What the AI Bubble Popping Is Likely to Look Like
This piece is a follow-on to What Dot-Com Bubble?
In my earlier piece I argued that the dot-com era wasn’t a bubble in any meaningful sense — the transformation was real, the value created was real, and the NASDAQ sits at roughly 30 times where the “bubble” started. The internet was not a bubble. But Pets.com was. Webvan was. The people who bought the Netscape IPO and held it didn’t get rich — even though Netscape was real, mattered enormously, and made the transformation visible to the world.
AI is real. The transformation is real. That is not the question. The question is how to value new and existing companies that participate in this transformation.
The Exit Is Already Being Built
Let’s be blunt. The excitement around AI is lighting a fire under everyone, and the capital markets are going to capitalize on that excitement. They are counting on the fact that most people either cannot or will not ask the right questions to make sound investment decisions. There is no Consumer Reports for IPOs. There is no independent rating agency telling you what the unit economics actually are. There is the prospectus, in four-point font, and the analyst coverage from the banks that underwrote the deal.
Anthropic closed a $30 billion Series G in February 2026 at a $380 billion post-money valuation. That number did not emerge from a discounted cash flow model. It was preceded by a Wall Street Journal report citing anonymous sources familiar with the deal. Pre-IPO narrative management is standard practice. The WSJ story is how it begins.
The institutional investors who led the Series F got in at $183 billion five months earlier with preferred shares and liquidation preferences.
The investment community can sell the growth everyone can see. The $19 billion revenue run rate growing 10x annually is real — the best possible backdrop for moving preferred shares into a public float. The LP does not care whether Anthropic is profitable in 2030. The LP cares whether the fund returns capital on schedule. They will be gone before the math catches up.
The retail investor buying at IPO is not entering the same trade. They get common shares with no downside protection, no fund lifecycle exit, and no access to the unit economics the institutional investor reviewed before writing the check. They have the revenue growth number in the press release and analyst coverage initiated at “buy” by the banks that underwrote the deal. That conflict is disclosed in the prospectus, in four-point font, across two hundred pages nobody will read.
Wall Street will do fine. They just have to sell the general public on it long enough to not make it look like a scam. That is not cynicism — it is a description of a mechanism that has worked the same way for a hundred years.
The Nifty Fifty Pattern
Not every part of the AI bust will look like dot-com. Some of it will look like the Nifty Fifty.
The Nifty Fifty were the “one decision” stocks of the early 1970s — Xerox, Polaroid, Avon. The growth was so certain that any price was justified. The companies were real. The earnings were real. Then the multiple compressed — some businesses became dinosaurs displaced by the next technology wave, others simply never met the expectations baked into the price. Polaroid fell 90% and eventually went bankrupt. Avon fell 86%. The investors who bought at peak multiples spent a decade underwater — if they got their money back at all.
Anthropic at $380 billion — roughly 20 times annualized revenue, pre-profit, pre-moat, and pre-proof that the business model works at scale — is a Nifty Fifty setup with the downside more acute, not less.
Anthropic is doing genuinely impressive work. The revenue growth is real, the enterprise adoption is real, and the models are good. The question is not whether they are executing well. The question is what that execution is actually worth — and how hard it is for others to replicate it. The team that built Anthropic walked out of OpenAI and reconstructed it in months. What stops the next team from doing the same? Is there a real moat, or is the moat just being first and well-funded in a race where the tools to catch up are the very thing being sold? The revenue is real: $1 billion in December 2024 to $19 billion by March 2026, a trajectory with no precedent in enterprise software history. The company is not Pets.com. But roughly 20 times revenue in a compute arms race with no durable moat requires heroic assumptions about what comes next.
Apply the framework from my earlier piece on Buffett-speak: will the cash you take out comfortably exceed the price you paid before the moat runs out? At roughly 20 times revenue, that math has never been demonstrated at this scale in this competitive environment. Every dollar raised goes to compute. Skip a training cycle and you fall behind. It is Sun Microsystems with better press.
Most people buying AI stocks are not doing this math. If you are not asking whether the discounted free cash justifies the price, you are not making an investment. You are making a bet that someone will pay you more than you paid. That is the greater fool theory — and it works until it doesn’t.
The Gold Bar in Every Box
There is a question nobody is asking loudly enough: will there ever be any free cash?
The standard narrative is that AI losses are investment losses — spend now, revenue catches up, margins follow. That is the Amazon story. Amazon’s losses were building the warehouse network, the logistics infrastructure, AWS. Deposits against future dominance.
Anthropic’s losses are structurally different. The cost of delivering the product is the compute required to run inference on every query. That cost scales with revenue. If unit economics are negative at the product level, more revenue makes the losses larger. It is not an investment loss. It is a product loss. If every box comes with a gold bar inside, you cannot fix that by selling more boxes.
