Jefferies warns of a "2008 moment": AI may trigger a "capital destruction" in the US, with capital flowing to Chinese open-source models

Debt risks in the AI industry are triggering intensive warnings from veteran Wall Street strategists. Chris Wood, Global Head of Equity Strategy at Jefferies and author of "GREED & fear", drew parallels between the current AI boom and the 2008 subprime mortgage crisis in his latest quarterly report, highlighting a structural vulnerability: hyperscale cloud providers are extending large amounts of “unsecured credit” to two consistently loss-making AI labs.

Wood believes: “The most likely long-term market outcome for the AI story will be large-scale capital destruction in the US, with market share shifting to cheaper Chinese open-source large language models.”

He argues that the AI story “will face a major reset at some point in the not too distant future.”

Chinese Models: From "Cheap" to "On Par"

Wood’s warning is not new, but the scale is escalating.

According to OpenRouter data tracked by Wood, the leading Chinese models processed 761.2 trillion tokens in the week ending September 27, up from only 4.37 trillion at the end of April; during the same period, leading US models processed 14.2 trillion tokens. China’s figure doubled within ten weeks.

Quality is also shifting. Wood notes that Moonshot’s Kimi K3 “is, in AI industry parlance, ‘competing on the frontier,’ with estimated performance reaching 95% of Anthropic Claude Fable 5.” A model achieving 95% performance at extremely low cost is not a discount brand—it is a true competitor.

Price consequences are already apparent. According to Silicon Data’s LLM Token Spend Index, the market average price for AI tokens has dropped 52% from its May peak, reaching a historical low of $0.98 per million tokens.

US $2.3 Trillion Backlog: Credit Concentration Risk

Wood likens the AI boom’s credit risk to the 2008 real estate crisis, laying out the logic as follows:

By the end of Q2 2026, the combined remaining performance obligations (RPOs) of Microsoft, Google, Amazon, and Oracle reached $2.34 trillion, an increase of 186% year on year, with an increment of about $1.52 trillion.

The issue is that about half of these massive orders are owed by OpenAI and Anthropic—neither of which is currently profitable.

Wood writes: “Essentially, hyperscale cloud vendors are extending centralized unsecured loans to still loss-making tenants, because neither Anthropic nor OpenAI is profitable. For neo-clouds, the risks are larger still, because they themselves borrow money to rent out chips.”

Jefferies technology analysts estimate that by Q2 2026, these two cutting-edge AI labs will account for about 45% of Google Cloud Platform’s revenue, 25% of Microsoft Azure's revenue, and their share in AWS is expected to double or even triple in Q3 from 6% to 12–18%.

Wood’s analogy is sharp: “Hyperscale cloud providers’ investments in data centers are based on assumed future payments from tenants—exactly the same logic as the 2008 real estate credit crisis.”

Off-Balance-Sheet Liabilities: Another Layer of Hidden Risk

Wood also points to risks that sit off-balance-sheet.

Moody’s tallied in February that the five biggest hyperscale cloud providers had $662 billion in not-yet-commenced data center lease commitments, up from just $152 billion at the end of 2023. A study by Nikkei estimated their “implicit debt” by Q2 2026 at $1.65 trillion, surpassing their reported balance sheet debt of about $1.35 trillion.

Meanwhile, the five largest hyperscale cloud providers have already issued $229 billion in bonds this year, compared with just $108 billion for all of 2025. Wood notes that credit spreads on Amazon, Google, and Meta have risen 13 to 16 basis points since early June; “Demand for long-term financing from hyperscalers is competing with the federal government.”

His conclusion: Credit spreads for AI-related borrowers “must now be closely monitored to see if investors start to show reluctance to finance AI capex.” Data center asset securitization “is an area that needs close attention.”

The Truth About Profits: Circular Financial Engineering?

Wood’s questioning of the profit quality of hyperscale cloud providers might be the most striking part of his report.

Data shows that the four hyperscalers increased annualized profits by $222 billion over the four quarters to Q2 2026, reaching $581 billion. However, over the same period, “other non-operating income/loss”—mainly capturing unrealized mark-to-market gains from investments in unlisted companies—increased by $224 billion.

In other words, the full increase in hyperscalers’ profits can be explained by mark-to-market gains from their equity in AI startups.

Wood describes a circular logic: hyperscalers invest in AI labs, the AI labs purchase computing power from the same set of hyperscalers, hyperscalers then mark up their profits with higher private valuations of these labs.

He also points out that reported EPS for the S&P 500 has grown at an 18% annualized rate over the past three years, while national income-accounts-based non-financial corporate profits have only increased 5%. The last divergence this large was during the peak of the internet bubble. Wood writes: “The above circumstances indicate financial engineering.”

Capital Expenditure: Who Will Pay the Bill?

Even though free cash flow has plummeted from $60 billion in Q4 2025 to $7 billion in Q2 2026—and is expected to turn negative $13 billion in Q3—capital expenditure has not slowed.

Consensus forecasts project the big four hyperscalers’ capex to rise from $723 billion this year to $993 billion in 2027.

Wood’s question is pointed: “Will these four hyperscalers really spend $993 billion in capex in 2027? That is the current consensus estimate.”

His baseline scenario is: The AI story “will face a major reset at some point in the not too distant future.”

“Safety” or Cartel?

Wood’s interpretation of the recent call for an AI “safety pause” is equally noteworthy.

Dario Amodei of Anthropic, Sam Altman of OpenAI, and Elon Musk have all recently called for a slowing of frontier AI development, though Trump immediately rejected the idea.

Wood is skeptical of the call for “safety,” writing: “Look at the incentives: a coordinated pause in the name of ‘safety’ can be seen as a way to build regulatory moats, particularly if, subsequently, open source models—and especially Chinese open source models—are banned. Another possibility—not mutually exclusive with the first—is that the AI giants have realized compute costs are becoming unsustainable and/or demand for models is plateauing. If so, why not try and lock in a cartel now?”

Notably, Meta’s Mark Zuckerberg rejected calls for a pause, after which Meta’s stock price rose 18%.

Credit Markets Are the Key

Wood’s investment advice has not changed since 2023: hold “picks and shovels” (that is, chip manufacturers collecting capex from hyperscalers), rather than the hyperscalers themselves. This strategy works as long as the buying continues.

But he believes what will ultimately bring all this to an end is the credit market—not the stock market.

His conclusion: “The duration bonds currently being issued by the AI ecosystem already surpass those by the US Treasury, at a time when the bond market is turbulent and the Fed is raising rates. The financing is going to build data centers while the largest tenants remain loss-making.”

Wood draws a historical analogy to the 1997 Asian Financial Crisis—SocGen’s Albert Edwards proposed a similar framework at the time—arguing that Thailand’s “miracle” did not end because of disappointing productivity, but because creditors noticed the problems.

Novogratz says the bubble is “not yet spectacular” enough to burst. Wood says it doesn’t have to be: it just needs to exhaust willing lenders.