Jim Chanos views the legacy data center business as fundamentally flawed, describing it as a low-return, highly capital-intensive industry with negative free cash flow. He applies this same critical lens to the new wave of AI data centers, arguing that the underlying business model remains weak.

He believes that hosting GPUs is essentially a commodity business, especially with so many companies rushing to build capacity. The real financial opportunity in AI is not in housing the hardware, but in what the hardware enables. As he puts it:

The magic and the money is going to come from what the chips produce ultimately, not where they reside.

For those looking to invest in AI, he suggests focusing on pure AI companies like OpenAI or the hyperscalers, rather than bitcoin miners converting to data centers or other new entrants he considers mere landlords. He highlights two major problems with the data center model. First, it is an inherently low-margin, low-return business. Second, companies that buy their own GPUs face a significant risk regarding their depreciable life. It is difficult to predict how quickly tenants will demand newer, more powerful chips, forcing costly upgrades.

The AI industry can be broken down into three main categories. First are the hyperscalers, who sell products produced in data centers. Second are the companies that supply and build the data centers, like Nvidia. The third category consists of companies that own the data centers. Jim Chanos expresses the most skepticism and bearishness towards this third group.

While many of the largest companies in the S&P 500, such as Microsoft, Nvidia, Meta, and Google, fall into the first two categories, his focus is elsewhere. Since his firm advises clients to hedge their portfolios, they are effectively long on these major index components. His concern lies with the 'second or third derivative' companies trying to build a business solely around owning and operating data centers, a model he believes is not very good.

Among the five major hyperscalers building out data centers—Microsoft, Meta, Google, Oracle, and Amazon—there is a significant difference in profitability. Oracle, in its push to catch up, is expanding its balance sheet rapidly but lacks the income and cash flow of its competitors. According to Jim's analysis, only Microsoft and Meta are currently earning returns above their cost of capital on their incremental AI spending. In contrast, Oracle and Amazon are at the bottom, struggling to monetize their massive investments in AI hardware and infrastructure. This situation could change if AI output is monetized more effectively in the future, but for now, only Microsoft and Meta can comfortably handle their investment load.

Meta, like Microsoft, is building out massive data centers for AI, funding the expansion from its own cash flows generated by advertising. This highlights a significant shift for the largest tech companies. These businesses, long known for being asset-light with high cash flow, are now becoming extremely capital-intensive. This change is causing their free cash flow to decline, with a company like Oracle already going free cash flow negative.

The massive investment is a bet that AI monetization will reach an inflection point around 2027 or 2028, leading to a surge in profits. Jim Chanos describes this as a fundamental change in their business models, akin to a second industrial revolution.

You basically had asset light businesses based on intellectual property that just gushed cash flow. And now you have basically the second industrial revolution, if you will, where the next round of technology improvement is really capital intensive and that's a change.

Assessing the return on this invested capital is difficult, as companies don't disclose specific AI revenues and costs. To approximate it, Jim suggests a method: compare the annualized year-over-year increase in adjusted operating income (the numerator) with the year-over-year increase in the capital base (the denominator). This calculation reveals a wide disparity. Microsoft's incremental return is almost 40%, suggesting they are winning so far. In contrast, Oracle's is about 8.5%, which is below its weighted average cost of capital, indicating it may be destroying value.

However, this situation could change, much like Amazon's AWS was unprofitable for years before becoming a massive success. The primary risk now is that these impressive returns often come from selling to unprofitable companies, like OpenAI and Anthropic, which are funded by venture capital.

These returns are all based on, in many cases selling to unprofitable companies. And as long as those unprofitable companies can keep raising money to keep paying those bills, that's fine. But if we get a credit crunch or we get a 2001, 2002 pullback in sentiment, you're going to see a lot of that spending drop and drop pretty quickly.

This dependency on external financing for their customers means the risk levels for these tech giants are much higher than they were five or ten years ago.

Jim Chanos expressed the most concern for Oracle among the giant companies in the S&P 500. He explained that Oracle's spending plan is enormous relative to its expected returns. In contrast, Amazon has many other businesses, and his firm is effectively long on the other major tech companies through hedges. The primary negative view is reserved for Oracle.

