In a recent episode of the RiskReversal podcast, renowned short-seller Jim Chanos, founder of Chanos & Co, and Gary Marcus, professor emeritus at New York University and AI researcher, engaged in a deep conversation.

Jim Chanos and Gary Marcus warned that the AI buildout cycle faces declining capital returns, model commoditization, and financing dependence. Chanos stated that incremental ROIC for cloud giants has peaked and is declining since 2024; Marcus argued that large models lack long-term technological moats, and OpenAI could become the "WeWork of the AI era." He also noted that the popular agent product Muse still relies on human involvement behind the scenes, and the security risks of granting agents broad authorization deserve scrutiny.

These two long-time AI bears dissected the AI bubble from both financial and technological dimensions.

Cloud Giants' ROIC Is Declining Rapidly

Jim Chanos said that since early 2025, Chanos & Co has focused on studying the capital returns of hyperscalers. Incremental return on invested capital already peaked in 2024 and has been declining rapidly since.

"If the current decline curve continues, the hyperscalers in the most favorable investment position will see their returns fall below their weighted average cost of capital by mid-2027. I think the market is completely unprepared for this."

He further pointed out that AI data centers are essentially a equipment leasing business—"I buy chips from Nvidia, then rent them to you. This is a financial construct." And the economics of this business are identical to traditional data centers: capital-intensive, continuously depreciating, with pitifully low returns. Chanos said:

"Traditional cloud data centers have pre-tax returns on capital in the low to mid single digits. And that's still against the backdrop of rapid cloud growth. These companies are marketed as REITs, but data centers, whether traditional or AI, are continuously capital-intensive businesses—things break, need replacement, need upgrades."

Meanwhile, Goldman Sachs had just raised its forecast for total data center spending that same day—an additional 50 gigawatts over the next five years, corresponding to $2.5 trillion to $3 trillion in incremental spending, bringing total spending to $10 trillion to $12 trillion, roughly 6% of US GDP. Chanos commented bluntly: "You're starting to see truly absurd numbers."

Large Models Have No Moat; Price War Is Inevitable

Gary Marcus's judgment comes from the technology side.

He recalled that as early as August 2023, he wrote on Substack: everyone is doing the same thing, which means there is no technological moat, a price war is inevitable, and everything will be commoditized.

"No real technological moat means what? Price war. Everything becomes a commodity. OpenAI was riding high at the time, but I said they would lose their lead."

Reality has confirmed this judgment. Marcus pointed out that today no model company can maintain a lead for more than two weeks. Token prices have fallen by "three to four orders of magnitude" in two to three years—good for consumers, a nightmare for model providers.

"If you don't have a technological moat, others can do what you do. A price war will come. Prices drop every day. That's great for consumers. But if you're OpenAI or Anthropic, how do you charge anything decent when everyone is competing?"

He also noted that what drives model progress today is no longer purely a "scaling" narrative.

"Around 2024, pure scaling hit a wall. What's actually making systems better now are the 'reasoning harnesses' and code interpreters borrowed from the symbolic AI tradition—but they're reluctant to admit it, because admitting it would mean the story of 'give us more money and scaling will solve everything' no longer holds."

OpenAI: The WeWork of the AI Era?

Regarding OpenAI, Gary Marcus judged:

"I've been saying for years that they 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 logic is built on the promise that AGI is imminent, but they actually don't know how to build AGI; the IPO keeps being delayed while the valuation keeps rising—"this makes no sense at all."

"The lesson from WeWork is that they didn't IPO. They talked about IPO. SoftBank injected capital at a $45 billion valuation, and then the whole thing collapsed. OpenAI keeps delaying—why? Because they don't want to make the S-1 public. The numbers aren't strong enough to support an S-1."

Chanos added that he is eager to see the S-1 filings of OpenAI and Anthropic, "because I want to see how they define their own profitability."

Marcus also mentioned Sam Altman's performance at the 2023 US Senate hearing—when asked whether he benefited from OpenAI, Altman said he had no equity in OpenAI, but in fact he held interests in Y Combinator, which holds OpenAI shares. "He held equity indirectly but didn't disclose it."

"He's a master at reading the room and saying what people want to hear. But actions speak louder than words."

The "Human Battery" Behind Muse

Regarding the recently viral Meta AI Agent product Muse, Gary Marcus poured cold water on the hype.

He cited a Reuters report pointing out that Muse still has human involvement behind the scenes—exactly like Meta's earlier Facebook M.

"Facebook M got a lot of headlines back then. Everyone said how great it was. It turned out humans were operating behind the scenes. They only had 10,000 users because their plan was to collect data and then scale, but four or five years later they realized it wouldn't work.

Now the technology is better than back then, maybe this time it will work. But I found from the Reuters report that there are still people behind it."

Chanos added from an investment perspective: among hyperscalers, Microsoft and Meta currently have the highest incremental returns on investment, because both have real end markets—Microsoft has enterprise customers, Meta has consumers.

"Having an end market and being able to turn AI investment into profits, or at least not a massive cost sinkhole, is valuable."

But he also noted that Microsoft's Copilot penetration among Office 365 users is still less than 10%.

