The technology is real. The financial scaffolding around it may not be.
That distinction is the organizing insight of a carefully reasoned paper by Paul Kedrosky and colleagues, published in February 20261 under the banner of Man Group. The authors are unequivocal that artificial intelligence ranks among the most consequential technologies of the last century. Their concern is not with the technology itself, but with the financial architecture built on top of it, and whether that structure can hold.
Their answer, in short, is no.
A loop that feeds itself
The report's most arresting observation is structural. A small cluster of mega-caps, Microsoft, NVIDIA, Amazon, Meta, Google, OpenAI, and Anthropic, function simultaneously as suppliers, customers, investors, and validators in what Kedrosky describes as "a closed, recursive financing loop." Each node in the system pays another. Revenue looks spectacular because internal demand appears endless. Capex looks justified because the demand signal is circular.
Kedrosky notes that individual participants are not behaving irrationally in isolation. Given the existential stakes of falling behind in AI, heavy investment appears defensible for any single firm. "Yet collectively," he writes, "the behaviour is deeply irrational." Multiple companies are training similar models, building overlapping infrastructure, and bidding up the same constrained resources, high-bandwidth memory, power capacity, and data centre space, with no independent market validation anchoring any of it.
The leverage hiding in plain sight
The financing has quietly migrated off balance sheets. Using asset-backed structures common in commercial real estate, hyperscalers are pushing liabilities onto special purpose vehicles, private credit funds, insurance balance sheets, and infrastructure REITs. Meta's Hyperion project is the illustrative example: a US$30 billion commitment with only 20% sitting on Meta's own books.
The mismatch at the core of these structures is what Kedrosky calls the central danger. Private equity and venture investors underwriting these projects expect power-law returns from a handful of winners. Creditors, banks, insurers, pension funds, believe they are financing long-lived infrastructure comparable to commercial real estate. The problem is that GPU chips have an effective economic life of roughly one year. A data centre filled with H100s in 2024 faces severe disadvantage against Blackwell chips in 2025. Depreciation schedules are too long. Collateral values in default are illusory. Kedrosky is direct: "This duration and usage-risk mismatch is being masked by the financing templates."
A capex-to-revenue gap that compounds
The numbers are stark. Kedrosky places AI infrastructure capex above US$200 billion against roughly US$12 billion in identifiable revenue, and the gap is not narrowing. It is widening. Historically, bubbles contract as adoption catches up to investment. In AI, the opposite is happening: GPU orders are accelerating as AI revenue growth stalls.
Token costs are falling more than 70% per year. To generate flat revenue on a collapsing price per unit would require demand to grow more than 225% annually, a bar that current inference workloads, dominated by role-playing, conversational entertainment, and price-sensitive coding assistance, are nowhere near clearing.
Where the reckoning lands
Kedrosky identifies 2027-2028 as the likely window for the first wave of stress, as initial lease terms come up for renewal and the gap between underwriting assumptions and reality becomes undeniable. By then, first-generation data centres will be two technology generations behind. Token prices will have fallen more than 90% from initial underwriting assumptions. Collateral that creditors believed was worth 70-80 cents on the dollar may be worth 20-30 cents, or less, with no viable secondary tenant market. "This scenario is not a liquidity crisis," Kedrosky writes, "it is a solvency crisis."
The second-order effects extend well beyond tech. Utilities, data centre operators, insurers, retail investors through interval funds and REITs, and pension funds have all quietly absorbed exposure to a risk they do not fully see.
5 Key Takeaways for Advisors and Investors
- Distinguish the technology from the trade. AI as a technology is durable and transformative. AI as a financial architecture, built on recursive demand loops and off-balance-sheet leverage, is not. Conflating the two leads to poor risk assessment.
- Watch private credit and REIT exposure closely. Retail investors and pension beneficiaries hold indirect exposure to AI infrastructure SPVs through interval funds, data centre REITs, and yield products. Advisors should examine underlying holdings for concentration in single-tenant data centre assets.
- The capex-to-revenue divergence is the primary signal. Hyperscaler GPU orders are accelerating while AI revenue growth stalls. This gap, $200 billion in capex versus $12 billion in revenue, is not a temporary lag. It compounds with each hardware cycle.
- Position toward cost reducers and efficient distributors, not builders. Winners are likely to be companies reducing AI token costs (chip designers, software optimization firms) and those distributing AI efficiently (enterprise software platforms, edge inference). Data centre operators and power infrastructure built to 2024-2025 demand assumptions face stranded-asset risk.
- Microsoft's pullback from OpenAI is a leading indicator, not a footnote. When the original AI flag-bearer becomes "more methodical and discerning" on AI investment, that is a signal about where the recursive loop is beginning to fracture. Advisors should treat it as an early warning, not an isolated corporate decision.
Footnote:
Kedrosky, Paul, et al. "The AI Bubble: Hidden Risks and Opportunities." Man Group Insights, 19 Feb. 2026, www.man.com/insights/the-ai-bubble.