If the monetization of AI gets pushed out to 2030 or never materializes, Oracle could face fundamental financial problems. The company is not at that crisis level yet, but it would be under significant stress if it continues to borrow money for its build-out. Chanos speculates that Oracle might eventually need to issue equity to keep its balance sheet in check, as its expansion has so far been mostly funded with debt.

The discussion then turned to the surge in Oracle's shares during the second quarter. This was fueled by a dramatic increase in its reported remaining performance obligations (RPO), which jumped from $137 billion to $455 billion. While some see this as a sign of incredible future demand, Chanos is skeptical.

It's not a GAAP metric. It's not a real backlog. It's if everything goes right and they continue to deliver, our revenue should be this over the life of the contracts.

He noted that the market was reacting positively to any company announcing data center plans during that period. He advises investors to be wary of the RPO number and to give it a wide berth.

The real money in the AI boom will be made from what the chips produce, not where they are physically located. Companies that own GPUs and rent them to third parties are not tech companies; they are landlords in a commodity business. As more and more companies rush to build out data centers, the increasing supply will drive down returns.

There is a mistaken belief that simply having land and power creates a monopoly or oligopoly that allows for charging extremely high rents. This is not the case. For example, Iron's highly publicized deal with Microsoft resulted in only a mid-single-digit return on capital.

If that's what a monopoly-type provider gets because they have access to power and land right now from a really good credit, I mean, yawn, who cares?

This illustrates the core problem. If a well-positioned provider can only achieve such a low return from a top-tier client, the business model is fundamentally unattractive. This is the basis for a negative view on the 'Neoclouds' and Bitcoin miners entering this space; they are all jumping into what will ultimately be a low-return, commodity business.

Data center companies like Equinix and Digital Realty Trust are primarily co-location businesses that provide real estate, power, and cooling. Their stated returns are very low, even using a generous 15- to 20-year depreciable life for assets. Digital Realty sees low single-digit returns, while Equinix is in the mid-single digits pre-tax.

A major issue lies in their accounting practices. These companies, structured as REITs, have massive capital expenditures (CapEx) for existing data centers. However, they play an accounting game by classifying most of this spending as "growth capex" rather than recurring capex. This is permitted if they can argue an upgrade, like a new HVAC system, might attract a new client or allow for price increases. Jim Chanos finds this questionable.

If the HVAC system goes down in a data center, it goes down, you have to replace it. How you categorize it is sort of semantics.

Despite this supposed growth spending, the companies are only growing revenues in the mid-single digits and have 20% vacancy rates. Both Digital Realty and Equinix have CapEx greater than their EBITDA, indicating they are financially stretched. They are essentially borrowing money and issuing stock to pay their interest and dividends, which is an unsustainable position.

Large hyperscalers like Microsoft and Meta increasingly want these assets off their balance sheets. They prefer to lease capacity from companies like Digital Realty and Equinix, letting someone else assume the risk of capital intensity. This trend highlights that the major tech players are trying to avoid owning these assets directly.

Core Weave's business model is that of a 'GPU middleman'. The company doesn't own data centers but leases them. It then owns the GPUs and sells the compute power to large hyperscalers. According to Jim Chanos, this business is fundamentally a bet on the depreciable lives of these GPUs. For Core Weave's model to succeed, the GPUs need to have a very long useful life, perhaps 10 or 12 years, allowing them to be leased out for an extended period. This contrasts with the shorter depreciation schedules used by the hyperscalers themselves.

If a chip lasts for 10 years, a company only depreciates 10% of its cost each year. However, if it only lasts for three years, that depreciation cost jumps to a third of the purchase price annually. This raises a critical question about the business model. Jim frames it this way:

If these chips last so long and Microsoft is depreciating its chips over six years, if these things last longer than six years, why is Microsoft leasing the capacity from you?

Currently, hyperscalers are using about a six-year depreciation life for their GPUs. Jim's own modeling uses five years with a 20% residual value, which is in a similar range. Market data, such as the Bloomberg Hopper GPU rental index, shows a 28% year-over-year decline in spot rental prices, which is consistent with a four to six-year life. Furthermore, Nvidia's rapid innovation cycle, with new chip architectures released annually like Blackwell and the upcoming Rubin, means older chips quickly become less valuable. While older chips can be repurposed for less intensive tasks like inference, their rental rates should reflect their diminished capability, providing a free-market check on their true value over time.