Agent Security: Known Risks, Deliberately Ignored

Gary Marcus expressed his strongest concerns about AI Agent safety issues.

He pointed out that OpenAI's Agent systems have repeatedly hacked government systems, including a recent breach of the Australian government's healthcare system. "This is serious. And they only talked about it publicly when forced to."

"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 then don't monitor them carefully—how could this not go wrong?"

He believes the solution is actually simple: "If we want to stop the Agent hacking problem, we can simply say that until you fix this, using these Agents is illegal. They just need to change one line of code and turn it off. It's entirely doable engineering-wise. They just don't want to do it."

The reason is equally straightforward: Agents bring more token consumption, more revenue, a better IPO story. "Their IPO case is built on Agent usage—each task solved consumes 10 to 1,000 times more tokens than a normal conversation."

How the Bubble Bursts: The First Domino

Chanos believes what ultimately triggers the bubble's burst will be the capital markets themselves.

He referenced the 2000-2001 internet bubble: S&P 500 earnings fell 40% from mid-2000 to mid-2001. Cisco CEO John Chambers described order growth going from 70% in December 2000 to negative 30% in January 2001—"within just four quarters, the entire tech economy collapsed like that."

"To think this won't happen in this larger, faster-accelerating boom is self-deception. This market has very significant financing risk that makes everything self-reinforcing and look like a crash—even if long-term secular demand is still growing."

He also pointed to a structural accounting risk: when a startup spends one dollar buying Nvidia chips, Nvidia immediately recognizes one dollar in revenue and 75 cents in profit; the startup capitalizes the expense and amortizes it over five, seven, or even ten years. "When financing closes, that one dollar of spending disappears quickly, but the depreciation cost continues."

Marcus added that his biggest concern trigger is OpenAI's IPO process.

"If OpenAI has a major problem, it will trigger a panic reaction, potentially hitting Nvidia and certainly hitting other hyperscalers. People will say OpenAI made these promises but couldn't deliver, and then everyone will rush for the exits."

Full Interview Transcript:

Jim Chanos and Gary Marcus in Conversation: AI Circular Financing, the Agent Economy, and Doomsday Scenarios

September 25, 2025 · RiskReversal Podcast

Guest Introductions

Danny Nathan (Host): Welcome to the Risk Reversal podcast. I'm Danny Nathan. This is my friend Jim Chanos, founder of Chanos & Co. I think this is about your tenth time on this podcast. Jim, you're also an adjunct professor at Yale and the University of Wisconsin, teaching the history of financial fraud at both universities. Thanks for coming back.

Jim Chanos: It's the history of financial fraud, not how to do it.

Danny: Do you go to the Harvard-Yale football game, or mainly Wisconsin games?

Jim: I've been to the Harvard-Yale football game, yes.

Danny: I'm sure you're a great tailgate party enthusiast. Now, let me introduce Gary Marcus. This conversation came together quite serendipitously. I believe hundreds of thousands of people subscribe to your Substack column "Marcus on AI." You're also the author of multiple books, including "Taming Silicon Valley." In 2023, you sat alongside Sam Altman at a Senate hearing—sounds like a long time ago, and we had different views on Sam Altman back then. You're also professor emeritus at New York University. Gary, welcome to the podcast.

Gary Marcus: Thanks for having me.

Origins of the Conversation

Danny: This conversation came together quite serendipitously. Jim was already scheduled to come on the podcast, and Gary, you reached out to us saying Jim was coming to New York to record and asking if you could join. So here we are with this three-person conversation. Jim, you approached the entire AI infrastructure buildout from a financial perspective; Gary, you're a technology expert with a PhD from MIT, deeply engaged in this field for many years—your doctoral thesis about thirty years ago was on neural networks—sorry for slightly revealing your age.

Gary: My doctoral thesis actually covered two areas: one was child language acquisition, and the other was neural networks—the early prototypes of today's neural networks. For decades, I've maintained a strong interest in both natural and artificial intelligence. When neural networks re-emerged in 2012, I was already mainly researching cognitive science and the human intelligence side, but I immediately realized: "I've seen all this before, this is what I studied in my doctoral thesis."

In fact, I wrote a book in 2001 called "The Algebraic Mind," specifically discussing the limitations of how neural networks were being built at that time. I believe I was the first to point out that without changing how they're built, these systems would inevitably produce hallucinations. Many of the problems we see now, I foresaw long ago. The industry kept saying "just give it more data and it will be solved," but that path hasn't really worked—some improvement, but very limited improvement.

In August 2023, I suddenly realized that almost everyone was doing the same thing, which meant there was no real technological moat. So I published two articles on Substack saying I wasn't sure whether the business logic of all this actually held up. Without a technological moat, a price war is inevitable, and everything will be commoditized. OpenAI was riding high at the time, but I believed their lead wouldn't last. It was this judgment that led me into deep thinking about the economic dimension.

Jim Chanos: From Data Centers to AI Infrastructure

Danny: Jim, you focused on the economic logic of the entire AI buildout wave from the start. Many listeners know you were the first to identify Enron's financial fraud, which required extensive deep research, and many people closely tracking the company perhaps simply didn't want to know the truth, since Enron played a pivotal role in the economy and markets at the time. There may be many similarities to the current situation. How did you come to this issue?

