In September 2004, Paul Graham wrote an essay that should be required reading for anyone trying to understand what's happening with AI right now. He'd had a front-row seat at Yahoo during 1998 and 1999, watching the company trade at $200 while calculating its true value at $12. He knew the dot-com era was a bubble. But his essay wasn't called "What the Bubble Got Wrong." He titled it "What the Bubble Got Right."
Graham's central insight: You need something real at the center to get a really big bubble. The internet was genuinely important. The bubble was right about that. It was wrong about nearly everything else, including which companies would matter and how long things would take. But the pendulum swinging against everything associated with the bubble, he argued, was an even bigger mistake than believing the hype.
Three decades later, we're watching a similar dynamic unfold. The AI bubble is either forming, inflating, or about to pop, depending on whom you ask. The Shiller CAPE ratio sits near 40, a level exceeded only at the peak of the dot-com mania. Nvidia's market capitalization reached $5 trillion, making it the most valuable company in the world. And the money flowing between a handful of AI companies has become so circular that some analysts warn that a weak link could threaten the viability of the whole industry.
So let's ask Graham's question: What is the AI bubble getting right?
The Ponzi Scheme That Wasn't (Or Was It?)
Graham described Yahoo's position during the bubble in terms that sound uncomfortably familiar:
"Investors looked at Yahoo's earnings and said to themselves, here is proof that Internet companies can make money. So they invested in new startups that promised to be the next Yahoo. And as soon as these startups got the money, what did they do with it? Buy millions of dollars worth of advertising on Yahoo to promote their brand. Result: a capital investment in a startup this quarter shows up as Yahoo earnings next quarter, stimulating another round of investments in startups."
Now consider what's happening with AI. Nvidia plans to invest $100 billion in OpenAI. OpenAI will use that money to buy Nvidia chips. AMD struck a similar arrangement, offering OpenAI warrants for 160 million shares while OpenAI commits to buying billions in AMD processors. Oracle agreed to spend about $40 billion on Nvidia's chips to power OpenAI's data centers. The interconnections multiply: Nvidia has a stake in CoreWeave, which provides AI infrastructure to OpenAI, which buys cloud computing from Oracle, which buys chips from Nvidia.
Graham called Yahoo's arrangement "in effect, the center of a Ponzi scheme," then immediately added: "What made it not a Ponzi scheme was that it was unintentional."
The AI circular financing debate hinges on the same distinction. Acadian Asset Management's analysis noted that Graham's scenario wasn't truly circular because "there are outside investors who are pumping money into the system." A Ponzi scheme always needs fresh outside money. By that definition, "No outside investors, no Ponzi."
But that argument has a problem. Microsoft has poured over $13 billion into OpenAI. Venture capital flooded in: in early 2025, 58% of all global VC funding went to AI startups. SoftBank's $500 billion Stargate project pulls in yet more outside capital. The outside money exists. It's just flowing through increasingly convoluted channels before landing, eventually, at Nvidia.
What the Numbers Say (And What They Don't)
The valuation comparisons cut both ways, and the people making them tend to cherry-pick whichever metrics support their argument.
Bears point to the Shiller CAPE ratio hovering near 40, which is higher than at any time besides the dot-com peak. They note that 30% of the S&P 500 and 20% of the MSCI World index is now held by just five companies, the greatest concentration in half a century.
Bulls counter that the top four tech leaders of early 2000 traded near 70 times 2-year forward earnings, while today's hyperscalers trade at about 26 times. Today's giants generate enormous cash flows. Fed Chair Jerome Powell noted that AI companies "actually have earnings and stuff like that," unlike the profitless wonders of 1999.
Both sides are correct. The valuation metrics are elevated but not insane. The market concentration is concerning but perhaps justified. The real question isn't whether current prices are sustainable. It's whether the money being deployed will generate returns.
The $400 Billion Question
Amazon, Google, Meta, and Microsoft are collectively sinking around $400 billion on AI this year, mostly for data centers. Some of these companies are devoting half their current cash flow to construction.
"Every iPhone user on earth would have to pay more than $250 to pay for that amount of spending," noted Paul Kedrosky. "That's not going to happen."
Here's where the bubble skeptics make their strongest case. An MIT study published in August 2025 found that 95% of enterprise AI pilots deliver no measurable profit impact. Despite $30-40 billion in enterprise investment, the vast majority of organizations saw zero return.
The methodology drew criticism. One analyst noted the study defined success narrowly as "deployment beyond pilot phase with measurable KPIs" within six months, ignoring efficiency gains, cost reductions, and other benefits. Fair point. But the study asked the question that matters: are companies actually making money with this technology?
The answer, so far: mostly not yet. The MIT research found that over half of generative AI budgets go to sales and marketing tools, even though the biggest measurable ROI comes from back-office automation. Companies are investing where AI feels exciting, not where it actually works.
The DeepSeek Tremor
On January 27, 2025, the AI bubble received its first serious stress test. Nvidia lost nearly $600 billion in market value in a single day, the largest one-day loss for any company in stock market history. The trigger wasn't a financial scandal or earnings miss. It was a chatbot from China.
DeepSeek claimed it built an AI model rivaling ChatGPT for less than $6 million, while American companies had spent hundreds of millions or billions on theirs. Worse for the bulls, DeepSeek trained on mid-range Nvidia chips, not the latest high-end hardware.
Venture capitalist Marc Andreessen called it "a Sputnik moment." The comparison was apt. Just as Sputnik forced Americans to question whether their technological lead was as secure as they'd believed, DeepSeek raised uncomfortable questions about whether the hundreds of billions being thrown at AI infrastructure might be unnecessary.
The market recovered, mostly. Nvidia's shares bounced back 8.8% the next day. But the DeepSeek shock revealed something important: the AI investment thesis depends heavily on the assumption that training frontier models requires nearly unlimited compute. If that assumption is wrong, a lot of expensive data centers become white elephants.
Michael Burry Enters the Chat
When Michael Burry bets against something, people notice. The investor portrayed in "The Big Short" has been spectacularly right once, about the 2008 housing collapse, and wrong several times since, including calls that market crashes were imminent when they weren't. His track record is mixed. But his arguments deserve examination.
Burry called the AI boom a "glorious folly," comparing Nvidia to Cisco, the company at the center of the dot-com bubble that lost 80% of its value when the bubble burst. In November, his hedge fund disclosed over $1 billion in put options against Nvidia and Palantir.
His specific critique: hyperscalers are extending the estimated useful lives of their AI hardware to lower depreciation expenses and inflate earnings. If companies assume GPUs last six years when they actually need replacing every two or three due to rapid technological advancement, their reported profits are overstated.
Nvidia pushed back in a seven-page memo to analysts, saying its hardware "remains productive far longer than critics say, thanks to efficiencies driven by the company's CUDA software system." CEO Jensen Huang dismissed bubble concerns, noting his investments represent "a tiny percentage" of Nvidia's revenues.
Palantir's CEO Alex Karp had a more colorful response to Burry's short position: "The idea that chips and ontology is what you want to short is bats--t crazy."
Maybe. But Burry's argument about depreciation is worth taking seriously, even if his timing proves wrong. Companies have an incentive to be optimistic about how long their expensive hardware will remain useful. The dot-com era saw similar creative accounting.
What the Bubble Is Getting Right
Graham identified ten things the dot-com bubble got right. Some apply directly to AI.
The technology is genuinely important. Graham wrote: "The Internet genuinely is a big deal. That was one reason even smart people were fooled by the Bubble. Obviously it was going to have a huge effect." The same is true of AI. ChatGPT reached 400 million weekly active users by early 2025. OpenAI's revenue is running around $13 billion a year, and Anthropic is targeting $9 billion. Real products. Real users. Real revenue. Whether the valuations are justified is a separate question from whether the technology matters.
Most winners will be indirect. Graham observed that "most of the money to be made from big trends is made indirectly. It was not the railroads themselves that made the most money during the railroad boom, but the companies on either side." The AI equivalents might be companies using AI to become more efficient, not companies building AI. The MIT research found that back-office automation delivers the biggest ROI, not flashy customer-facing AI products. The winners might be boring.
Capital allocation is shifting. AI-related capital spending accounted for over 1 percentage point of U.S. GDP growth in early 2025. That's a real economic phenomenon, not just stock market froth. Whether it's misallocated is debatable. That it's happening is not.
Productivity gains exist, even if measuring them is hard. UC Berkeley researchers argued that the 95% "failure" rate might reflect organizations "measuring the wrong things at the wrong time." Individual productivity gains from AI are real and documented, even if P&L impact at the enterprise level remains elusive. The internet looked similarly unproductive in traditional metrics for years before its effects became undeniable.
What the Bubble Is Getting Wrong
The parallels to dot-com failures are also striking.
The circular financing is concerning. Graham noted that Yahoo's Ponzi-like arrangement was "unintentional." Today's AI circular deals look considerably more deliberate. Nvidia's equity stakes in companies like OpenAI and CoreWeave enable those companies to access debt financing at lower interest rates. That's not a side effect. It's a strategy. When chip companies are financing their own demand, distinguishing real growth from artificial inflation becomes nearly impossible.
The vendor financing echoes Cisco. During the late 1990s, Cisco launched extensive vendor-financing schemes that allowed customers to buy equipment they couldn't afford. When the telecom bubble burst, Cisco faced excess inventory, price collapses, and plummeting demand. The structure of AI deals, where Nvidia invests in customers who then buy Nvidia products with the investment, rhymes uncomfortably.
The off-balance-sheet arrangements are creative. Meta recently structured a data center deal where Blue Owl Capital took out a $27 billion loan backed by Meta's lease payments, but the debt never shows up on Meta's balance sheet. These special purpose vehicles aren't necessarily nefarious, but they make it harder to understand the true financial exposure across the AI ecosystem.
The "picks and shovels" bet may be overdone. The conventional wisdom is that the safest AI investment is in the infrastructure providers, not the AI companies themselves. Nvidia, the argument goes, makes money regardless of which AI company wins. But Burry's Cisco comparison points to a problem: the picks and shovels companies were the hardest hit when the dot-com bubble burst. In the early 2000s, less than 5% of U.S. fiber capacity was actually operational. Today's data centers could face a similar overbuilt future.
Three Scenarios
How does this end? The most likely paths:
The soft landing. AI adoption accelerates, enterprise ROI improves, and valuations gradually become justified by actual earnings. This is the bull case, and it requires AI to deliver on its productivity promises faster than skeptics expect. BlackRock notes that investor behavior remains "measured" compared to the dot-com era, with net outflows from U.S. equity funds even as AI stocks surge. If exuberance is contained, a correction could be mild.
The dot-com replay. Investment exceeds realistic returns, a catalyst triggers panic, and the weakest players collapse while taking the ecosystem with them. Economists warn that household stock market participation is higher now than in 2000, meaning a burst could be more painful. The interconnected nature of AI financing means failures could cascade quickly.
The slow deflation. The most probable outcome may be neither crash nor soft landing but a prolonged period of disappointing returns. Analysts at IMF and Goldman Sachs suggest that cash-funded AI investments make a sharp 2000-style crash less likely, but a multi-year correction more probable. Some companies thrive. Most don't. Investors slowly realize that valuations were too high, but there's no single moment of reckoning.
The Meta-Lesson
Graham's essay contained a warning that applies equally now:
"Now the pendulum has swung the other way. Now anything that became fashionable during the Bubble is ipso facto unfashionable. But that's a mistake, an even bigger mistake than believing what everyone was saying in 1999."
The dot-com crash didn't prove the internet was unimportant. It proved that investors got the timing, the companies, and the business models wrong. Google went public in 2004, after the crash. Amazon survived to become one of the most valuable companies in history. The internet mattered. Most internet stocks didn't.
The same will likely be true of AI. The technology is real. The productivity gains are coming. But not all AI companies will survive to capture them. Not all AI infrastructure being built will be needed. Not all valuations will be justified.
The question isn't whether AI is in a bubble. Bubbles are definitional, and you can argue either side with the available data. The question is what the bubble is getting right, because those things will persist after whatever correction comes. And the question is what positions you want to hold when we find out.
