On the morning of August 10, six of the most powerful men in global finance sat down together in a television studio alongside Jensen Huang. Goldman Sachs chief executive David Solomon was there.
So were Blackstone president Jon Gray, Apollo president Jim Zelter and Brookfield chief executive Bruce Flatt. KKR sent Waldemar Szlezak, who runs its digital infrastructure business. Larry Fink of BlackRock joined by video link from the road.
The segment ran for more than half an hour and contained remarkably little detail. What it contained instead was a message, delivered with the theatrical confidence that has become Huang’s trademark.
Nvidia had signed memorandums of understanding (MoU) with all six firms to create what it called independent compute financing platforms, with the aim of mobilising more than $500 billion of third-party capital for the construction of AI data centres and the purchase of Nvidia hardware.
No deals had actually been signed. There is no fixed timetable. The USD 500 billion figure, as Bloomberg later reported, is a round number combining transactions already under discussion with a forecast of demand still to come. Each lender will vet borrowers individually before committing a cent.
And yet the announcement may prove to be one of the most consequential financial events of the AI era. Because what Huang was really doing was not raising money. He was proposing a new asset class.
The problem nobody could keep paying for
To understand why Nvidia needed to stand on a stage with six financiers, look at what has happened to the balance sheets of its biggest customers.
For most of the last decade, Big Tech funded its own expansion. Cloud businesses threw off enormous operating cash flow, and capital spending, however large, stayed comfortably inside it. That relationship has now broken.
Alphabet, Amazon, Meta and Microsoft have collectively guided to something close to USD 700 billion of capital expenditure in 2026, a rise of roughly three quarters on the previous year’s already record figure.
Bank of America projects aggregate hyperscaler capex will top USD 860 billion this year and approach USD 1.2 trillion in 2027. Goldman Sachs now models more than USD 5 trillion of combined capex for the big four between fiscal 2025 and fiscal 2030.
The cash consequences arrived faster than most investors expected. Alphabet posted its first negative free cash flow quarter since its 2004 listing in the second quarter of 2026, burning USD 5.9 billion as capital spending surged past USD 44 billion in three months.
It then raised the top end of its full year capex guidance by as much as USD 15 billion. Amazon’s trailing 12-month free cash flow swung to negative USD 7.6 billion after three consecutive positive years.
Research house Epoch AI, fitting growth curves to quarterly filings, calculated that aggregate hyperscaler cash capex would overtake operating cash flow somewhere around the third quarter of 2026. That crossover point is now behind us.
Microsoft remains the outlier, the only one of the American hyperscalers still generating meaningful free cash flow, and it has managed that partly by leasing rather than buying, adding some USD 26 billion of finance lease assets over four quarters rather than issuing senior bonds.
The rest have gone shopping for outside money, and at extraordinary scale. FactSet calculates that incremental annual debt has risen from 9% of hyperscaler capex in fiscal 2024 to 32% by mid-2026.
Equity has returned to the funding mix too. Alphabet priced an USD 84.75 billion raise in June 2026, the largest equity capital transaction ever completed by a listed company, including a USD 10 billion private placement with Berkshire Hathaway.
Oracle, the most leveraged of the group, raised USD 43 billion of debt and USD 5 billion of equity in fiscal 2026 and plans roughly USD 40 billion more.
Then there is the arithmetic that hangs over the whole sector. Morgan Stanley’s widely circulated estimate puts global data centre capital expenditure through 2028 at around USD 2.9 trillion, against hyperscaler operating cash flow capable of covering perhaps USD 1.4 trillion of it.
The remaining USD 1.5 trillion has to come from somewhere else. In Morgan Stanley’s own bridge, the largest single share, about $800 billion, is allocated to private credit, with roughly USD 200 billion from corporate bonds and USD 150 billion from securitised products.
That USD 1.5 trillion hole is the reason six financiers were sitting in a television studio in August.
Why Nvidia cannot simply write the cheque
Nvidia is not short of money. It reported record revenue of USD 81.6 billion in the first quarter of fiscal 2027, up 85% year on year, with data centre revenue of USD 75.2 billion and gross margins around 75%.
It has authorised a further USD 80 billion of share buybacks and raised its dividend 25-fold. Its market capitalisation sits around USD 5.5 trillion.
But Huang has said publicly that AI infrastructure spending could reach USD 3 trillion to USD 4 trillion a year by the end of the decade. At that scale, no single corporate balance sheet is adequate, including his own.
There is a second problem, and it is arguably more urgent. Nvidia’s growth increasingly depends on customers who are not hyperscalers. Frontier laboratories such as OpenAI and Anthropic, specialist AI clouds, sovereign projects and enterprises want compute at scale, but many of them lack the credit rating or the cash to buy millions of dollars of silicon outright.
Meanwhile the hyperscalers, Nvidia’s traditional customers, are busy designing their own accelerators. Broadening the buyer base is a strategic necessity, and the constraint on that broadening is no longer chip supply or data centre shells. It is financing.