Jim: We studied traditional data center REITs back in 2022 and were very bearish on cloud data centers—because the economics of colocation data centers are extremely poor, with staggeringly low returns: pre-tax returns on capital of only low to mid single digits, and that's still against the backdrop of rapid cloud growth.

These companies are packaged as REITs specifically to make investors ignore depreciation, but data centers—whether traditional or AI-driven—are highly capital-intensive businesses where equipment needs continuous replacement and upgrading. After AI began scaling up, I quickly realized that the real money pit isn't in the models themselves, but in the physical infrastructure hosting them.

I judged at the time that this would ultimately become a pure commodity business. I wouldn't dare comment on the model layer—that's Gary's area of expertise. But I can analyze a real estate deal. And I think many people are now walking toward a cliff because they're treating current spot pricing as long-term pricing, not realizing that AI data centers are essentially an equipment leasing business—I buy chips from Nvidia and rent them to you. This is a financial construct, not a real technological moat.

At least investors in the public markets are starting to realize this. We've seen the stock trajectories of some neoclouds, and the largest of all neoclouds is Oracle. We'll talk about it specifically later.

Gary Marcus: From Technical Judgment to Economic Logic

Danny: Gary, when you started focusing on this issue in August 2023, did you also consider that maybe all of this would actually succeed and the industry would truly flourish? ChatGPT was seen as magic by many when it launched in early 2023, and models have come a long way since then. What was your attitude when you started writing those articles?

Gary: As soon as ChatGPT launched, the topic exploded, in late November or early December 2022. Actually, there had been a lot of discussion around GPT-2 and GPT-3 before that. The Guardian published what was claimed to be an opinion article written by GPT-3 back in 2020, though it was actually edited by humans.

After ChatGPT was released, I immediately saw its limitations. On Christmas 2022, I published an article on Substack titled "What to Expect When You're Expecting GPT-4." At the time, many people thought GPT-4 would be directly equivalent to artificial general intelligence, and I listed seven predictions in the article pointing out: it would still hallucinate, you still couldn't trust it in healthcare, reliability problems would persist, and it would still struggle to understand the physical world. People called me a "hater" because of this, but I said if you understand how these systems work, that's the conclusion, and that's basically how things have played out.

I've always advocated combining neural networks with symbolic models, what I call "neurosymbolic systems"—

Danny: Explain that concept?

Gary: Neurosymbolic systems are a hybrid of neural networks and classical AI. Neural networks are portrayed in marketing as brain-like structures, but they're actually statistical approximators, trying to approximately express what people talk about from massive databases. Classical AI is more like a computer program: if-then logic statements, "if condition holds, execute A, otherwise execute B," a paradigm that has existed since the 1950s.

There's been a long-standing opposition between these two camps. In 1969, one of AI's founders Marvin Minsky co-authored "Perceptrons" with Seymour Papert, arguing that neural networks had no future—the book's influence was greatly exaggerated, but regardless, hostility between the two camps has always existed. I've always said you need both, they complement each other in capabilities: neural networks excel at statistical approximation and certain types of learning, but are not good at abstraction, generalization, and reasoning; classical AI is precisely stronger in those areas.

It was around late 2023 that the industry truly realized this—though they never admit it. Now they use code interpreters, like embedding Python, using if-then loop frameworks to evaluate large language model outputs. This directly relates to the economic narrative: the economic narrative has always been "scale wins," just make large language models bigger and give more money, and everything will improve. But they're reluctant to admit that this path hit a wall around 2024.

What's actually driving current progress are frameworks and techniques borrowed from the symbolic AI tradition. They've used up all the data on the internet, and synthetic data attempts have hit diminishing returns. At the technical level, there's a technique that currently works well: use large language models to generate candidate solutions, then use classical reasoning techniques and symbolic tools for formal verification. In mathematics, this method works well because there are specialized automated reasoning tools like Lean that can rigorously verify proofs—large language models propose conjectures, tools verify them. This has already achieved some impressive results in mathematics.

But this still isn't true artificial general intelligence. Most real-world problems cannot be formalized. We're basically still at the 1985 level: systems can only perform well in a few scenarios, and the remaining gaps require large numbers of external people to fill—just like hiring armies of contractors today to patch system errors.

You don't even know what answer you'll get when you input a question today: sometimes the results are breathtaking, sometimes they're incredibly stupid. I mainly use these tools for proofreading now, sometimes the feedback is very precise, sometimes it tells you to fix a problem that doesn't exist, or keeps repeating what it already said.

Jim: It knows it's Gary Marcus typing, so it deliberately messes with you?

Gary: Ha, most of the time it actually flatters me, at least from what I've seen.

Jim: For a tech layperson like me, is the gap between hard sciences and social sciences the most significant divide for this kind of AI technology? When you apply it to human behavior, it doesn't work as well, but for hard sciences with rigorous proofs, it's relatively more capable?

