You Go First

THE NUMBER: $518 billion. That’s the compute Anthropic has committed to buy, over the coming years, and most of it cannot be canceled. Hold it against the other numbers in the same filing: about $4.59 billion of revenue last year, and roughly $8 billion of operating loss running the business. The company is spending something like $1.60 on compute for every dollar it earns, and it has signed for half a trillion more. For every dollar of revenue in 2025, it committed about $113 of future infrastructure. Read that ratio twice. Under any normal sky, that is not a company you can finance. It only pencils as something else: a wall the whole market has decided to lean on, because the alternative is finding out what’s behind it.

The scary sentence is the decoy

Everybody spent Tuesday quoting the same part. Anthropic’s S-1 leaked Monday, Reuters and the FT got a look, and by the afternoon the internet had the screenshot: of 261 pages, roughly 80 are risk factors and only about 48 describe the actual business. The filing says advanced AI could pose “catastrophic or existential risks to humanity.” It names self-preserving behavior, models that resist shutdown, systems that conceal or manipulate. A senior safety researcher went on the record putting the odds that AI kills everyone inside a decade north of ten percent.

Michael Burry read the same document and said the quiet part out loud: “IPOs need hype and puffery; ‘we are so awesome it could become dangerous’ is hype and puffery.” He’s right, and it’s the most useful sentence written about this filing. The doom is a flex. We might be so smart we end you. We might be so successful we devour the sun. It’s the risk-factor section as a humble-brag, and it does real work: it fixes your eyes on the apocalypse so you don’t do the arithmetic on the income statement.

So do the arithmetic. That’s where the actual risk lives, and it isn’t extinction. It’s insolvency dressed as inevitability.

The math nobody screenshots

Four and a half billion in revenue. Eight billion in operating loss — and forget the $42 billion headline number, which is mostly accounting noise from warrants and convertibles. The business itself lost about two dollars for every one it took in, and the losses are getting bigger, not smaller.

Then the commitments. Five hundred eighteen billion dollars of cloud and compute, spread across the people who also fund the company: roughly $161 billion to Broadcom, $111 billion to Google, $110 billion to Amazon, $31 billion to Microsoft. About 80% of it binding. Anthropic ended the year with roughly $20 billion of cash. So the mismatch isn’t subtle. It’s a company that could write a check for one year of its own promises and then would be out of money, staring at a decade of bills it can’t cancel.

And the customer base under all of it is thinner than the growth rate makes it look. Nearly a quarter of last year’s revenue came from two customers, and neither is on a long-term contract. Those two customers are Amazon and Google — the same two names on the other side of the compute ledger, the ones whose cloud Anthropic is renting with the money those companies put in. Investment goes in one door and comes back out the other as cloud spend. There’s nothing illegal about it. But it makes it very hard to tell how much of this is real outside demand and how much is Big Tech funding a market around itself and booking the revenue on the way through.

We wrote this exact shape on Thursday, in Other People’s Money: the best model on somebody else’s balance sheet becomes a tenant of its own competitor. The S-1 is the itemized lease.

Why it gets done anyway

So why does a deal like this get done? Because it has to. This isn’t a company going public. It’s a load-bearing wall, and the people holding the beams above it need it to hold.

Look at the tape around it, because the tape is doing the talking this week. OpenAI shelved its next model, GPT-6.1 Astra, over safety failures, and it’s been sitting on its own confidential IPO filing for months. Too dangerous, not ready — you first. Then the cleaner tell: Oura, a real company with a real product and roughly $5 billion of prior valuation, pulled its $2 billion IPO hours before pricing on Tuesday and blamed “market uncertainty.” A profitable-ish consumer hardware business with actual customers looked at the window and climbed back inside. Meanwhile the 10-year Treasury just printed a 19-year high, and Goldman says the breadth of the S&P 500 — how many stocks are actually holding the index up — is the thinnest it’s been since the dot-com top. A handful of AI names are carrying the whole tape.

Put those together and you get the real question of the week. If the market won’t take Oura, how does it take Anthropic? A company losing two dollars on the dollar, with half a trillion in commitments, going public into the narrowest market in twenty-five years? The only answer that works is: because it must. Somebody has to walk to the edge of the pit and go in first, to prove to everyone still standing on the rim that there’s a bottom down there. Snakes. Why’d it have to be snakes. Asps, very dangerous. You go first.

Don’t tell me SpaceX proves the window’s open

Here’s where a bull stops me, and it’s worth taking the argument head-on rather than waving it off. SpaceX went public in June at about $1.77 trillion and it held — closed its first day at $160, trades near $147, above the $135 offer. So the window’s open. Giant private tech can list and stick. