In a wide-ranging interview, Chanos, founder of Kynikos Associates and one of Wall Street’s most experienced skeptics, laid out a thesis that cuts against the prevailing enthusiasm. His argument is not that AI lacks transformative potential. In fact, he is careful to separate belief in the technology from belief in the businesses being built around it. As he put it bluntly, “We think the magic and the money is going to come from what the chips produce ultimately, not where they reside.” That distinction, simple on its face, has profound implications for how trillions of dollars are currently being allocated.
Chanos’ skepticism is most concentrated not on Nvidia or Microsoft, which dominate headlines, but on what he describes as the second- and third-derivative plays: the data center owners, the so-called “neoclouds,” and companies recasting themselves as AI infrastructure providers. These firms, he argues, are stepping into what is fundamentally a commodity business—one that is capital intensive, low margin, and historically prone to disappointing returns. Hosting GPUs, in his view, is not meaningfully different from being a landlord. “These are landlords,” he said. “They’re not tech companies.”
This perspective is informed by history. Chanos has been studying and shorting data center businesses for years, long before AI became the dominant narrative. In mid-2022, his firm placed significant short positions on legacy data center operators, and his view has not softened since. He describes the traditional data center model as “a very low return on capital business, highly capital intensive,” adding that many of these companies “capitalize a lot of things that we believe should be expensed” and remain “negative free cash flow.” Even before AI entered the picture, the economics were fragile.
The arrival of AI has not improved those fundamentals; it has amplified the risks. As demand for compute has surged, capital spending has exploded. Hyperscalers such as Microsoft, Meta, Google, Amazon, and Oracle are building massive new facilities, while smaller players rush to carve out niches. Yet Chanos sees a dangerous asymmetry emerging between who is spending and who is actually profitable. “The unprofitable companies as a percent of the spend is higher in this cycle than it was in the telecom cycle,” he warned. That comparison is deliberate and ominous.
To understand why Chanos keeps returning to the late 1990s and early 2000s, it is important to revisit what actually happened during the dot-com and telecom boom. The popular narrative blames excesses on cartoonish examples like Pets.com, but Chanos insists this misses the point. “The companies everybody sort of laughs about in the IPO boom of the dotcom era were tiny,” he explained. The real spending power came from highly profitable corporations—Fortune 500 companies, telecom giants, and enterprise customers—that built out networks based on overly optimistic assumptions about demand growth. When those assumptions collapsed, orders were cut abruptly, and profits across the sector fell sharply. “With a very mild recession,” Chanos recalled, “we saw corporate profits drop 40%.”
What makes the current AI boom potentially more dangerous, in his view, is that the underlying customers driving demand are often unprofitable from the outset. While Microsoft, Amazon, and Meta are the ones buying GPUs, the reason they are buying them is to serve companies like OpenAI, Anthropic, and other AI startups that burn enormous amounts of cash. “A very large amount of that underlying demand is inherently quite unprofitable,” Chanos said. As long as venture capital flows freely, this system can sustain itself. But if credit tightens or sentiment shifts, spending could fall “pretty quickly.”
This is where Oracle becomes a focal point in Chanos’ analysis. Among the hyperscalers, Oracle stands out as particularly vulnerable. While it has a profitable legacy software business, its aggressive push into AI infrastructure has dramatically expanded its balance sheet without generating commensurate returns. “They’re expanding their balance sheet really, really quickly but don’t have the income and cash flow levels that someone like Meta or Microsoft do,” Chanos said. According to his firm’s calculations, Oracle is one of only two hyperscalers that currently do not earn their incremental cost of capital, even after recent earnings releases.
Chanos’ method for evaluating this is straightforward but revealing. He looks at the year-over-year increase in adjusted operating income, annualizes it, and compares it to the year-over-year increase in the capital base. For Oracle, that ratio recently came in around 8.5%. “That’s below their weighted average cost of capital,” he noted. “It’s positive, but basically it’s telling you so far they’re destroying value.” By contrast, Microsoft’s comparable figure is close to 40%, highlighting a stark divergence in capital efficiency.
The implications for Oracle extend beyond equity performance into credit risk. Chanos does not claim an imminent crisis, but he is clear about the trajectory. “If AI monetization gets pushed out and if it’s not a 2027 or 2028 occurrence but 2030 or never, then Oracle will have fundamental financial problems,” he said. Continued borrowing to fund data center expansion could force the company to issue equity simply to stabilize its balance sheet. “So far it’s been mostly done with debt,” he added, but that option is not unlimited.