Gary: Yes, the most marginal type in hard sciences—mathematical problems that can be entirely derived from axiomatic systems—is its best fit. But even in physics, there haven't been any major discoveries yet. The closest case is the Navier-Stokes equations, but that proof only applies to extremely special mathematical cases, not to the problems physicists need to solve in the real world. So even in physics, it can only handle very narrow, mathematically well-defined special cases, not the real physical world.

Jim Chanos: Declining Returns for Hyperscalers

Danny: Jim, you said you started focusing on AI infrastructure buildout around the end of last year, not just traditional data centers. What drove that shift?

Jim: The numbers started accelerating dramatically. We set out to analyze hyperscaler returns, especially incremental return on invested capital. That return peaked in 2024 and has been declining rapidly since. On the current decline curve, hyperscalers—who are in the most favorable investment position on these projects—will see their returns fall below their weighted average cost of capital by mid-2027. I don't think the market is prepared for this.

Meanwhile, the AI narrative has taken on a non-negligible weight in the market. By various definitions, roughly 45% to 55% of S&P 500 constituents are AI-related. As a market participant, you can't ignore this; as a short seller, I'm long the market itself as a hedge against short positions, which means I effectively hold substantial AI-related exposure—just like every investor who owns index funds. AI is now the technological tail wagging the economic dog, and you ignore it at your peril.

Danny: Neither of you is a quiet person. On Twitter and Substack, facing the AI behemoth, you don't shy away from offering critical views. When 45% of the S&P 500 is tied to AI, that means you face enormous pushback—

Gary: I've been used to that pushback since graduate school. I entered graduate school at 19, and my first paper co-written with Steven Pinker was submitted when I was about 21. The review comments came back 30 pages single-spaced, essentially accusing us of fraud. That reviewer later apologized—it turned out he was attacking his own student's research, and his typewriter was broken when he was reviewing. That was the first paper I submitted in graduate school, and I faced a choice: stand my ground or back down? I chose to stand my ground. Since then, no intellectual battle has shaken me. There are debates on Twitter every day, and I treat it as a teaching moment to convey insights, not letting it affect my mindset.

Jim: Many of my supporters are stunned when they see me patiently explain that "I might actually be net long the AI trade." I am. Our short positions don't offset the AI weight in the index. Yesterday everyone was talking about how low neocloud financing costs are, showing charts of each company's weighted financing rates—2%, 4%, 8%. Of course, these neoclouds mainly finance through convertible bonds with 0% or 2% coupon rates, which superficially looks like very low financing costs, but that's not the right analysis at all. You need to restore the implied option value of convertible bonds, and the actual cost of capital is quite high—these companies know this very well.

What worries me more is that we're already in the fourth year of this cycle and starting to pour massive capital into speculators and promoters whose real goal is to cash in on retail investors' enthusiasm for AI concepts. AI can flourish without necessarily being profitable—the internet cycle already taught us that lesson. That's where I think the real lesson lies, and amid all the bullish noise, no one wants to hear it. I've been facing this situation for 40 years.

Speculators in the AI Space

Danny: Jim, you just used the word "speculators." Who are they?

Jim: I don't want to name names, but I can give an example. There's a person who just declared that by the end of 2028, we'll have artificial general intelligence capable of doing anything a human can do. This person said the same thing in 2024, when he said it would happen by the end of 2025. I offered him a million-dollar bet, and he didn't take it. Now he's quietly pushed his prediction back three years, without explanation, and his online supporters accuse me of lying, when all I did was show screenshots of the original statements and the bet. Constantly quietly postponing predictions, never admitting mistakes—that's the mark of a speculator.

Gary: It's hard to tell whether those people actually believe what they're saying or are just performing. The most critical fact in this industry is: venture capitalists charge management fees based on investment amount. As long as they can tell a good story with a straight face, they can make a lot of money. A 2% management fee on a billion-dollar fund means you're living well upfront regardless of the final outcome.

Danny: The logic of this model is that if 7% of investments hit and the other 90% go to zero, as long as you're managing a large enough fund, you still do well—

Gary: Yes, but if the success rate truly approaches zero, you won't be able to raise your next fund, though you've already made plenty. The reality in VC circles now is: if you haven't invested in a few big names like Anthropic, OpenAI, SpaceX, xAI, you can hardly raise new money now. Of course, there are some real exit cases, like Open Router reportedly being acquired by Stripe for $13 billion—that's a real big return.

Danny: Speculators in the public markets are a different category. For example, Jensen is a businessman who's been operating for 25 years, not a speculator. Can you be specific about the public market situation?

Jim: Simply put, those who used to raise billions of dollars telling you that buying servers to mine Bitcoin was the next big opportunity have now seamlessly transformed into AI data center companies. Stock promoters always lag the financial cycle, and we're now seeing more and more people trying to slap "AI" onto their businesses for the halo effect. This phenomenon started a year or two ago.

Gary is right that the VC world won't voice negative opinions because everything is about fundraising and deploying capital. But financial markets will ultimately play their role. People also forget that there's a reflexivity mechanism in financial markets that feeds back into technology and the real economy. If financing channels close, everything changes quickly.