Nvidia’s earlier attempts to solve this itself produced exactly the reaction it feared. The company has invested in customers including CoreWeave, contributed billions to an OpenAI funding round, and joined a consortium backing xAI. Analysts began describing the pattern as circular financing, the vendor funding its own demand, and comparisons to the telecom vendors’ financing collapse of the dot com era followed quickly.
The reaction sharpened when reports emerged that Nvidia was weighing a USD 250 billion guarantee for an OpenAI data centre project in Ohio. Nvidia shares fell 5%, and the price of credit default swaps on Nvidia bonds recorded their largest intraday move since they began trading actively. The company subsequently trimmed that guarantee to under USD 120 billion, covering only the first phase.
Seen against that background, the six-way partnership is a deliberate correction. Nvidia will still provide credit support, but Huang clarified after the announcement that its guarantees would cover as much as 25% of an opportunity, assessed project by project.
The other 75%, and the origination, structuring, distribution and warehousing of the risk, belongs to Wall Street. The chipmaker keeps the demand and sheds most of the balance sheet.
The intellectual move at the centre of the deal
Huang’s contention is that a rack of Nvidia GPUs should be treated the way a lender treats a warehouse, a toll road or a power station.
In his framing, Nvidia compute is an investable infrastructure asset, productive, revenue generating and fungible across the entire market.
Nvidia’s own statement described its compute as broadly adopted, transferable between customers and operators, and continuously improved by CUDA software updates that extend its useful life.
If that classification holds, everything else follows. Loans can be secured against the hardware itself alongside the offtake agreements that guarantee its use. Special purpose vehicles can own chips and lease them to Nvidia’s customers, keeping the debt off the customer’s balance sheet and off Nvidia’s.
Those vehicles can then issue bonds, some expected to run to tens of billions of dollars each. If a borrower fails, the chips can be re-rented to somebody else, which limits the damage from any single default. Insurance capital, pension money and sovereign wealth funds can buy the resulting paper, because it looks and behaves like infrastructure debt.
If the classification does not hold, the whole edifice is a very large pile of fast depreciating electronics dressed up as real estate.
The case against
An H100 that changed hands for roughly $30,000 in 2023 was trading at around $8,000 by the middle of 2026, a fall of about 73% in three years. Hourly rental rates for the same chip peaked near USD 8, collapsed to between USD 1 and USD 2 as supply arrived, recovered, then softened again.
CUDA’s ecosystem of more than six million developers may guarantee that a buyer exists for repossessed hardware. It does not guarantee the price.
Michael Burry, who made his name calling the last credit crisis, has attacked the depreciation schedules underpinning the sector, arguing that a two-to-three-year hardware upgrade cycle cannot support five- and six-year useful life assumptions, and estimating that understated depreciation could distort reported earnings by around USD 176 billion between 2026 and 2028.
Accounting specialists have pushed back on the strongest version of that claim, but the debate has moved from technical footnotes to the front of investor decks.
Then there is China. Bernstein Research expects Nvidia’s share of the Chinese AI chip market to collapse from roughly 40% to around 8% by the end of 2026, with Huawei approaching half the market. Should Chinese production flood the world with cheap compute, the collateral behind these loans could erode faster than the loans amortise.
One analyst estimate suggests investors will price GPUs as high depreciation equipment rather than property, and demand yields of 11% to 17% depending on their position in the capital structure. That is high yield pricing, and it sits well above what a hyperscaler pays in the corporate bond market.
Rating agency methodology for GPU backed securitisations, meanwhile, is still being worked out. Fitch has yet to publish a settled approach.
What Wall Street actually gets
Fees, and a lot of them. Alternative managers earn management fees on committed capital, typically 1.5% to 2%, plus carried interest on profits. Fee related earnings are what analysts prize, because they are recurring and predictable.
Apollo reported record fee related earnings of USD 785 million in the second quarter of 2026, up 25% year on year, on USD 74 billion of originations. Strikingly, that figure excluded the USD 35 billion Broadcom AI infrastructure financing entirely, because Apollo books volume at closing rather than announcement, leaving roughly USD 50 billion of signed deals to feed later quarters.
Management has also noted a shift towards structures that recognise fees across multiple quarters or years rather than upfront, smoothing earnings in a way public shareholders reward. Goldman, the only participant with a full investment banking apparatus, collects the underwriting and distribution economics on top.
A home for permanent capital. The deeper motivation is a liability problem. The five largest listed alternative managers now oversee about USD 1.5 trillion of perpetual capital, roughly 40% of their combined assets, much of its insurance and annuity money gathered through platforms such as Apollo’s Athene, KKR’s Global Atlantic and Blackstone’s insurance mandates.