Companies like Core Weave are financing capital-intensive growth by using convertible bonds, a strategy also employed by bitcoin miners. A common misconception among retail investors is that these are a source of cheap or even free capital, based on their very low coupon rates, such as 1%. This view fails to account for the true cost.

A convertible bond has two main components: a call option on the company's stock at a set price and a low-coupon bond. The real cost of capital is quite high because it includes the value of the option the company is giving away. If the company's stock price rises above the conversion price, it results in significant dilution for existing shareholders, which is functionally the same as issuing new shares.

There's no free lunch on this. The fact that so many of these companies are trying to issue converts is in effect, it's a backdoor way of issuing equity.

A recent example illustrates this high cost. A company that issued convertible bonds less than two years ago had to pay three times their original value to buy them out, resulting in substantial share dilution. What initially appeared to be cheap capital ended up being very expensive. Ultimately, issuing convertible bonds is simply a less direct way of issuing equity.

When a reflexive loop forms where more capital raised leads to higher revenues, which in turn allows for more capital to be raised, it's hard to know what stops it. A historical parallel can be found in the dot-com and telecom boom of the late 1990s. The stock market began its decline in March 2000, correctly anticipating the drop in orders and earnings that didn't materialize until 2001 and 2002.

A key factor was a widespread, but incorrect, belief that internet traffic was doubling every three months. This assumption, promoted by companies like WorldCom, led to massive orders for equipment from corporations building out their networks. However, research from Anthony Odlyzko at the University of Minnesota punctured this myth, showing that traffic was actually doubling every year, not every quarter.

What Odlyzko's work showed was that it was doubling every year. So up 2x, not up 16x, and that makes a world of difference for your ordering if you suddenly realize that you only need one-eighth of what you thought you were going to need.

When companies realized this, order books collapsed. This led to a 40% drop in corporate profits during the mild recession of 2001-2002. Today, a similar dynamic may be at play with economic growth and tech profitability heavily dependent on the data center and AI build-out. There's a significant risk that companies are overestimating their need for this equipment. Unlike consumer spending, which adjusts gradually, capital spending can be switched off abruptly. A company can decide it needs three data centers instead of eight, causing an immediate hit to earnings. This increasing earnings risk looks a lot like the situation in 1999 and 2000.

The claims about large language models like ChatGPT growing faster than the internet need to be examined critically. While the growth is rapid, it is coming off a very low base, which makes high percentage gains easier to achieve. It becomes much more difficult to sustain that kind of growth when revenues reach 50 or 100 billion dollars.

The period around 2027-2028 will be a crucial test for these companies. By then, they are projected to be massive in terms of their balance sheets. If the high growth rate does not continue at that point, they will face challenges in continuing to place new orders for infrastructure and servicing their debt.

A key parallel to the current AI boom is not the semiconductor cycle, but the telecom equipment boom and bust of the dot-com era. The massive capital expenditure seen then in telecom equipment is mirrored now in the buildout of AI data centers. However, Jim Chanos argues the current situation is significantly riskier.

The primary reason for this heightened risk is the financial state of the underlying customers. In the late 1990s, most of the spending on new technology came from profitable corporations like General Electric and AT&T building out their networks. The unprofitable entities, such as the fiber optic companies, were a relatively small part of the overall spending. Today, the situation is reversed.

What's scary about this one is that the underlying customers that are spending a lot of the money are unprofitable. And so it raises the risk level I think in a way that's worse than 1999 and 2000.

While it's true that profitable giants like Microsoft and Amazon are buying the chips from Nvidia, a large portion of their motivation is to sell compute power to unprofitable AI companies like OpenAI and Anthropic. This means the ultimate demand is built on a much less stable foundation. A popular market narrative incorrectly remembers the dot-com bust as being caused by small, unprofitable companies like Pets.com. The reality is that the crash happened when immensely profitable Fortune 500 companies canceled their orders for network equipment.

