institution, lending against GPU compute as collateral the way an infrastructure lender treats a power station or an office block. Jensen Huang called it an investable asset class, which is exactly what the collateral treatment implies.
Nvidia attaches two caveats and they're worth keeping: the partnerships are subject to final agreements, and the money deploys over time. None of this lands as orders this quarter.
Why a Financing Story Belongs on a Sourcing Desk
Component demand has always had a ceiling. A buyer can only build what it can pay for, and even hyperscaler cash flow runs out somewhere.
Off-balance-sheet financing moves that ceiling. Frontier labs, enterprises and AI clouds that couldn't have committed to a multi-billion-dollar buildout last year can now do it the way a utility funds a plant. The pool of credible buyers for your constrained parts just got deeper, and what they can spend is no longer bounded by what they earn.
Demand was never the shaky part of this forecast anyway. TrendForce puts combined capex at the nine largest cloud providers near $886.7 billion, roughly 90% above 2025, with five North American hyperscalers accounting for close to 90% of that. Their 2027 figure is about $1.32 trillion.
What the Number Tells You About Duration
Nobody builds permanent financing infrastructure for a shortage they expect to clear.
Six of the largest asset managers on earth do not stand up lending platforms around an asset class they think gets commoditised in eighteen months. The capital markets have looked at AI compute and priced it as multi-year and structurally supply-short. That is a longer view than most procurement plans currently carry.
The Assumption This Breaks
Plenty of 2027 plans quietly assume competitive pressure eases once rivals hit their budget limits.
That assumption now has a counterparty list attached to it, and the counterparties are Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR. If your plan works only because you expect the other bidder to run out of money, it doesn't work.
What Buyers Should Do Now
- Start here: Stop sizing constrained-part risk against competitor budgets. The question worth asking is who holds written allocation, and whether you're on that list.
- Within the month: Turn critical memory, HBM-adjacent and enterprise SSD demand into manufacturer-confirmed allocation with committed ship windows, mapped to exact package, density and approved die. Verbal ATP isn't coverage here.
- Before you commit 2027: Re-run the supply assumptions with the funding constraint taken off the demand side, and see whether the plan still stands up.