In the current wave of AI infrastructure investment, capital is flooding into a sector whose economic logic remains unproven at a historically unprecedented scale and speed, with the market choosing to price in "promises" rather than "reality."
Recently, Wall Street's prominent short-seller and founder of Chanos & Co., Jim Chanos, issued a stark warning on the Risk Reversal podcast: the scale of current AI infrastructure investment far surpasses that of the dot-com bubble era. Hundreds of billions of dollars in capital expenditure are being committed based on short-term spot pricing, yet are being used to make asset investment decisions spanning two decades. Massive amounts of equipment sit idle in warehouses, not yet deployed, with depreciation expenses deferred through accounting treatments. Upstream chip suppliers are swamped with orders, while downstream buyers are scrambling at any cost to lock in computing power—but fundamentally, Chanos argues, the numbers simply do not add up. He stated this is one of the greatest divergences between bullish and bearish views he has witnessed in his career.
The Core Economic Mismatch
This assessment directly impacts capital market pricing logic. Chanos believes the incremental return on invested capital (ROIC) for hyperscale cloud providers has fallen from around 40% roughly 18 months ago to about 20% currently. If the pace of spending continues, this figure could drop further to 10%. At that point, management at these tech giants will face a genuine capital allocation dilemma—and if they hit the brakes, the entire ecosystem of Neocloud (emerging cloud computing) companies that depend on their orders will face a chain reaction of shocks.
A More Perilous Structure Than the Dot-Com Bubble
Chanos directly compares this AI infrastructure boom to the internet infrastructure bubble of the late 1990s, but his conclusion is that the current situation is worse.
The data he cites is stark: between 1998 and 2002, during the dot-com bubble, the two most typical "cash-burning" industries—Competitive Local Exchange Carriers (CLECs) and fiber optic cable laying—accumulated about $100 billion in spending over five years, roughly $20 billion per year. In the current AI cycle, the annual data center and AI infrastructure spending of a single company already far exceeds the combined five-year total of those two industries.
More critical is the nature of the funding. Chanos points out that during the dot-com bubble, the actual paying enterprise customers—whether Bank of America, General Electric, or Coca-Cola—remained profitable throughout the cycle; they simply "cut their orders." In this cycle, however, the hyperscale cloud providers are the only entities that are truly profitable and bearing the expenditures. Almost every other participant in the ecosystem relies on venture capital or high-leverage financing to stay afloat.
He also draws an analogy to the asset-liability mismatch preceding the 2008 financial crisis: back then, institutions funded long-term derivative books with short-term repo market money, a maturity mismatch that ultimately triggered a systemic crisis. "It's basically 'Finance 101' level of mistake," Chanos said. "People are committing to long-term capital projects based on short-term spot pricing."
Accounting Tactics Masking Real Depreciation Pressure
Chanos specifically highlights a widely overlooked accounting issue: massive amounts of purchased but unused GPUs and data center equipment are currently recorded under the "Construction in Progress" line on balance sheets, where depreciation has not yet begun.
"That means these assets are undergoing both economic and technological obsolescence, and it's not reflected in the income statement at all," he said. Factoring in an approximate 18-month lag from GPU procurement to actual deployment, even with a nominal depreciation schedule of five to six years, the actual amortization period is stretched to six and a half to seven and a half years.
Chanos noted that Chanos & Co. uses a 10-year depreciation schedule in its own models for conservatism. "Even then, we still can't make the economics of most data centers work." He also pointed out that this accounting treatment creates a macro-level distortion: the S&P 500's expected earnings growth this year is a lofty 28%, with about 20% next year, far exceeding the long-term historical trend of about 6% annually. "Part of the reason is that massive capital spending by one party is being recognized as revenue on another party's income statement, while the spender capitalizes that cost and doesn't expense it currently."
The "Light Asset" Pivot of Neocloud Players
This AI infrastructure boom has spawned a batch of emerging cloud computing companies (Neocloud) operating with heavy asset models, but recent market moves are undermining this narrative.
Chanos specifically named Nebius in the interview. The company recently announced a pivot to a "light asset" business model—it will no longer own data centers and GPUs, but instead transform into a "computing power management service provider" akin to a hotel franchisor, offloading capital expenditures to third parties and collecting management fees. Chanos's assessment of this move was "a pretty significant self-own."
"For the last two years, these Neocloud companies have been telling us that heavy assets are the core competency, the 'money printing machine.' And now one of the biggest players comes out and says, 'You know what, we don't need to own these assets,'" he remarked.
He offered a deeper explanation: whether traditional data center firms or new Neocloud players, they are starting to realize that due to labor and equipment shortages, new construction costs are skyrocketing, and future maintenance capital expenditures will far exceed the numbers previously described to investors. "So they're trying to get these assets off their books as fast as they can."
He also noted that multiple deals for compute rental from SpaceX's XAI data center—including agreements with Anthropic and Google, each around $10 billion—come with extremely short exit clauses, allowing withdrawal in as little as three months. "I think those deals are very promotional in nature."
Interest Rates as the Real "Time Bomb"
Within Chanos's analytical framework, interest rate risk is the most underestimated systemic threat to the entire AI infrastructure bubble.
He points out that a vast amount of assets—whether office buildings, data centers, or warehouses—are trading at capitalization rates (cap rates) of 5% to 7%, while the U.S. 10-year Treasury yield is around 4.5%. With such a razor-thin spread, project sponsors are stacking leverage and using aggressive mezzanine financing to promise equity investors 15% returns.
"Once rates go to 6% or 7%, this whole thing blows up. It blows up asset class by asset class."
He believes the credit market currently shows no signs of concern, but that could change abruptly if rates move noticeably toward 5%—widening credit spreads would be a key warning signal. He also observes that spreads for the lowest-rated CCC tier of junk bonds have begun to widen, but the BBB to BB range has not yet followed.
Chanos also poured cold water on the so-called "power bottleneck" narrative. He stated that data center power costs constitute only about 5% to 6% of revenue, the smallest of all cost items. "We are absolutely not short of electricity in this country." He expects the premium logic built around power scarcity for related assets is not solid.
The Hyperscaler's Dilemma When Returns Fall Below "Treasury Yield"
Chanos's core judgment is that the capital efficiency of hyperscale cloud providers is systematically declining and will trigger a real strategic shift within the next 12 to 18 months.
According to his calculations, as a group, Google, Meta, Amazon, Microsoft, and Oracle have seen their incremental ROIC drop from around 40% a year and a half ago to about 20% currently. If capital expenditure growth continues, this figure could fall further to 10%.
"At that point, management at these companies faces a real question: Do we keep spending like this, or is it better to just buy Treasuries?" Chanos said. He anticipates this question will become unavoidable by late 2026 or 2027.
He cited Oracle as an example: the company's AI infrastructure-related incremental capital returns are the lowest among its peers, and its stock is down 65% to 70% from its highs.
"Any CEO who is actually doing their job is looking at Oracle and saying to themselves: We can't get to that point."
Chanos concluded by summarizing the current market's pricing logic and potential inflection point in one sentence:
"In a bull market, people pay a premium for 'promise.' In a bear market, people only discount for 'reality.' We are clearly in the former right now. Do we get to the latter? I don't know."
Comments