Beyond the hyperscalers, Chanos is even more critical of neoclouds and former bitcoin miners pivoting into AI infrastructure. These firms often own or lease GPUs and rent compute capacity to others, positioning themselves as essential intermediaries in the AI ecosystem. Chanos sees this as a deeply flawed bet. “Hosting GPUs or owning them and renting them out to third parties is just not a particularly good business,” he said. “It’s a commodity business.” With more players entering the market, supply is rising rapidly, undermining pricing power and returns.
A central risk in this model lies in assumptions about GPU lifespan. Companies must decide how quickly to depreciate extremely expensive hardware in an environment where performance improvements arrive annually. While hyperscalers have extended depreciation schedules from five to six years, Chanos remains skeptical. His firm uses a five-year life with a 20% residual value, but he acknowledges uncertainty. “It could be six years, it could be three or four years. Time will tell.” If GPUs last longer than expected, returns improve modestly. If they become obsolete faster, many of these businesses are in serious trouble.
Market data already offers hints. Rental prices for older-generation GPUs have fallen sharply, with indices showing declines of roughly 28% year over year. This aligns with Chanos’ view that technological progress erodes the value of existing hardware faster than optimistic models assume. While some argue older chips can be repurposed for inference or less demanding tasks, Chanos counters that such uses should still be reflected in lower rental rates. “We will have a check on this in by the free market,” he said.
Financing adds another layer of fragility. Many neoclouds rely heavily on convertible bonds, which are often misunderstood by retail investors. Chanos is particularly dismissive of claims that low coupon rates imply cheap capital. “A convert is composed of really two components,” he explained, describing the embedded call option and the bond itself. When the option value is properly accounted for, the true cost of capital can be quite high. “There’s no free lunch on this,” he said, noting that converts often function as “a backdoor way of issuing equity.”
This reflexive loop—rising valuations enabling more capital raises, which fund more capacity, which boosts reported revenues—can persist for years. But history suggests it eventually breaks. Chanos points to academic work from the early 2000s that debunked wildly optimistic assumptions about internet traffic growth during the telecom boom. When reality intruded, capital spending collapsed. “If demand is only growing at two times a year but capex is four times a year, that’s too much,” he said. AI may be transformative, but transformation does not exempt industries from basic economic constraints.
Importantly, Chanos does not position himself as an AI doomsayer. He acknowledges that companies like OpenAI could eventually become highly profitable or even monopolistic, much as Microsoft did in earlier computing eras. “If AI is truly the technology that its proponents believe it is, then it’s most likely that those companies will become inherently profitable,” he said. His critique is narrower and more surgical: the infrastructure being built around AI, particularly by weaker balance sheets and commodity operators, is where value destruction is most likely to occur.
This distinction often gets lost in market discourse, where skepticism about specific business models is conflated with disbelief in the technology itself. Chanos repeatedly emphasizes that one can be bullish on AI’s long-term impact while bearish on the economics of its current buildout. “I don’t have to have an opinion yet on that,” he said of AI’s ultimate monetization. “I can look at the hyperscalers on one side and the data center companies on the other.”
The coming years, particularly the period between 2027 and 2028, loom large in his analysis. By then, balance sheets will be enormous, depreciation will be unavoidable, and revenue growth will need to justify the capital already committed. If monetization disappoints or is delayed, stress will appear quickly. Unlike consumer spending, capital expenditure can be turned off abruptly. “We don’t need eight data centers, we just need three,” Chanos said, capturing the brutal simplicity of capex cycles.
In that sense, his warning is less a prediction than a reminder. Financial history is littered with moments when confidence outran cash flows, when belief in a transformative narrative justified extraordinary spending, and when the reckoning came not because the technology failed, but because the economics did. “Predicting the future is hard,” Chanos said, with characteristic understatement. What is clear, however, is that today’s AI boom carries risks that many investors seem unwilling to confront.
At its core, this is why Chanos calls it a confidence game. As long as capital remains abundant and belief remains strong, the system functions. But confidence is not a substitute for return on invested capital, nor does it immunize companies against balance sheet realities. Whether AI ultimately fulfills its most ambitious promises is a question for technologists. Whether today’s AI infrastructure investments will reward shareholders is a question for accountants, cash flow statements, and, eventually, markets.
For now, the spending continues, the narratives remain bullish, and skepticism is a minority position. Yet if history is any guide, it is precisely at moments of maximum confidence that such skepticism becomes most valuable.
Comments
Post a Comment