NVIDIA turns the AI factory into an asset class with $500 billion financing push

NVIDIA has partnered with six of the world's largest institutional investors to create independent financing platforms designed to mobilise more than $500 billion of third-party capital for AI infrastructure, a structure that moves the buildout of compute capacity away from corporate balance sheets and towards the long-term capital markets that have historically funded power grids, ports and toll roads.

The partnerships, announced earlier this month, involve Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, and are intended to serve AI labs, enterprises and cloud operators that have compute demand without the balance sheet capacity to fund it. The $500 billion figure represents aggregate capital the platforms are designed to mobilise over time rather than committed funds, NVIDIA revenue or a commitment to any single customer.

"We have moved from an era in which companies bought chips and built data centres project by project to one in which AI factories can be financed as productive infrastructure, with repeatable platforms, long-term institutional capital and a diverse customer base that uses compute to create revenue," said Jensen Huang, chief executive of NVIDIA.

The structural significance sits in what the arrangement asks the capital markets to accept. Infrastructure investors underwrite assets with predictable cash flows and long useful lives, which is why they have historically favoured toll roads over technology hardware. Persuading them that a rack of accelerators belongs in the same category as a gas turbine requires an argument about durability that the semiconductor industry has never previously had to make at this scale.

Depreciation assumptions carry the weight of the entire structure

NVIDIA's case rests on the claim that its installed base remains commercially productive well beyond the depreciation schedules applied to it. The company points to the Ampere-based A100, introduced in 2020 and still in active commercial use six years later for training, fine-tuning, inference and high-performance computing, with customers continuing to commit capacity for multi-year deployments that extend the product's economic life towards a decade.

Software carries much of that argument. "Every generation of NVIDIA software improves the performance, efficiency and total cost of ownership of already-installed infrastructure," Huang explained, describing a compounding effect through which CUDA optimisations allow older hardware to produce more output at lower cost across its operating life.

The claim matters because useful-life assumptions have become one of the most scrutinised variables in the AI trade. Short sellers and equity analysts have spent much of the past year questioning whether hyperscalers depreciating accelerators over five or six years are overstating earnings, and whether a two-year product cadence renders prior generations commercially marginal faster than the accounting allows. Institutional financing platforms convert that accounting debate into an underwriting one, where the answer determines whether a project clears its return hurdle.

Rental pricing has become the evidence base for residual value

The market data NVIDIA has put forward runs against the depreciation bears. One-year H100 rental pricing rose from roughly $1.70 per GPU-hour in October 2025 to about $2.35 by March 2026, while cross-provider on-demand median pricing moved from approximately $2.00 to $2.70 per GPU-hour between October 2025 and June 2026. Blackwell capacity commands a substantial premium, with reported B200 cloud rates spanning roughly $5.30 to $7.05 per GPU-hour.

Rising rental rates for a two-generation-old part suggest scarcity rather than obsolescence, which is precisely the condition residual-value underwriting requires. Fungibility does similar work in the argument, since capacity built for one tenant can be redeployed to another operator when requirements change.

"One NVIDIA AI factory can serve many customers and many workloads," Huang noted, pointing to an architecture adopted across every major cloud provider, systems manufacturer and enterprise deployment as the mechanism that gives the asset a deep secondary market of potential offtakers.

The circularity question changes shape under independent underwriting

Scrutiny of vendor-adjacent financing has intensified across the sector, and NVIDIA has addressed it directly. Under the platforms, the financial institutions independently assess each opportunity, evaluating the customer, demand, utilisation, cash flow and residual value, with NVIDIA supplying the platform rather than the capital.

The company has disclosed one qualification. NVIDIA may provide a residual-value support mechanism covering up to 25% of an opportunity, assessed on a project-by-project basis, which it describes as substantially lower than comparable compute-financing arrangements in the market. That contingent exposure sits with the vendor whose hardware determines the residual value being supported, and the industry will judge the platforms partly on how visible that exposure remains as deals accumulate.

"Our role is to help unlock a very large pool of independent capital while maintaining disciplined risk exposure," Huang stated, positioning the support as a complement to independent underwriting.

Access to capital, rather than access to chips, now separates the field

The competitive implication reaches well beyond NVIDIA's own order book. Supply constraints have eased relative to the acute shortages of the past two years, and the binding constraint for most buyers has become financing cost and tenor. Frontier labs backed by hyperscaler capital and sovereign wealth have never faced that constraint; the tier below them consistently has.

"The demand for AI infrastructure is extraordinary," Huang said, acknowledging that access to capital remains uneven across the companies, enterprises and AI clouds that need compute at scale.

Gulf sovereign AI programmes stand among the clearer beneficiaries of a repeatable financing template, given that the region's national compute strategies already pair long-horizon state capital with the institutional investors named in these partnerships. National AI buildouts in the UAE and Saudi Arabia have been structured around exactly the kind of long-lived infrastructure economics these platforms are designed to underwrite, and a standardised route to third-party capital lowers the execution risk on programmes that have so far depended heavily on balance-sheet commitment.

The real test comes when the assets perform without the vendor behind them

For the wider ecosystem, the arrangement signals that AI compute is being priced, underwritten and syndicated by the same institutions that finance energy and transport, which brings both discipline and consequence. Infrastructure capital demands contracted revenue, transparent utilisation and defensible residual assumptions, and projects that cannot demonstrate them will fail to secure funding regardless of how strong headline demand appears.

Whether the model holds will become apparent when the first cohort of financed factories reaches refinancing, and when residual values are tested against a market rather than a projection. NVIDIA has staked a position on that outcome, and so, now, have six of the largest pools of private capital in the world.

"This is the beginning of an open capital market for AI infrastructure," Huang added.