The debt risks tied to the AI industry are triggering a wave of warnings from veteran Wall Street strategists. Jefferies global head of equity strategy and author of GREED & fear, Chris Wood, compared the current AI boom to the 2008 subprime crisis in his latest quarterly report and flagged a structural hidden danger: hyperscale cloud providers are extending enormous amounts of unsecured credit to two continuously loss-making AI labs.
Wood argued that the most likely long-term outcome of the AI story at the market level will be large-scale capital destruction in the United States, with market share shifting to cheaper Chinese open-source large language models.
The AI story will face a major reset at some point in the not-too-distant future.
Chinese Models: From Cheap to Equivalent
Wood's warning is not new, but the scale is escalating.
OpenRouter data he tracks shows that in the week ended September 27, top Chinese models processed 76.12 trillion tokens, compared with just 4.37 trillion at the end of April this year. Over the same period, top US models processed 14.2 trillion tokens. China's figure doubled within ten weeks.
Quality has also changed. Wood noted that Moonshot's Kimi K3 is, in AI industry terms, competing at the frontier, with performance estimated at 95% of Anthropic Claude Fable 5. A model delivering 95% of the performance at an extremely low price is no longer a discount brand, but a genuine competitor.
The consequences for pricing are already visible. According to Silicon Data's LLM Token Spending Index, the average market price of AI tokens has fallen 52% since its May high, dropping to a record low of $0.98 per million tokens.
Credit Concentration Risk Behind $2.3 Trillion in US Backlog Orders
Wood likened the credit risk of the AI boom to the 2008 housing crisis, with the following logic:
By the end of the second quarter of 2026, the combined remaining performance obligations, or RPO, of the four major platforms Microsoft, Google, Amazon and Oracle reached $2.34 trillion, up 186% from a year earlier, an increase of about $1.52 trillion.
The problem is that about half of this enormous order book is owed by two companies, OpenAI and Anthropic, and neither is currently profitable.
Wood wrote that hyperscale cloud providers have effectively provided concentrated unsecured loans to tenants that are still burning cash, because neither Anthropic nor OpenAI is profitable. For neo-clouds, the risk is even greater because they themselves borrow money to rent out chips.
Jefferies technology analysts estimate that these two frontier AI labs accounted for about 45% of Google Cloud Platform revenue and 25% of Microsoft Azure revenue in the second quarter of 2026, while their share of AWS is expected to double or triple from 6% in the third quarter to 12%-18%.
Wood's analogy is direct and sharp: hyperscale cloud providers' investment in data centers is based on commitments assuming future payment by tenants, which is exactly the same as the 2008 real estate credit crisis.
Off-Balance-Sheet Liabilities: Another Hidden Layer of Risk
Wood also pointed to a layer of risk outside the balance sheet.
Moody's calculated in February this year that the five major hyperscale cloud providers had $662 billion in data center lease commitments not yet started, compared with just $152 billion at the end of 2023. A Nikkei study estimated the five companies' hidden debt at $1.65 trillion by the end of the second quarter of 2026, exceeding their on-book debt of about $1.35 trillion.
Meanwhile, the five major hyperscale cloud providers have issued $229 billion in bonds this year, compared with just $108 billion for all of 2025. Wood noted that credit spreads for Amazon, Google and Meta have risen by 13 to 16 basis points since early June, and that hyperscale cloud providers' demand for long-term financing is now competing with the federal government.
His conclusion is that credit spreads for AI-related borrowers must now be closely monitored to see whether investors are starting to become unwilling to finance AI capital expenditure, and that data center asset securitization is an area that needs close attention.
The Truth About Profits: Circular Financial Engineering?
Wood's questioning of the quality of hyperscale cloud provider earnings may be the most striking part of the report.
Data shows that in the four quarters ended in the second quarter of 2026, the four major hyperscale cloud providers' annualized profit increased by $222 billion to $581 billion. However, over the same period, the line item for other non-operating income/loss, which mainly captures unrealized gains on equity investments in unlisted companies, rose by $224 billion.
In other words, the entire increase in hyperscale cloud provider profits can be explained by mark-to-market gains on equity stakes in AI startups.
Wood described the circular logic: hyperscale cloud providers invest in AI labs, the AI labs buy computing power from the same hyperscale cloud providers, and the hyperscale cloud providers then book the labs' higher private valuations into profit.
He also noted that reported S&P 500 EPS has grown at an annualized rate of 18% over the past three years, while non-financial corporate profits on a national accounts basis grew only 5%. The last time such a large divergence appeared was at the peak of the internet bubble. Wood wrote that the above situation suggests the presence of financial engineering.
Capital Expenditure: Who Pays?
Although free cash flow has plunged from $60 billion in the fourth quarter of 2025 to $7 billion in the second quarter of 2026 and is expected to turn negative at $13 billion in the third quarter, capital expenditure has not slowed.
Market consensus forecasts that the four major hyperscale cloud providers' capital expenditure will rise from $723 billion this year to $993 billion in 2027.
Wood's question is direct: are these four hyperscale cloud providers really going to spend $993 billion in capital expenditure in 2027? That is the current consensus estimate.
His base case is that the AI story will face a major reset at some point in the not-too-distant future.
Safety or Cartel?
Wood's interpretation of recent calls for a safety pause in the AI industry is equally noteworthy.
Anthropic's Dario Amodei, OpenAI's Sam Altman and Elon Musk have recently called for slowing frontier AI development, but Trump immediately rejected the idea.
Wood does not accept the safety narrative. He wrote that it is worth focusing on incentives. A coordinated pause in the name of safety can be seen as a way to build a regulatory moat, especially if the follow-up plan is to ban open-source models, particularly those from China. Another possibility, and the two are not mutually exclusive, is that the AI giants have already realized that computing costs are becoming unsustainable and/or that demand for models is starting to plateau. If so, why not try to lock in a cartel now?
It is worth noting that Meta's Mark Zuckerberg rejected calls to slow down, and Meta shares have risen 18% since then.
The Credit Market Is the Key
Wood's investment advice has been unchanged since 2023: hold the picks and shovels, meaning the chipmakers that charge hyperscale cloud providers for capital expenditure, rather than the hyperscale cloud providers themselves. This strategy works as long as buyers keep buying.
But he believes the credit market, not the stock market, will determine when all of this ends.
His conclusion is that the AI ecosystem is now issuing more duration bonds than the US Treasury, in an environment of bond market turmoil and Fed rate hikes, with the funds used to build data centers while the largest tenants continue to lose money.
Wood invoked the historical analogy of the 1997 Asian financial crisis, and SocGen's Albert Edwards also offered a similar framework at the same time, pointing out that Thailand's miracle did not end because productivity disappointed, but because creditors noticed the problem.
Novogratz said the bubble is not spectacular enough to burst. Wood said it does not need to be spectacular: it only needs to exhaust the people willing to lend.