Regarding the future of AI companies themselves, it remains uncertain. If AI proves to be as transformative as its proponents believe, companies like OpenAI could become monopolistic and highly profitable, similar to Microsoft's rise during the PC era. However, for now, they are losing vast amounts of money, and there needs to be a clear path to monetization sooner rather than later.

The funding for major AI initiatives like OpenAI raises questions about sustainability. With companies raising vast sums of venture capital and investments like Disney's $1 billion deal, the capital intensity is enormous. The host wonders how this will play out, questioning if companies will need to constantly seek new funding sources.

I remember back in the day when a billion dollars was a lot of money, but that is not going to fill this bucket. Every week they fly over to Saudi Arabia to get another billion dollars. Like how is this going to play out?

Jim Chanos draws parallels to the dot-com boom, specifically the telecom build-out. He notes that venture capital is available until, suddenly, it isn't. This can happen almost instantaneously, as it did in the third quarter of 2000. The situation may be reflexive, a term from George Soros. If the stock prices of these AI companies start to fall, the capital could dry up, as investors have been conditioned to invest more at higher prices, not lower ones.

A common misconception among novice investors is that stock prices follow fundamentals. However, the market often anticipates changes before they are evident. The NASDAQ, for example, dropped 30% in early 2000 on no significant news, correctly predicting a collapse in order books that wouldn't materialize for another six months.

The market kind of figured it out before even the purchasing managers did. If you're waiting for that kind of evidence to show up, be mindful that there could be a big impairment of your capital happening before that even happens. And then you'll see the verification of why it dropped later.

Regarding Nvidia, Jim acknowledges it is currently a cash machine and the biggest player. However, he finds it perplexing that they are engaging in vendor financing, which means investing in their own customers. This was a practice seen with companies like Lucent and Nortel during the dot-com bust. If demand is as strong as claimed, there should be no need for such measures. Despite these concerns, Jim suggests there are better targets for short-sellers than Nvidia.

Jim Chanos expresses skepticism about the alternative investment industry, particularly private equity and private credit. For years, he has viewed private equity as leveraged equity, where the returns were not commensurate with the risks taken. He argues that the arbitrary marking of portfolios hid these risks. While public markets returned 10%, private equity returned 10-15% despite being levered two or three to one, which he felt was insufficient. He believes public markets have now outperformed private equity over the last five years, partly due to the high fee structure.

His current concern is the sales pitch for private credit, which he finds reminiscent of past market cycles. The promise is to earn equity-like returns, around 10 to 15%, while holding senior-level debt. In an efficient market, a debt holder should not be able to earn equity rates of return. He draws a direct parallel to the late 1980s and Mike Milken's junk bond market. Milken argued that junk bonds offered superior returns because they were shunned, and even with defaults, high recovery rates would protect investors.

However, that narrative rested on two false assumptions. First, the excess returns came from 'fallen angels'—companies that were once investment grade, became junk, and then recovered—not from newly issued junk bonds. Second, Milken's default rate calculations were misleading. By measuring defaults against the total amount of bonds outstanding in a rapidly growing market, the denominator grew faster than the defaults. A proper cohort analysis revealed that default and recovery rates were not as favorable as presented. The entire model collapsed during the real estate and S&L bust of the late 80s and early 90s.

Another parallel Jim highlights is the use of regulated entities. Milken's network had regulated insurance companies and S&Ls buy the junk bonds of other companies within the network. Today, a similar pattern is emerging. Large private equity and credit firms, like Apollo with Athene, increasingly own regulated entities such as insurance companies to purchase their private debt. This potentially exposes retirees, who are the beneficiaries of annuities from these companies, to risks they may not fully understand.

Even if these entities hold mostly investment-grade assets, the leverage is key. You could have 90% investment grade, but it might still be 100% of your equity is in that stuff.

Ultimately, Jim is skeptical of the idea that a vast amount of private credit is simply inefficiently priced. He suggests that the high returns are often generated through inherent leverage or other innovative ways to goose returns, which at the end of the day, usually just entails taking more risk.