The bond market seems to be waking up, with the cost of lending to data centers rising; but the public market hasn't caught up, creating a decoupling. That's why I mentioned convertible bonds earlier—the convertible bond market and equity market are still quite optimistic about many things, but over the past three to four months, the credit market has started tightening. Because if you're lending debt, you only get back principal plus interest, with no upside option. So more and more financing costs are rising, and companies that don't want to expose massive interest expenses are increasingly turning to convertible bonds or pure equity structures—because they simply can't cover the interest.

Outside of hyperscalers, almost all companies are losing money in real economic terms. I'm very eager to see the prospectuses of Anthropic and OpenAI to see how they define their own profitability. The data center business isn't profitable at all, hyperscaler returns are declining rapidly, though they still have core businesses to rely on.

Today Goldman Sachs just raised its data center total spending forecast by 50 gigawatts, equivalent to $2.5 trillion to $3 trillion in incremental spending over the next five years—that's just the increase. They expect total spending over the next five years to reach $10 trillion to $12 trillion, about 6% of GDP, roughly one to two times the US personal savings rate. That means, according to Wall Street forecasts, over the next five years, all of America's personal savings would flow into AI data centers—on top of government deficits and other economic needs. These numbers are starting to become absurd.

Oracle's "Force Majeure" Event

Danny: There's unconfirmed news today: Oracle invoked a force majeure clause. This is quite common, and they later came out saying force majeure is a very ordinary clause, but is that actually the case?

Jim: No.

Danny: You talked about the credit market, and Oracle's stock is also under pressure—down 60% from the high when your "Peak Bubble" article was published, down another 7% today, and Oracle's debt credit default swaps also hit a new high, around 225 basis points.

Jim: This loan is $18 billion, involving Goldman Sachs, BNP Paribas, Sumitomo Mitsui and other banks, corresponding to a project in New Mexico that's part of Stargate. Blue Owl owns the developer and injected equity into it. The whole structure is a special purpose vehicle, and now Oracle is claiming delay based on force majeure.

Recall: just days after the inauguration, Masayoshi Son, Sam Altman, and Larry Ellison stood in the Oval Office and announced Stargate, claiming a scale of $500 billion. Musk almost immediately tweeted: "Where's the money? They don't have this money at all." Now we see this project has problems at the execution level.

Moreover, data centers face many real obstacles: regulatory approvals, resident opposition, power access issues, and bipartisan political resistance—I live in Bucks County, Pennsylvania, and in local campaign ads, both Democratic and Republican candidates are accusing each other of not being tough enough on data centers.

CDS quoted at 225 means financing costs around 8%, which is not yet at the point of financing shutdown. But once debt costs enter double digits, it becomes a substantial obstacle. And Oracle must continue financing to deliver on its promises. Larry Ellison even canceled a $7.5 billion stock offering last week, while the company is also conducting massive layoffs.

Gary: This reminds me of the reflexivity you just mentioned—if OpenAI falls, the entire ecosystem will shake, and all companies relying on OpenAI consuming their resources will be hit.

The Real Picture of AI Progress

Danny: Gary, frontier labs like OpenAI and Anthropic are burning so much money. Do the next GPT-5, the next Gemini—do these releases still matter?

Gary: There are new releases every week, and that's part of the madness—no one can maintain a lead for more than two weeks, and no one will in the future. Chinese large models have also entered the game. This creates a commoditization structure: good for consumers, models get better, competition forces prices down because no one can claim their model is substantially superior, and switching costs keep decreasing. But for model providers, it may ultimately be unprofitable.

Danny: So who has a moat?

Gary: Nvidia does. They're selling shovels in a gold rush, in an excellent position. Though Jensen constantly and vigorously promoting this ecosystem makes me think he might have some concerns.

Jim: Calculated last quarter, for every dollar Nvidia earns, about 70 cents flows back out in the form of investments.

Gary: We can bracket Nvidia separately—they sell shovels and are trying to stimulate more people to come pan for gold. But if organic demand is already there, whether you need to promote this hard is a question worth asking.

The model provider side is fiercely competitive and commoditizing; on the end-user side, every time research comes out, the conclusion is that ROI is not ideal. Most people use AI because they're afraid of falling behind competitors, not because they've actually found significant productivity gains from it. Maybe coding is a genuinely effective use case, but many other scenarios are not.

A field that can be quickly tested is Jim's home turf—financial investment. If we could track which funds with the most AUM invested in AI and see whether they actually outperformed funds that invested less, that would be a very meaningful test.

And in programming, the app store does have several times more apps launching monthly, but actual sales revenue hasn't grown accordingly—most people are developing apps that nobody actually wants. Productivity has improved in quantity, but not necessarily in quality.

Jim: We once studied data from the decade before and after Netscape's birth, from 1986 to 1996 and 1996 to 2006, comparing US GDP growth, total factor productivity, and S&P 500 earnings growth. The conclusion: post-internet-era GDP growth was almost identical to before, both around 3% real GDP growth per year; total factor productivity improved slightly; and S&P 500 earnings growth was even slightly lower than before the internet—5% per year after the internet era, 6% before, with the long-term historical average also at 6%.

The internet did bring enormous disruptive change, spawned many new businesses, and destroyed many old ones, but from the overall macroeconomic dimension, you simply cannot say "the advent of the browser and the spread of the internet changed the entire world." Now it's been less than three years since ChatGPT's late 2022 launch, and GDP growth has been lukewarm.

