The $965B Ghost in the Machine: A Forensic Read of the Anthropic Bubble
The chart says the AI trade is invincible. The math says someone is pricing a company at nine times the most expensive public listing in American history and calling it fair value. When a number that large lands on my desk, I stop reading the press release and start reading the receipts.
Anthropic — the safety-first AI lab — is being whispered toward a $965 billion private valuation and an IPO before Thanksgiving. On paper, that is a company worth 78 profitable S&P 500 corporations combined. Michael Burry, the man who called the 2008 housing collapse and then misjudged the 2023 rebound, threw the first rock. He called it a bubble. The internet erupted, as it always does.
I have spent the last decade doing something unglamorous: reading the receipts. Gas receipts. Treasury transfers. Pool balances. When a $965 billion number arrives with no audited ledger behind it, my instinct is not to argue about whether AI is overhyped. My instinct is to trace where the money actually goes — because the ghost in a valuation always leaves a trail, and that trail is almost never in the narrative. It is in the cash flow.
This is that trail.
Context: Setting the Board
Let me lay out the pieces before I start pulling threads, because a forensic case lives or dies on the quality of its evidence chain.
Anthropic is one of a handful of frontier AI labs. It builds large language models, sells API access, and — until recently — looked like a software company with software margins. Then the capex number leaked: $518 billion in cloud and infrastructure commitments. That single figure changes the species of the animal. Anthropic stops being a model company and becomes a compute company — a heavy-asset, capital-hungry infrastructure bet dressed in the clothing of a lean software startup. The costume is convincing. The balance sheet underneath it is not.
The reported numbers, all of them second-hand from a leaked prospectus and a chain of Fortune, Reuters, and CNBC retellings, sketch a strange picture:
- Private valuation: $965 billion
- Q2 2026 revenue: $11.5 billion, annualizing near $46 billion
- 2025 net loss: $42 billion — most of it, per Reuters, accounting charges rather than operating cash out the door
- Two consecutive quarters of operating profitability
- A $65 billion equity round led by Altimeter, Dragoneer, Greenoaks, and Sequoia
- Potential fair value floated by prospective investors: $1.8 to $2 trillion
Burry's bear case is arithmetic. At $46 billion in annualized revenue, a $965 billion valuation implies roughly 21x price-to-sales. The largest IPO in history — UPS, at $119 billion — priced at 2.4x sales. Anthropic's multiple is about nine times that record. His rhetorical anchor — 78 S&P 500 companies — is a category error dressed as a warning: it compares a hypergrowth AI lab to mature value stocks and lets the sheer size of the comparison do the emotional work.
Before I go further, I have to flag something, because it determines how much weight any of this deserves. The core events in this story — the Burry tweet dated October 2026, an IPO targeted before Thanksgiving, a Q2 2026 earnings release — sit beyond the range I can independently verify. The headline figures — $965 billion, $42 billion, $518 billion — come from a leaked prospectus and secondary reporting, with no primary document I can pull. There is no on-chain ledger for a private AI lab. There is no block explorer for a leaked prospectus. A $965 billion valuation is an announcement. A $518 billion capex commitment is a settlement in progress. Those are two very different kinds of claims, and only one of them shows up in a cash flow statement.
So treat what follows as a structural read of the logic, not a verification of the facts. I spent six weeks in 2017 pulling apart ERC-20 contracts for a Riyadh venture fund, and the lesson that stuck was brutal: the events that matter are the ones that settle, not the ones that get announced. The numbers here smell too clean. That is exactly when I go looking for the circularity.
Core: The On-Chain Evidence Chain
Here is the thing about AI revenue that nobody in the bull camp wants to say out loud, and that the bear camp keeps gesturing at without naming: a huge slice of it may be moving in a circle.
Think about the structure. A cloud provider — call it a hyperscaler — supplies compute to an AI lab. That same hyperscaler, or a fund affiliated with it, invests equity into the lab. The lab then spends that invested capital buying compute back from the supplier. Revenue is recognized. Investment is booked. The money orbits.
I have seen this exact shape before, and I have seen it used to paint a lie. In the DeFi Summer of 2020, I deployed $50,000 of my own ETH across Uniswap V2 and SushiSwap to study yield volatility. The most important thing I learned was not about impermanent loss. It was about how easily volume can be manufactured. Wash trading — a wallet swapping with itself, or with a coordinated cluster — inflates the activity metric that everyone uses as a proxy for health. The pool balance barely moves. The volume counter spins like a slot machine. Traders chase the number; the number was never real.
This is the forensic lens I now apply to AI revenue. When your supplier is also your investor and also your distribution channel, the revenue line stops being a measurement of external demand and becomes a measurement of internal circulation. I am not saying Anthropic's revenue is fake. I am saying the burden of proof is on whoever claims it is clean, and no one has shown me the external customer concentration, the net revenue retention, or the gross margin that would settle it. The bull case leans on revenue quality it has not demonstrated. That is the gap. Reading the pulse in the pool balance means asking not how big the number is, but where the number came from and who else touched it on the way.