Traditional software economics work because marginal delivery cost approaches zero — a SaaS company at scale runs 80% gross margins because the servers are a rounding error. AI inference does not work that way. Every query costs real money in GPU time, power, and cooling. Underneath that is a capital expenditure layer unlike anything in software history: GPUs at tens of thousands of dollars each, purpose-built cooling infrastructure, power consumption measured in gigawatts. A normal software company’s infrastructure is a line item. An AI company’s infrastructure is the business.
The math that would justify these valuations is rarely shown explicitly. Anthropic’s gross margin is currently estimated at 40 to 55 percent. Mature software companies run 75 to 85 percent. The gap is compute cost — and that is not a temporary inefficiency. It is physics. Transformers require matrix multiplications. Matrix multiplications require GPUs. None of that goes to zero. Discount the projected free cash at a reasonable rate, subtract the capital expenditure to stay competitive, adjust for open source pricing pressure, and ask what number you get. Nobody is publishing that number because it is uncomfortable. At honest assumptions the discounted free cash flow does not get you to $380 billion. It probably does not get you within a country mile.
The gold bar is load-bearing. Inference costs are falling — but competitors’ costs fall at the same rate. And the bar may get heavier, not lighter: each new frontier model generation has required more training compute than the last. Agentic workflows compound this further — an agent completing a multi-step task consumes dramatically more tokens than a simple query, so the cost curve may not fall as fast as hardware improvements suggest even as individual token prices drop. Stay at the frontier and your costs go up. Fall behind and the cheaper alternatives — open source, Chinese labs, distilled models — eat your premium pricing from below. The squeeze runs in both directions simultaneously.
There is a deeper problem underneath this. Traditional accounting valuation was built for a world of physical assets — factories, inventory, land. Things you could walk up to and appraise. Nobody has a valuation methodology that works reliably for frontier AI companies.
The Drug Dealer Business Model
Drug dealers give the first hits away free. The economics only work once the customer is addicted and has no clean exit. What the AI companies are doing is not different — just legal, and the product being pushed is productivity.
Right now, corporate America is being hooked. Agentic AI is being woven into workflows, hiring decisions, and organizational structures. Companies are eliminating headcount on the assumption that AI handles it. The institutional knowledge of how to do things the old way is walking out the door with the people who knew how. Once that restructuring is complete, the switching cost is no longer an API key. It is organizational reconstruction. At that point the AI company can charge what it needs to charge.
This may be the only viable business model available. You cannot charge cost-covering prices today because a competitor is running the same subsidy play. Every major AI company races to get deepest into the enterprise stack before the pricing era arrives. It is a rational strategy — it requires surviving years of losses, which requires the capital raises, which requires the narrative, which requires the IPO. The whole apparatus is in service of buying enough time to get deep enough into enough enterprises that pricing power finally materializes.
The darker version: what if the addiction sets but the pricing power never arrives? Open source is always one generation behind but good enough for most enterprise use cases at zero marginal cost. If Meta or other players give away models for free or as open source, the ceiling on closed model pricing may never cover costs. You have restructured the entire global economy around a product that nobody can profitably deliver. That is not a bubble in any historical sense. That is something stranger and harder to name.
The Spinout Economy
The other part of the bust looks like 1999 — real technology, unwinnable unit economics, and a credential bubble layered on top.
Every senior researcher who left OpenAI got a lavishly funded startup. Ilya Sutskever left — funded at a $32 billion valuation with no product. Mira Murati left — funded. The VP of Research left — funded. The funding happens inside the Sand Hill Road network — the Stanford Mafia — the same tight circle recycling credentials into valuations for thirty years. They went to the same schools, sit on the same boards, validate each other’s numbers. The market cannot absorb fifteen frontier model companies each burning nine figures a year on compute. The network doesn’t care. The network gets its fees regardless.
Yann LeCun’s AMI raised $1 billion at a $3.5 billion valuation — twelve people, Paris headquarters, no product, no demo. The Turing Award is a receipt for work done decades ago, not a forecast. The man who funded LeCun for thirteen years — Mark Zuckerberg — is not in this round. In 1999 every ex-Netscape engineer got funded. The pedigree was real. The market wasn’t big enough for all of them.
What the Bust Looks Like
The dangerous game is picking AI pure-plays at IPO valuations on narrative alone. It is like going to the horse races and trying to pick the winner from the race bill. You studied the form. The professionals who set the odds have forgotten more about those horses than you will ever know. The house takes its cut regardless of who wins.
If AI transforms the economy the way the optimists believe, the S&P 500 will reflect it. The index owned Amazon, Google, Microsoft, and Nvidia through all the volatility and captured every dollar of value they created. It will own whatever the AI winners turn out to be. Do what Buffett says: buy an S&P index fund and let the sorting happen there — without making yourself crazy listening to market pundits, without losing sleep over which frontier model survives, without being on the wrong side of a trade that was designed before you sat down.
The sorting mechanism does not announce itself. It just sorts.
Related reading: What Dot-Com Bubble? · The Buffett Indicator: What It Is — and What It Is Not · The Most Misunderstood Phrases in Buffett-Speak · LeCun’s AMI: What Is the Proposition? · The OpenAI-Anthropic Story