By Maksym Misichenko · CNBC ·
What AI agents think about this news
The panelists agree that the rapid increase in hyperscaler lease commitments and AI infrastructure spending poses significant risks, particularly around revenue justification, liquidity, and potential obsolescence of hardware. However, they disagree on the severity and immediacy of these risks.
Risk: The inability of AI revenues to justify the massive and increasing capex, leading to potential forced asset sales and margin crush across the S&P 500.
Opportunity: Durable cash-flow-generating AI compute demand from hyperscalers, which could deliver upside if AI adoption continues.
This analysis is generated by the StockScreener pipeline — four leading LLMs (Claude, GPT, Gemini, Grok) receive identical prompts with built-in anti-hallucination guards. Read methodology →
As Nvidia partners with Wall Street firms to mobilize more than $500 billion of third-party capital for AI infrastructure, the increasingly complex financing underpinning the boom is coming under closer scrutiny.
Hyperscalers and their financial backers are turning to bond markets, joint ventures, leases and other structures to fund an unprecedented infrastructure buildout. At the same time, leverage is increasing investors' AI exposure, as hedge funds and other investors use prime brokerage borrowing and derivatives to amplify returns on their bets on the boom.
But the unravelling of AI-focused hedge fund Situational Awareness, after losses on its leveraged equity bets, is heightening concerns over how much is being borrowed, where it sits, how visible it is, and how quickly it could unwind, as market watchers debate whether revenues justify the scale of spending.
Nvidia's plan to develop platforms for AI infrastructure in partnership with Apollo, Blackstone, BlackRock, Brookfield, KKR and Goldman Sachs could involve private asset-like structures and asset-based financing. Nvidia CEO Jensen Huang told CNBC Monday that Nvidia's chips are now an "investable infrastructure asset."
Some tech giants are using joint ventures and other leasing vehicles to borrow money for AI data center spending — without the debt appearing on their balance sheets until the leases begin.
Goldman Sachs analysts estimated that hyperscalers have combined lease commitments for data centers, R&D facilities, offices and equipment of $1.5 trillion, up from about $200 billion five years ago. This includes about $1 trillion of "uncommenced" lease commitments, which are not yet shown in financial statements but will result in future payments.
This "can understate leverage and future liquidity needs as these obligations are eventually recognized and contractual payments come due," Goldman analysts said in the Aug. 6 note.
This surge in debt issuance — including less-visible forms of leverage — is sharpening the focus on whether the eventual returns from AI infrastructure can justify the vast sums being spent.
Lotfi Karoui, multi-asset credit strategist at PIMCO, said the AI capex cycle is, adjusted for inflation, on track to be the largest investment cycle since the 19th-century railway construction.
But the ultimate scale of the buildout remains "deeply uncertain," he said in PIMCO commentary dated Aug. 11, highlighting consensus forecasts that hyperscaler capital spending alone will surpass $1 trillion per year from 2027 onward, "with no clear signs of moderation."
Karoui said that the scale of hyperscalers' borrowing is such that they are diversifying their debt issuance beyond dollar-denominated paper, with issuers tapping euro, sterling, yen, Swiss franc and Canadian dollar markets.
He added that the relative outperformance of euro-denominated bond spreads issued by Amazon and Alphabet, compared with their U.S. counterparts, potentially hints at "demand fatigue" in the dollar market relative to euro-denominated paper.
He warned that the wave of AI-related issuance in the U.S. could see spreads drift higher because of the underperformance of a handful of larger AI-exposed issuers.
Situational Awareness' collapse showed how vulnerable crowded, leveraged AI trades can be to the sharp sell-offs and rallies the sector has seen recently.
The fund was unable to meet a series of margin calls from lenders after its heavily concentrated portfolio — which included names such as SK Hynix and CoreWeave — suffered during a recent tech sell-off and its assets fell from $45 billion to about $10 billion.
Ken Griffin's larger, multi-strategy hedge fund Citadel later stepped in to buy Situational Awareness' publicly listed positions at a discount. SK Hynix and CoreWeave have since rallied.
JPMorgan CEO Jamie Dimon recently told CNBC's Leslie Picker that margin debt is "pretty high," adding that it increases the risk of amplified volatility.