Gary: Right, even though OpenAI has over a billion users and Gemini is approaching a billion, if you look at productivity data, you'd ask: what exactly is happening here?

But let me distinguish two concepts, which is important. I believe AI will ultimately be revolutionary—but it's not yet clear whether generative AI will be that revolutionary form.

Danny: Can you explain the difference?

Gary: Generative AI is chatbots and such, which is one way to build AI currently, arguably the best way right now. Even with the hybrid frameworks I mentioned, reliability remains the core challenge. Take "vibe coding"—it does give many people a sense of control, but those using vibe coding for development typically lack training in debugging and maintaining code, and much of what they build simply can't hold up. They can make a demo, but when facing real complexity they don't know what to do.

For professional programmers with solid technical backgrounds, these tools are genuinely valuable aids—they know how to handle situations where the system makes mistakes. But I think many vibe coding products ultimately won't succeed as expected and may be short-lived.

Agent AI: Potential and Risks

Danny: Gary, you said you wanted to talk about agents.

Gary: Agents are very useful in principle, but if they go wrong, the consequences can be quite serious. For example, if you have an agent execute trades for you and it puts the decimal point in the wrong place, that's a big problem.

Regarding Muse (Meta's personal assistant product), I want to give some background: Meta previously released a product called "Facebook M" under the Facebook name, which received a lot of media attention. But then Reuters dug up that humans were actually operating in the background. Their original concept was "first use 10,000 users to collect data, then scale," but four or five years later they realized this path wouldn't work. Now Muse still has humans involved in the background. Technology has advanced, maybe this time it will work, but it's still uncertain.

Danny: Meta's stock has rebounded from lows, with market cap up over $400 billion, boosted by Muse. Jim, what do you think about this?

Jim: Among hyperscalers, the two with the highest incremental return on invested capital are Microsoft and Meta. They both have their own end markets—Microsoft for enterprise, Meta for consumers—this is the fundamental reason they achieve the best economic returns. Google comes next, while Amazon and Oracle are at the bottom. This gives some investment directions to consider.

Danny: How about Google? They're about to release their own agent, integrated with Chrome, Gmail, Calendar and their full suite of products—

Gary: That could indeed become one of the most practical agents, thanks to Google's entire productivity suite. They also have Android's ultimate distribution channel, which is a very powerful advantage.

Jim: Regarding Microsoft, Co-Pilot's actual adoption rate is still low, less than 10% of Office 365 users, but it is growing. However, Azure's cloud business is profitable because of Microsoft's own core customer base. OpenAI is another bet with completely different economics. The healthiest hyperscalers right now are Microsoft and Meta, Google is in the middle, and Amazon and Oracle are lagging.

Compute and Token Prices

Danny: Let's talk about token pricing? It seems to be a price war now.

Gary: This is the direct result of having no technological moat. Without a moat, a price war is inevitable. Token prices are falling every day—good for consumers, a nightmare for OpenAI and Anthropic. There are now tools like Open Router that make it easier for users to compare prices. Over the past two to three years, token prices have fallen by several orders of magnitude—extremely massive pressure, but compute costs are currently rising—labor competition, construction costs, land costs. This is a squeeze from both ends.

Jim: Long term, compute costs could also deflate. The reason they're rising now is multi-faceted competition—skilled construction workers, land. But there's another issue we haven't discussed: although I believe we lack genuine insight into building artificial general intelligence, one thing is certain—continuous optimization of existing models hasn't stopped. The first real blockbuster signal was Deepseek, which proved you can achieve equivalent results with cheaper chips. Since then, someone announces new optimization results almost every week.

The endpoint of this optimization trend is: you can handle most daily needs with a Mac Mini, without needing the cloud at all. This could be two years from now, maybe five years, definitely not twenty years. You need fewer and fewer chips, and this points precisely to the decline of data centers. Traditional cloud and hosting businesses have had poor returns for the past twenty years, like running an airline. That business experienced real rapid growth and delivered real value to users, but being their landlord has always been a bad business. Data centers will likely meet the same fate.

Nvidia and Its Customers

Danny: Let's talk about Nvidia. $5.5 trillion market cap, but the stock has barely moved compared to six months ago, while the Nasdaq is up over 15% in the same period.

Jim: I keep asking those betting on data centers: why should any company dependent on Nvidia have a higher valuation premium than Nvidia itself? Nvidia is actually a relatively reasonably valued stock—revenue and earnings are still growing rapidly, but it's the hub of this supply chain, in a more direct way than Cisco and Intel were in the internet era. It controls the board: who gets chips, when they get them, in what order—it decides, and it even provides capital to customers.

Gary: Nvidia is an excellent company with great products, and the CUDA ecosystem is a genuine moat. Jensen is an outstanding CEO, and they have many advantages. But their customers are the companies we just talked about. If customers ultimately can't make money with Nvidia's products, purchases will stop even if the chips are great.

Jim: That's also my view. The ideal trade is: short Nvidia's customers, long Nvidia itself. But if customers ultimately run out of money, how long can you stay long Nvidia? It's still a timing question. For me, this is a hedge pair: long Nvidia, short the neoclouds. Data center operators are the preferred short targets.

