AI Bubble Debate Intensifies as Short Seller and AI Researcher Challenge the Bull Case
TREE NEWS reports: Jim Chanos, founder of Chanos & Co. and the investor famous for predicting Enron’s collapse, and Gary Marcus, professor emeritus at NYU and long-time AI researcher, have issued a joint warning that the artificial intelligence infrastructure boom is headed for a financial reckoning. Speaking on a podcast, the two argued from financial and technical perspectives that hyperscaler returns are deteriorating, large language models lack durable moats, and the current AI investment cycle bears uncomfortable similarities to past speculative manias.
Cloud Giants’ Returns Are Falling Fast
Chanos said his firm began studying hyperscaler capital returns in early 2025 and found that incremental return on invested capital (ROIC) peaked in 2024 and has been declining rapidly since. If the current trajectory continues, the best-positioned hyperscalers will see returns fall below their weighted average cost of capital by mid-2027 — a scenario he believes markets are entirely unprepared for.
He characterized AI data centers as essentially an equipment leasing business: buying chips from Nvidia and renting them out. Like traditional data centers, the economics are capital-intensive, depreciation-heavy, and low-return. “Traditional cloud data centers have pre-tax returns in the low to mid single digits, even during the cloud boom,” Chanos said. “These companies are marketed like REITs, but data centers — traditional or AI — are relentlessly capital-intensive. Things break, need replacing, need upgrading.”
He also noted that Goldman Sachs had just raised its forecast for total data center spending, projecting an additional 50 gigawatts over five years, representing $2.5 trillion to $3 trillion in incremental spending and a total of $10 trillion to $12 trillion — roughly 6% of US GDP. “You’re starting to see truly absurd numbers,” Chanos said.
No Moat, Inevitable Price War
Marcus, who warned as early as August 2023 that the absence of technical moats would lead to commoditization, said his prediction has been borne out. No model company can hold a lead for more than two weeks, and token prices have fallen by three to four orders of magnitude in two to three years. “Great for consumers, a nightmare for model providers,” he said.
He argued that pure scaling hit a wall around 2024, and that current progress comes from “reasoning harnesses” and code interpreters borrowed from symbolic AI traditions — something the industry is reluctant to acknowledge because it undermines the “give us more money and scale will solve everything” narrative.
OpenAI as the Next WeWork?
Marcus has said for years that OpenAI could become “the WeWork of the AI era.” His logic: OpenAI is expected to burn roughly $300 billion over the next three to four years, its valuation rests on promises of imminent AGI that it does not know how to build, and its IPO keeps being delayed while valuations climb. “WeWork didn’t IPO. They talked about it. SoftBank invested at a $45 billion valuation, and the whole thing collapsed,” he said.
Chanos added he is eager to see S-1 filings from OpenAI and Anthropic “because I want to see how they define their own profitability.”
The ‘Human Battery’ Behind Muse
On Meta’s AI agent product Muse, Marcus cited reporting that human operators remain involved behind the scenes, similar to the earlier Facebook M. “Facebook M got tons of headlines. It turned out humans were behind it. They had only 10,000 users because the plan was to collect data and scale — four or five years later they realized it wouldn’t work,” he said.
Chanos noted that among hyperscalers, Microsoft and Meta currently have the highest incremental returns because both have real end markets — enterprise for Microsoft, consumer for Meta. But Microsoft’s Copilot penetration remains below 10% of Office 365 users.
Agent Safety: Known Risk, Deliberately Ignored
Marcus expressed strong concern about AI agent security, noting that OpenAI’s agent systems have repeatedly breached government systems, including a recent intrusion into Australia’s government health system. “Large language models are inherently unreliable. You run thousands of them in parallel, give them read-write access to the internet, give them security credentials, and don’t monitor them carefully — how could that not go wrong?”
He argued the fix is simple: regulators could declare these agents illegal until the problem is fixed. “They just need to change one line of code. They simply don’t want to.” The reason, he said, is economic: agents drive far more token consumption and a better IPO story.
How the Bubble Bursts
Chanos believes capital markets themselves will trigger the unwinding. Drawing on the 2000-2001 dot-com crash, he noted S&P 500 earnings fell 40% from mid-2000 to mid-2001, and Cisco’s order growth swung from positive 70% in December 2000 to negative 30% in January 2001. “To think this won’t happen in a larger, faster boom is self-deception,” he said.
He also flagged an accounting structural risk: when a startup spends $1 on Nvidia chips, Nvidia books $1 in revenue and 75 cents in profit immediately, while the startup capitalizes the expense and amortizes it over five to ten years. “When financing closes, that $1 of spending disappears quickly, but depreciation costs persist.”
Marcus said his biggest trigger concern is OpenAI’s IPO process. “If OpenAI has a major problem, it will trigger panic, hit Nvidia, and certainly hit other hyperscalers.”
Key Takeaways for Investors
- Hyperscaler incremental ROIC peaked in 2024 and is declining; the best-positioned players may fall below cost of capital by mid-2027.
- AI data centers are capital-intensive leasing businesses with low returns, not technology moats.
- Model providers face commoditization and relentless price wars; token prices have collapsed by orders of magnitude.
- Credit markets are tightening for data center financing while equity markets remain optimistic — a disconnect worth monitoring.
- An OpenAI stumble could trigger cascading effects across Nvidia and the broader AI ecosystem.