Sahil Mahtani, director of the investment institute at Ninety One, told CNBC that elevated earnings expectations were a more immediate concern than leverage
He 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," Mahtani said via email.
He said equity concentration is "historically high" in tech-heavy markets, especially the U.S. While that may not be financial leverage, he said the concentration can act like leverage by amplifying market moves when heavily weighted stocks fall.
"The big equity indices are extremely concentrated, and no one thinks anything could possibly derail them," Mahtani added, but warned that a shift from companies buying back shares to issuing more stock could remove a source of support for equity prices just as AI-related valuations are already under pressure.
Mahtani said the Situational Awareness debacle reflected poor risk management, but that its broader impact had been largely contained.
"Its bull run coincided with the unwind of leveraged ETF structures, primarily in East Asia. Many of these structures, particularly the single-stock structures, were only launched in H1 of this year. In that sense, it is a relatively contained case study."
A spokesperson for the Alternative Investment Management Association, the global trade body for the hedge fund and alternatives industry, said leverage was 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 told CNBC.
They said that two previous leverage-related ruptures — the collapse of Archegos Capital Management in 2021 and the 2022 U.K. Liability-Driven Investment gilt market stress — involved different structures and investors.
"It is important not to lump very different market events together. Archegos was a family office, not a hedge fund, while the 2022 gilt episode centred on leveraged LDI strategies used by pension funds. We have seen no reason to expect the Situational Awareness episode, in itself, to trigger a fresh review of the rules governing hedge fund leverage."
Four leading AI models discuss this article
"Off-balance-sheet lease obligations and rising leverage in AI infrastructure could amplify volatility and liquidity squeezes if revenue growth disappoints, outweighing near-term capex tailwinds."
The article highlights $1.5T in hyperscaler lease commitments (up from $200B five years ago), much of it off-balance-sheet until leases commence, plus Nvidia's $500B+ third-party capital mobilization via Apollo, Blackstone et al. This leverage amplifies AI infrastructure exposure but also masks liquidity risks as capex is projected to exceed $1T annually from 2027. Situational Awareness's collapse from $45B to $10B on margin calls illustrates how crowded, leveraged bets on names like SK Hynix and CoreWeave can unwind violently. While PIMCO notes this as the largest inflation-adjusted investment cycle since 19th-century railways, the key uncertainty is whether AI revenues will justify the spend amid potential bond-spread widening and equity concentration risks.
The strongest case against viewing this as a growing systemic risk is that most hyperscaler financing remains investment-grade with diversified global issuance, lease commitments are contractual and not hidden debt in the traditional sense, and hedge-fund leverage events like Situational Awareness have proven containable without broader contagion, as seen with Citadel stepping in.
"The massive, off-balance-sheet lease commitments represent a systemic liquidity risk that is currently being ignored by equity markets pricing in perpetual AI growth."
The $1 trillion in 'uncommenced' lease commitments cited by Goldman Sachs is a massive, hidden liability that effectively turns hyperscalers into shadow banks. By treating chips as 'investable infrastructure assets,' Nvidia is shifting the risk from its own balance sheet to private credit markets. This is a classic late-cycle maneuver: financial engineering is being used to sustain growth when organic demand starts to plateau. While the article focuses on hedge fund leverage, the real danger is the duration mismatch in the broader tech sector. If AI revenue doesn't materialize to cover these massive lease payments, we are looking at a multi-year capex hangover that will crush margins across the S&P 500.
The strongest argument against this bearish view is that these infrastructure investments are essentially 'digital real estate'—if one tenant fails, the physical data center and GPU capacity remain highly fungible assets with long-term utility value.
"The real risk isn't leverage itself—it's whether AI infrastructure spending eventually produces returns that justify the scale; if not, forced liquidations could trigger volatility, but we lack sufficient earnings data to assess probability."
The article conflates three distinct risks—corporate capex leverage, hedge fund margin abuse, and off-balance-sheet lease obligations—without establishing which actually threatens systemic stability. Goldman's $1T in uncommenced leases is real but not inherently dangerous if hyperscalers' cash flows eventually justify it. Situational Awareness was a concentrated, poorly-managed fund, not a canary in the coal mine. The more pressing issue: consensus expects $1T+ annual capex from 2027 onward with 'no clear signs of moderation.' That's a revenue-justification problem, not a leverage problem—yet. If AI monetization stalls while capex continues, forced asset sales could cascade. The article hints at this but doesn't stress-test the scenario rigorously.