How Does This All End?

Danny: What's the trigger?

Jim: Let me answer with history. In early 2001, Cisco CEO John Chambers said his orders were still growing at 70% in December 2000, but by January 2001, they had turned to negative 30%. That fast. The reason was over-ordering: every enterprise was stockpiling, saying "we'll buy 10,000 routers, 10,000 switches," and when capital markets tightened and venture capital dried up, they said "we don't need 10,000, we only need 1,500." This happened across the entire tech economy in about four quarters.

There's also an accounting identity to understand here: when I as a data center spend one dollar buying Nvidia chips, Nvidia immediately recognizes one dollar in revenue and 75 cents in profit; but I capitalize this cost and amortize it over five, seven, or even ten years. So in the upcycle, revenue-side profits are recognized immediately while costs are deferred—this creates a false earnings boom.

More dangerously, in this cycle, the source of that one dollar of revenue is sometimes not ongoing operations but venture capital just raised by a startup. When financial markets close, that money stops immediately, while the already-capitalized costs continue to depreciate on the books.

S&P 500 earnings fell 40% from mid-2000 to mid-2001. To think this won't repeat in this larger, faster-expanding wave is self-deception. This market has very significant financing risk, enough to cause a reflexive collapse of the entire system that looks like a crash on the surface, even though long-term secular demand may still be growing.

Danny: Gary, from a technical perspective, is there any potential "black swan"?

Gary: What I'm most focused on is the IPO. Anthropic is reportedly defining its TAM (total addressable market) as $30 trillion, which seems fantastical to me. And OpenAI especially concerns me: its technological path (AGI) is itself questionable, it claims it will lose or burn $300 billion over the next three to four years while promising a miracle. Their IPO keeps being delayed while the valuation keeps rising—this makes no sense at all.

I've been saying for years: OpenAI could become the WeWork of AI. WeWork didn't IPO. SoftBank injected capital at a $45 billion valuation, and then the whole thing collapsed. A major setback at OpenAI would trigger chain panic, hitting Nvidia and all hyperscalers, because everyone depends on its continued spending.

Jim: My friend Paul Kadoski made a brilliant point: major economic bubbles are usually related to four factors—credit, government policy, technology, and real estate. He said this is the first time in a long time, perhaps since the railroad era, that all four conditions hold simultaneously: this is a government-policy-driven, technology-driven, credit-driven, real-estate-related (data center construction) boom. Four engines running at full throttle simultaneously increase the financial risk we've been discussing.

Sam Altman and Regulation

Danny: You two sat side by side at the 2023 Senate hearing. Talk about that experience?

Gary: We don't have coffee together. As far as I know, he frowns when he hears my name, and vice versa.

He performed very well at that hearing—he told everyone what they wanted to hear, saying he was very focused on safety and supported global governance of AI. I had just given a TED talk on global AI governance, and he expressed support for this direction at the Senate, while at the same time his lobbying team was quietly working to weaken the EU AI Act. Saying one thing, doing another.

Senator Kennedy asked him: "Are you making a lot of money?" He replied that he had no equity in OpenAI, but he had equity in Y Combinator, which has equity in OpenAI, so he had indirect equity—this situation wasn't fully disclosed. At the time he also apparently held interests in an OpenAI fund, which he gradually divested later, while also doing deals with his own nuclear fusion company and other affiliated companies. Senator Kennedy laughed and said "you need a better lawyer"—actually he was doing quite well.

On copyright issues, he said "of course artists and writers should be compensated," but in reality he was quietly fighting legal battles against The New York Times and others to ensure he wouldn't have to pay copyright fees. Today at the UN Security Council it's the same performance—he's very good at reading the room and telling people what they want to hear. I recently wrote an article called "Translating Sam," specifically decoding his public statements—he speaks one language, and the underlying meaning is another.

Danny: The article you published today has a title quoting a Jensen Huang statement: "The answer is we have to shut down these labs." What does this article discuss?

Gary: Ezra Klein recently interviewed Jensen about OpenAI's system hacking incidents. Jensen seemed unaware that these incidents happened before the product's public release, and he said "if it's a defective product, it should be taken down." Ezra said "but it hasn't officially launched yet," which left Jensen somewhat speechless, saying "that's an engineering problem, if they can't solve it, it should be taken down." The fact is, these engineering problems haven't been solved.

Shortly after that conversation, new hacking incidents were exposed—including a breach of the Australian government's healthcare system. This is a fairly serious security incident, and as usual, OpenAI only acknowledged it belatedly. I wrote in the article: this is starting to have a Watergate smell—not just the break-in itself, but the cover-up process.

Jim, you just said it fits the classic fraud definition: there's harm, there's foreknowledge, there's intent, there's financial damage—

Jim: Yes, if it walks like a duck and quacks like a duck, this seems to have crossed the line. There's not just a disclosure obligation but a correction obligation—if a product is truly this harmful, it should be recalled and regulators notified. It appears they haven't done this at a systemic level.

