Nvidia’s $500B AI funding push puts hidden leverage back on the tape
Goldman pegs hyperscaler lease commitments at $1.5T, while a hedge-fund unwind shows how fast AI exposure can delever.
Nvidia is working with major Wall Street firms to mobilize more than $500 billion of third-party capital for AI infrastructure, reframing compute buildouts as financeable projects rather than just corporate capex. The pitch is landing as markets focus on where AI leverage actually sits, from off-balance-sheet lease commitments to prime-brokerage borrowing that can turn crowded equity themes into forced selling.
Key Takeaways
- Nvidia is partnering with large asset managers and banks to mobilize more than $500 billion of third-party capital for AI infrastructure.
- Goldman Sachs estimates hyperscalers’ lease commitments at $1.5 trillion, including about $1 trillion of “uncommenced” obligations that are not yet reflected in financial statements.
- AI-focused hedge fund Situational Awareness shrank from $45 billion to about $10 billion after leveraged, concentrated equity bets triggered margin calls, with Citadel later buying listed positions at a discount.
- PIMCO’s Lotfi Karoui cited consensus forecasts for hyperscaler capex to exceed $1 trillion per year from 2027, while calling the ultimate scale of the buildout “deeply uncertain.”
Nvidia’s $500B+ capital mobilization reframes chips as financeable infrastructure
Nvidia is partnering with Wall Street firms to mobilize more than $500 billion of third-party capital for AI infrastructure, a move that pulls the AI buildout closer to the playbook used for power, pipelines, and other long-duration projects. The mechanism is straightforward: instead of hyperscalers funding everything directly on balance sheet, capital can be raised and warehoused in vehicles that own infrastructure and lease capacity back to operators. The consequence is that AI compute starts to look less like a line item and more like an asset class.
Nvidia’s plan to develop AI infrastructure platforms names Apollo, Blackstone, BlackRock, Brookfield, KKR, and Goldman Sachs as partners, and it could involve private asset-like structures and asset-based financing. Asset-based financing means borrowing secured by specific assets, here the infrastructure and equipment, rather than only the borrower’s general credit. Nvidia CEO Jensen Huang framed the shift bluntly, calling Nvidia’s chips an “investable infrastructure asset.”
That framing matters because it invites more leverage and more intermediation. If compute is treated as financeable infrastructure, the market will naturally reach for structures that separate ownership, operation, and funding, which can lower the apparent cost of capital in good times and compress visibility when conditions tighten.
The hidden leverage bucket: $1.5T lease commitments and $1T uncommenced obligations
The cleanest example of “less visible until it isn’t” leverage is leasing. Goldman Sachs analysts estimated hyperscalers’ combined lease commitments for data centers, R&D facilities, offices and equipment at $1.5 trillion, up from about $200 billion five years ago. About $1 trillion of that total is “uncommenced” lease commitments, meaning obligations have been agreed but have not started yet, so they are not yet shown in financial statements even though they imply future payments.
The timing is the point. Uncommenced commitments become real cash outflows when leases commence, and they can shift reported leverage and liquidity needs quickly if multiple projects flip from “signed” to “in service” around the same window. Goldman’s warning was explicit: uncommenced lease commitments “can understate leverage and future liquidity needs as these obligations are eventually recognized and contractual payments come due,” the analysts wrote in an Aug. 6 note.
Joint ventures and leasing vehicles are part of the same toolkit. Some tech giants are using joint ventures and other leasing structures to borrow for AI data center spending without the debt appearing on balance sheets until leases begin. That works as long as funding stays cheap and demand assumptions hold. It gets messy when the market reprices credit or when the buildout’s utilization curve disappoints, because the obligations still arrive on schedule.
How AI exposure gets levered twice: corporate structures plus prime-broker and derivatives overlays
The AI trade can be levered at two different layers, and the layers can interact. The first layer is corporate and project finance: bond markets, joint ventures, leases, and asset-based financing used to fund data centers and equipment. The second layer is investor leverage: hedge funds and other investors using prime brokerage borrowing and derivatives to amplify returns on AI-related equity exposure.
Prime brokerage borrowing is financing provided by large broker-dealers that lets funds run larger positions than their cash would allow. A margin call is what happens when those leveraged positions lose value and the lender demands more collateral. In a crowded theme, that can turn a drawdown into a cascade because selling is driven by financing constraints, not by a fresh fundamental view.
Situational Awareness is the stress case markets are now using to map that cascade. The AI-focused hedge fund’s assets fell from $45 billion to about $10 billion after losses on leveraged, concentrated equity bets, and it was unable to meet a series of margin calls from lenders. Its portfolio was heavily concentrated and included names such as SK Hynix and CoreWeave. Citadel later bought Situational Awareness’ publicly listed positions at a discount, and SK Hynix and CoreWeave have since rallied.
The episode is not proof of systemic fragility by itself, but it is a clean illustration of the mechanism traders care about: leverage plus concentration plus volatility can force liquidation at the worst possible time, and the unwind can happen even if the underlying names bounce afterward.
Leverage vs expectations: the debate over what breaks first in the AI trade
The market argument is now split between two failure modes. One camp is focused on leverage and the speed of unwind. JPMorgan CEO Jamie Dimon said margin debt is “pretty high,” adding that it increases the risk of amplified volatility. In that framing, the risk is less about whether AI is real and more about how much exposure is financed and how quickly lenders tighten.
The other camp thinks the nearer-term risk is simpler: expectations. Sahil Mahtani, director of the investment institute at Ninety One, said expectations “of high and rising earnings in the years ahead” were “the main risk” the AI trade posed to markets. “That is primarily an expectations problem rather than a leverage problem,” he wrote via email. He also called equity concentration “historically high” in tech-heavy markets, especially the U.S., arguing that concentration can behave like leverage by amplifying index moves when the biggest weights fall. “The big equity indices are extremely concentrated, and no one thinks anything could possibly derail them,” Mahtani said, warning that a shift from buybacks to issuing more stock could remove a source of support as AI-related valuations are already under pressure.
On hedge funds specifically, the industry pushback is that leverage is not automatically a systemic threat. A spokesperson for the Alternative Investment Management Association said leverage is a “core tool” used by hedge funds to boost returns and provide market liquidity. “The key question is not whether hedge funds use leverage, but whether its use poses a material threat to financial stability. The available evidence does not support treating hedge fund leverage as an inherent systemic risk,” the spokesperson said. They also cautioned against treating different blowups as the same structure, saying, “It is important not to lump very different market events together,” and adding that they saw no reason to expect the Situational Awareness episode “to trigger a fresh review of the rules governing hedge fund leverage.”
My take: the tradable risk is the speed of recognition, not just the size of the buildout
The threshold that matters is when commitments stop being narrative and start being cash flow. Nvidia can call chips an “investable infrastructure asset,” and consensus can project hyperscaler capex above $1 trillion per year from 2027, but the market trades the path, not the destination. Goldman’s $1.5 trillion lease estimate, with roughly $1 trillion uncommenced, is the kind of number that can stay quiet right up until commencement schedules cluster and liquidity math becomes visible.
The real test is whether the financing wrapper stays boring under stress. If Nvidia and its partners formalize the platform with asset-based financing and private-asset-like vehicles while lease disclosures and credit spreads stay orderly, the setup starts to look structural rather than sentiment-driven. If prime brokers tighten and discounted block sales reappear, the AI trade’s downside will be defined by forced selling speed, not by the long-run size of the buildout.