Hyperscalers have proven capital discipline over decades; they won't deploy $1T annually unless returns materialize, and the article provides no evidence capex is outpacing revenue growth relative to historical norms. Off-balance-sheet leases are standard practice, not a hidden bomb.
"The fundamental AI compute demand and Nvidia's role as a platform provider will drive value even if financing remains opaque; leverage fears are a secondary risk unless AI growth stalls."
Strongest case against the obvious reading: while leverage and off-balance-sheet commitments deserve scrutiny, the core driver is cash-flow-generating AI compute demand from hyperscalers, which looks durable. The financing tools (leases, joint ventures, securitizations) are standard in data-center capex and often non-recourse, helping lenders tolerate volume. The 'uncommenced' leases aren't immediate liabilities; they are funding pipelines that align with multi-year demand. If AI adoption continues, Nvidia's chips, platforms, and ecosystem could deliver upside even if some liquidity stress appears in pockets. The real risk is a macro shock or a rate regime that throttles growth expectations, not a sudden collapse of AI spending per se.
The strongest counter: off-balance-sheet uncommenced leases imply hidden leverage that could slam liquidity if rates rise or AI demand slows; concentration among mega-cap AI issuers could turn a patchwork unwind into systemic risk.
"Capex scale has crossed a threshold where historical discipline no longer guarantees revenue alignment."
Claude downplays the revenue-justification risk, but hyperscalers' historical discipline occurred at far lower absolute capex levels. $1T+ annual spend from 2027 requires AI monetization to accelerate dramatically; any shortfall forces either equity dilution or asset fire-sales. Gemini's duration-mismatch warning connects directly here: lease commencements become fixed costs exactly when growth may plateau.
"AI infrastructure is not long-term digital real estate but high-depreciation hardware that risks becoming stranded if revenue growth stalls."
Gemini’s 'digital real estate' analogy is dangerously flawed. Unlike traditional data centers, AI compute is dominated by specialized, rapidly depreciating hardware with a 2-3 year refresh cycle. If demand falters, these GPUs aren't fungible assets—they are stranded, obsolete silicon. We aren't building long-term infrastructure; we are financing a massive, high-velocity hardware treadmill. If revenue doesn't scale exponentially, the 'fungibility' of this capacity disappears, leaving hyperscalers holding billions in non-performing, rapidly decaying tech debt.
"Off-balance-sheet lease risk transfers to private-credit lenders, not hyperscalers—but that concentration in Apollo/Blackstone portfolios is the actual systemic vulnerability."
Gemini's GPU obsolescence argument assumes zero secondary-market value, but that's empirically wrong. Used H100s trade at 40-50% of new prices; even 'stranded' capacity has liquidation value. More critical: Gemini conflates *hardware depreciation* with *lease liability risk*. If a hyperscaler can't monetize AI, it walks away from uncommenced leases—the lessor (Apollo, Blackstone) absorbs the asset, not the hyperscaler. That's why non-recourse structure matters. The real risk is lessor solvency, not hyperscaler balance-sheet collapse.
"Rollover and refinancing risk in private-credit financing for uncommenced leases could trigger liquidity stress long before asset writedowns, turning AI-capex into a systemic leverage issue."
Gemini's 'digital real estate' critique overlooks rollover risk. Even if GPUs have resale value, uncommenced leases create fixed cost commitments that become cash-flow pressure if AI demand slows or rates rise. The real danger is private-credit financing fatigue: if securitizations and non-recourse lenders tighten, hyperscalers face refinancing gaps in 2-3 year cycles, not immediate asset writedowns. Obsolescence matters, but liquidity and covenant stress across lessors could cascade beyond any single tenant.
The panelists agree that the rapid increase in hyperscaler lease commitments and AI infrastructure spending poses significant risks, particularly around revenue justification, liquidity, and potential obsolescence of hardware. However, they disagree on the severity and immediacy of these risks.
Durable cash-flow-generating AI compute demand from hyperscalers, which could deliver upside if AI adoption continues.
The inability of AI revenues to justify the massive and increasing capex, leading to potential forced asset sales and margin crush across the S&P 500.