Annuity liabilities are long dated and require long dated, contracted, investment grade style assets to match them. Those assets are scarce. A twelve-year lease on a GPU cluster with an investment grade offtaker attached is, in principle, exactly the instrument these balance sheets are hungry for.
Apollo’s private credit assets alone stand at roughly USD 405 billion, Blackstone’s credit and insurance arm at about USD 465 billion, BlackRock at around USD 220 billion after its HPS and GIP acquisitions, and KKR at about USD 140 billion. All of that money needs somewhere to go.
Ownership of a new market at its inception. Asset classes are created rarely. Whoever writes the first documentation, sets the advance rates, defines the residual value assumptions and builds the ratings dialogue tends to own the league tables for a decade.
Data centre securitisation issuance ran near USD 27 billion in 2025 and is projected by JPMorgan at USD 30 billion to USD 40 billion annually in 2026 and 2027, a rising share of the combined asset backed and commercial mortgage-backed market.
CoreWeave has already priced an USD 8.5 billion investment grade rated GPU collateralised transaction. Nvidia has now handed six firms a franchise position in the market that follows.
Better risk for the same yield. Nvidia’s willingness to backstop up to a quarter of a transaction materially changes the credit maths. A lender writing a loan against hardware alone is exposed to residual value.
A lender writing the same loan with a first loss cushion from a company with 75% gross margins and a $5.5 trillion market capitalisation is in a different business. Combine that with collateral that mixes the chips themselves with contracted offtake, and with the ability to re-rent hardware to a different tenant on default, and the risk adjusted return starts to look attractive even at spreads well inside 11%.
Distribution, which is where the real prize sits. These firms do not intend to hold the paper. They intend to originate it and sell it. Executives are already sounding out sovereign wealth funds, pension schemes and insurers, and indicated during the announcement that some of the capital could come from retail investors.
That last point matters more than it sounds. American regulators have recently opened the roughly $13 trillion defined contribution market to private credit managers, while Europe’s revised ELTIF regime has broadened what long term investment funds may hold.
Non traded business development companies and evergreen vehicles are growing quickly. A manufacturing line for long dated, contracted, AI linked credit feeding those channels is a business with obvious compounding characteristics.
Apollo is expanding a trading operation to sell down chunks of what it originates and make markets in the paper afterwards, which adds a second fee layer.
Adjacency. The financing will not be a single product. As Mercer’s global head of real assets observed after the announcement, the partnerships are likely to spawn strategies across infrastructure, real estate credit and possibly private equity, giving investors multiple access routes. Data centres need land, power, transmission, cooling and construction finance. A firm that anchors the compute layer is well placed to sell the rest.
Competitive necessity. Nobody wanted to be left out. Huang has said he approached only these six and none refused. Within minutes of the announcement, Morgan Stanley published a framework to facilitate USD 1.5 trillion of funding for American innovation and national security, with AI and advanced computing at the top of the list.
JPMorgan’s asset management arm is reportedly discussing how to participate. Broadcom set the template weeks earlier, tapping Apollo and Blackstone as anchor investors for more than 20 gigawatts of compute for frontier laboratories through 2028, with USD 35 billion already committed and the borrowing structured to sit off Broadcom’s balance sheet.
Where the win-win could break
The mutuality depends on one assumption holding for a decade. Chips must remain productive long enough, and generate enough revenue, to service the debt raised against them.
Apollo’s own published view illustrates the tension. The firm has argued that more than $5 trillion of expected data centre capital expenditure implies USD 1.5 trillion to USD 2 trillion of annual AI revenue by 2030, against USD 40 billion to USD 60 billion today. That is the gap the entire structure is betting will close.
The risk is no longer confined to technology shareholders. It now runs through special purpose vehicles, private credit originators, securitisation trusts and ultimately into pension portfolios and insurance reserves.
Insurance regulators have already tightened capital treatment for collateralised loan obligations and overhauled how collateral loans are charged, moving from a flat charge to a framework tied to what actually backs the loan.
American law firms are circulating client alerts on litigation risk in AI data centre financing. The Federal Reserve Bank of Chicago has noted that direct bank exposure to AI adjacent industries averages under 1% of assets, while cautioning that indirect exposure through lending to private credit funds is harder to see.
One person close to the announcement described Huang’s intention as building a debt shopfront, an advertisement aimed simultaneously at customers and at nervous investors. If the deals do not materialise as promised, or if they sour, the reputational cost lands on all seven names.
The final take
Nvidia has done something clever. It has kept the demand, capped its exposure at roughly a quarter, and persuaded the deepest pools of capital in the world to carry the rest.
Wall Street, for its part, has been handed a manufacturing line for exactly the kind of long dated, contracted, high yielding asset its insurance balance sheets and retail distribution channels have been starved of.
Both sides get what they want. Whether the arrangement is a win for the pensioners and policyholders who end up owning the paper depends entirely on a question none of the seven firms on that stage could answer, which is how long a graphics processor stays valuable.