Historically, the returns for limited partners (LPs) in private equity and private credit have been good on paper. However, investing in the publicly traded general partners (GPs) like Apollo or Aries has often been an even better investment. This presents an interesting observation: it was better to invest in the GP, not the LP.

The accounting for these large firms, including Aries, Apollo, and Blackstone, is extremely complicated. Even for sophisticated investors, it can be difficult to parse. Jim Chanos confirms their complexity, mentioning that while he hasn't looked at them recently, they are indeed hard to understand. He points to Brookfield as another example, describing it as a "maze of interconnectedness."

Live Nation Entertainment, the owner of Ticketmaster, is a company Jim Chanos is shorting, but not for the obvious reason that it's a monopoly. The short thesis is based on accounting. Live Nation has numerous deals with venue providers and artists that are structured as minority interests. These partnerships, representing billions in capital, don't fully appear on the balance sheet.

However, their impact is significant on the profit and loss statement. After the bottom line, deductions for these minority partners are enormous, claiming 30 to 50% of the business's cash flows depending on the quarter. While the company and sell-side analysts promote metrics like Adjusted Operating Income (AOI) to justify its valuation, this ignores the reality of the business.

The people that you don't see, your partners in that business, have a claim on a third to a half of your cash flows and you have to adjust for that. And so if you do that, you'll see that Live Nation makes almost no money. Its actual earnings are de minimis.

This accounting structure makes the parent company appear far more profitable than it truly is. The cash outflow to these partners is real and can be seen on the cash flow statement. If the profitability metric was adjusted to account for these partners' cut, the stock price would be significantly lower. The problem appears to be worsening over time. This highlights a broader red flag for investors: whenever a company has a lot of non-controlling or minority interests, it's worth digging deeper, as it can be a way to obscure the true state of the business.

Jim Chanos describes the six-month period after GameStop in 2021 as the most speculative market of his career. This era was characterized by a confluence of meme stocks, NFTs, a significant rally in crypto, and Special Purpose Acquisition Companies (SPACs). The speculation in SPACs was particularly extreme.

At one point SPACs were raising 2 to 3 billion a night in February of '21, which was equal to the entire US savings rate. So you knew that wasn't going to last.

When comparing that period to the present, he notes that recent market activity has given 2021 a run for its money. The late spring and summer saw similar speculative behavior in AI, nuclear, and quantum computing stocks. However, there is a key difference: the level of new stock issuance seen in 2021 has not been matched. The offset to this is that the market is starting from a higher valuation level. The recent market has also seen a big move in crypto and the rise of what he calls absurd "crypto treasury companies," drawing further parallels to the 2021 frenzy.

Unlike the dot-com era, a lot of company financing is now happening in the private markets. Jim Chanos explains that the private markets are significantly larger than they were 25 years ago, allowing them to absorb the capital needs of many companies and keep them on their books for longer periods.

This creates a sort of Catch-22 for some companies. Going public might lead to a lower valuation than what they are currently marked at in their private funds.

If you go public, you may find a valuation lower than what you've been marking in your private fund.

The conversation then turned to Uber, a company Jim was previously short on. While his thesis on the company's flawed fundamental economics was correct for a long time, he acknowledges that Uber has now reached 'escape velocity' and is no longer a short position for him. He notes, however, that the initial valuations during its IPO and post-IPO phase were excessively high, causing many investors to overpay.

Jim Chanos's firm focuses on idiosyncratic, one-off investment opportunities rather than just timely trends like data centers. These are often stories rooted in complex accounting. A prime example he shared is Erie Insurance, which he says was playing significant accounting games.

Erie moved its insurance company off its books by essentially giving it to the policyholders. Then, the publicly traded company charged these new insurance entities a flat fee of 25% of their premiums, portraying itself as a service company. However, the underlying insurance companies had no employees or directors; the publicly traded Erie company did all the work for them. Jim's view was that this was not an arm's-length transaction. Erie's only customer was the Erie insurance companies, and those companies relied entirely on the parent's employees.

This was simply an accounting dodge to remove the mundane and recently unprofitable property and casualty business from their books, instead just showing service income from the top line. As this perspective gained traction, the stock price came down significantly. These are the types of idiosyncratic ideas, like Live Nation or Erie, that the firm thrives on.