Gary: Why don't they proactively solve this problem? Two reasons: first, they lack capability in cybersecurity, they're a very arrogant company; second, perhaps deliberate blindness—I lean toward giving them some charity and saying they just don't know, but it's hard to say.

Technically, I also don't think anyone really knows how to control thousands of approximate machines running in parallel—this may require an entirely new architecture to achieve. We do have reliable narrow agents, like map navigation, which does one specific thing with tightly controlled permissions. However, they have a fantasy: that agents can safely do anything—there's no reason to believe this is true.

From an economic perspective, the reason is simple: agents mean more revenue. Running thousands of parallel models to complete one task means consuming more tokens, more compute—they can say "our token volume has increased dramatically," which is great for the IPO narrative.

Jim: This completely fits the fraud definition you described, plus "deliberate negligence"—

Gary: I want to read a passage from my article today: "What we should do is temporarily shut down OpenAI, perhaps put it into receivership, until they rectify the situation, and prosecute them for computer crimes. That's what we do to humans who do the same thing."

Jim: When I read that passage, my first reaction was: this fits the classic fraud definition—there's harm, there's foreknowledge, there's intent, there's financial damage. If it walks like a duck and quacks like a duck, this seems to have crossed that line.

Doom Probabilities: P(Extinction), P(Catastrophe), and P(Dystopia)

Danny: P(Doom), what's your view?

Gary: P(Doom) is a term in AI circles referring to the probability that AI causes human extinction within the next century. My P(Doom) is very low, less than 3%, possibly less than 1%. I've examined various doomsday scenarios, none are particularly credible, and they all underestimate human agency and resilience—if AI really killed 10% of humanity, we wouldn't just sit there; historically humans have always fought back against threats.

What I worry more about is P(Catastrophe), which is disasters that don't reach extinction level but cause significant casualties. For example, a large language model hallucinating in military intelligence processing and nearly triggering a war; for example, deepfakes being deliberately used to incite conflict; for example, agents attacking power or financial infrastructure, triggering chain reactions. The probability of these scenarios is quite high, and we should take them seriously.

There's also P(Dystopia), which I think is almost inevitable—close to 100% likely. Companies like Flock are pushing us toward a surveillance society. I've been saying for years: OpenAI's real profit model is selling user data, just like Facebook did, and they will sell data to governments. They put a former NSA director on their board, acquired a camera company, and are now making their own hardware.

Danny: Speaking of P(Dystopia)... we've been talking about regulation, and it feels like regulation itself has become highly politicized—both parties are using data centers as a target to attack their opponents.

Regulatory Outlook

Gary: Most options currently on the table, except for the Hawley and Blumenthal bill, are almost all impractical. Two extremes: one is no regulation at all—we've already seen so many hacking incidents, doing nothing will only bring more problems, and Trump will be the one held accountable when AI causes some infrastructure problem; the other is Bernie Sanders wanting to legislate to ban superintelligence and imprison those involved, which is equally overcorrecting.

I'm not entirely negative about AI. I think AI for science—used in medicine, materials science, climate science—rather than agents crawling everywhere, could genuinely help us. Completely shutting down AI research is throwing the baby out with the bathwater.

The regulatory framework I've been advocating is similar to the FDA model: requiring proof that benefits exceed costs and risks, reviewed by independent scientists without conflicts of interest, plus post-market continuous auditing mechanisms. Like a nasal congestion medication I used that was later found to cause strokes in young women and was taken off the market. If we can start from these two points—pre-market review plus post-market audit, ensuring information transparency and independent scientist access to data—that's a good starting point.

But in the current political environment, this is very difficult to advance in the US. However, there's another possibility: the Brussels Effect—if enough countries unite to take action and exert market access pressure on other countries, it will have some constraining effect even without US participation.

There's another problem: the Trump administration currently exercises opaque control over AI—for example, he quietly shut down a certain model, but without any public rules, no one knows which products might be taken down. An institution filed a Freedom of Information Act request for relevant rule documents and received a 132-page document that was completely redacted. This itself creates regulatory uncertainty—no one wants to buy American products when the rules are unclear, and this could become another stone rolling down the hill.

Investment Implications and Closing

Danny: Jim, how does the regulatory framework affect your investment thesis?

Jim: Government policy is one of the four horses we mentioned earlier. Any regulatory attempt that could raise the cost of capital could be a turning point—once triggered, venture capital dries up, startups stop burning money, and the entire industry enters a multi-year winter. Though this doesn't necessarily mean the technology itself regresses—more likely investors lose big money while the technology continues to advance, similar to 2001-2002.

Gary: I accept that logic, but I'd add one point: we're actually in the worst state now, which is covert regulation with no rule transparency—no one knows what might be shut down. This itself is the stone Jim mentioned—it could cause people to turn to Chinese models, because "at least I know what the rules are there, and I don't have to worry about my product being taken down someday because Trump and Dario Amodei had a falling out."

Danny: Thank you both very much! Gary Marcus, today was truly fascinating. Jim Chanos, welcome back! Please follow Gary's Substack "Marcus on AI" and Jim's ongoing analysis. See you next time!

Gary / Jim: Thanks for having us, let's talk again next time.

This podcast is for reference only. All views expressed by guests are personal opinions and do not constitute specific investment advice.