Now look at the loss. A $42 billion net loss that is, per the reporting, mostly accounting charges. Read that again. In accounting, a charge that is not cash leaving the building is a story about the future, not the present. If the bulk of that $42 billion is equity compensation or fair-value adjustments tied to the $965 billion mark, then the loss is a mirror of the valuation, not a symptom of the business. Mark the valuation up, and the loss balloons. Mark it down, and it vanishes — or flips into a paper profit. Either way, the headline number becomes a weapon for whichever side is holding it that week.
I have watched this movie. When Celsius froze withdrawals in June 2022, I sat in living rooms in Riyadh collecting retail investors' stories while my screens tracked the 6,000 BTC that moved out of the treasury. The financial statements and the on-chain truth told two different stories for months. The statements said yield. The blockchain said the yield is your principal, recycled. The signature is in the silent transfer — the movement that does not appear in the marketing, only in the ledger.
So let me do the arithmetic the way I would audit a treasury drain, because the numbers here tell a story that neither camp has fully told.
The capex commitment is $518 billion. The equity raise is $65 billion. That is an eight-to-one gap between what the company has committed to spend and what it has raised to spend it. I have followed capital through validator mazes and across bridge contracts, and one rule never breaks: a commitment larger than its funding source is not a plan, it is a dependency. It means Anthropic will need to keep raising — equity, debt, or operating cash flow — continuously, for years, just to service commitments it has already signed.
That reframes the IPO entirely. The bull camp reads the Thanksgiving IPO as a triumphant coming-out. The bear camp reads it as a bubble popping. The forensic read is colder: the IPO may not be an exit at all. It may be a lifeline. When your capex outstrips your raise by eight times, the public markets stop being a victory lap and start being a cash faucet. This is the motive nobody has stated plainly, and it is the one I would bet on. Following the money through the validator maze, you eventually learn that the loudest announcement often covers the quietest need.
Let me push further, because there is a second ledger hiding in plain sight — the crypto one, and this is where I have an unfair advantage over the equity analysts shouting about P/S ratios.
Every dollar of that $518 billion capex flows somewhere physical: GPUs, data centers, and — the part everyone underestimates — electricity. This is where blockchain data becomes genuinely useful as a cross-check on the AI narrative. The decentralized compute tokens — the Renders and Akashs and TAOs of the world — are priced, in part, as a leveraged bet on exactly this compute scarcity. When I read the pulse in those pool balances, I am reading the market's own forecast of how tight the compute supply will get. If the AI capex story is real, the demand for compute-adjacent assets should show up as sustained accumulation, not just a spike on a headline. If it is narrative, the pools drain the moment the story goes quiet. The pools are the lie detector the equity market cannot access.
And here I want to be honest about a bias I have earned. I have long argued that liquidity fragmentation in crypto is not a real problem — it is a manufactured narrative that venture firms use to sell new products. The same pattern is now visible in AI. Every week a new solution to an AI infrastructure gap gets funded, and the gap is often defined by the people selling the bridge. When I see a $1.8 to $2 trillion fair value floated by prospective investors, I do not see a valuation. I see a pre-IPO anchor — the standard investment-banking move of naming a big number early so the eventual pricing range looks reasonable by comparison. Hunting liquidity where the charts lie means recognizing that the chart itself was drawn by someone who wants you to read it a certain way.
I watched the same trick in 2024, during the Bitcoin ETF flow attribution. For three months I tracked daily flows from Grayscale and BlackRock custodians, following 120,000 BTC of movement, correlating ETF inflows against exchange reserves. What made that work was not the size of the flows. It was the persistence. Institutional accumulation does not announce itself with a spike; it announces itself with a slow, relentless drain of available supply that the price is eventually forced to honor. That is the standard I would apply here. If Anthropic's revenue is real, it should behave like institutional accumulation — steady, external, and boring. If it is circular, it will behave like wash trading — loud, self-referential, and gone the moment the music stops.
Here is the comparison that actually matters, and it is not UPS.
It is the ICO class of 2017. I audited 15 of those token contracts in six weeks and found reentrancy holes in three of the highest-profile projects — flaws that would have cost retail investors an estimated $4.2 million. What I remember most is not the bugs. It is the multiples. Tokens were priced at valuations that implied they would capture entire industries, and the pricing was set by private rounds with no public constraint, exactly like a $965 billion private AI mark. The public listing was supposed to be the moment of truth. For most of them, it was the moment of collapse, because the private price had been fiction all along and the public market was the first place anyone could short it.
I am not predicting Anthropic collapses. I am pointing at the mechanism. A valuation set in private, with no public price discovery, and defended by the same investors who profit from it, is not a valuation. It is a hypothesis. The IPO is the experiment that tests it. That is the whole game, and both camps are arguing about the result of an experiment that has not run yet.
Contrarian: Where I Disagree With Everyone
Now the part where I break from almost everyone, including people who would call themselves my allies.
The bear case — it is a bubble — may well be directionally right and still be the wrong reason. Burry's arithmetic is clean, but his anchor is dishonest, and a dishonest anchor can lead you to the right conclusion for the wrong cause. Comparing Anthropic's price-to-sales to UPS's is comparing a company whose entire value is a bet on future cash flows to a company whose value is present cash flows. Of course the multiples differ. That is not evidence of a bubble; that is evidence that they are different instruments. If your short thesis rests on the multiple is high, you have described every growth company in history and predicted nothing.
The real risk is not the multiple. The real risk is the capital structure. A company with an eight-to-one gap between commitment and funding is fragile not because it is overvalued but because it is dependent — on the patience of equity markets, on the willingness of hyperscalers to keep financing their own suppliers, on the continuation of a rate environment that tolerates cash-burning infrastructure. Break any one of those and the $965 billion mark is irrelevant; the question becomes whether the lights stay on. That is a solvency question, not a valuation question, and the two get confused constantly.
And here is the correlation-versus-causation trap that the crypto crowd is about to fall into. If AI corrects, the AI-adjacent crypto tokens will fall hard. People will say AI crashed, therefore crypto crashed, therefore crypto is correlated with AI. Wrong. Both are levered expressions of the same underlying variable: the cost of capital and the appetite for long-duration, cash-burning growth. When that variable turns, both fall — not because one caused the other, but because they share a parent. Trading the correlation as if it were causation is how you get liquidated on a headline that had nothing to do with you.
I learned this in the BAYC metadata dive of 2021. I clustered the wallets behind 10,000 apes and found that 40% of early sales traced back to five coordinated wallets. The organic community was a puppet show. But the useful part was not catching the manipulation. It was realizing that the price and the community were both downstream of the same small group of actors — and that when those actors stopped, both would fall together, which would look like correlation and would actually be shared dependence. Decoding the pixelated intent behind the PFP taught me to ask not what moves together, but what shares a cause.
The AI valuation debate and the crypto valuation debate share a cause. Nobody in either camp is trading that. They are all trading the symptom.
There is one more inversion I want to offer, because it cuts against the easy bubble narrative. In 2023, everyone called Bitcoin Ordinals a bubble — a frivolous fad that was clogging blockspace and inflating fees. The lazy take won. The correct take was that inscriptions injected a real, durable stream of fee revenue into Bitcoin at exactly the moment the security budget was starting to look thin. Strip the inscription wave out and Bitcoin's long-term security model gets materially weaker. The thing that looked like a bubble was, in part, the thing paying the bills. I am not saying Anthropic's capex is healthy. I am saying that dismissing a capital-intensive build-out as a bubble, because the number is big, is the same lazy reflex that almost missed what Ordinals actually did for Bitcoin. Size is not the same as waste. The question is whether the spend buys a durable asset or a temporary narrative, and that is answered by the receipts, not the vibe.
Takeaway: What I Watch Next
So what do I actually watch, and what would change my mind? Three signals, and I will be specific, because vagueness is how narratives survive.
First, the external revenue disclosure. If the IPO prospectus shows customer concentration and net revenue retention that prove the $46 billion annualized run-rate comes from outside the orbit of its own investors and suppliers, the circularity worry dies and the bull case earns its multiple. If those numbers are missing or obscured, the circularity worry is confirmed by omission. Silence is a data point. Tracing the ghost in the gas receipts always starts with noticing what the ledger refuses to show.
Second, the funding structure behind the $518 billion. Watch whether the capex is financed by operating cash flow, fresh equity, or debt. Operating cash flow means the business is real. Fresh equity means the business is a bet. Debt means the bet is levered, and levered bets on long-duration growth have a way of becoming forced sellers when the variable turns. Volatility is just data waiting to be tamed — but only if you know which input is driving it.
Third, the compute pools. If the AI capex story is genuine, decentralized compute demand should show sustained accumulation through the noise, the same way institutional ETF flows showed up in exchange reserves in 2024 — not as a spike, but as a slow, persistent drain of supply that the price eventually had to honor. If the pools only move on headlines, the demand is narrative, and narrative does not pay for GPUs.
The $965 billion number is not the story. It is the headline the story is hiding behind. Somewhere under it, capital is moving in circles, commitments are outrunning funding by eight times, and an IPO is being scheduled not as a coronation but as a cash requirement. Whether Anthropic is a bubble or a bargain is a question for the experiment. Whether the receipts will tell you the truth is not a question at all.
Audit trails do not lie. They only wait. And the question worth asking this week is not what the valuation says — it is who is still moving money into the pool when the announcement goes quiet.
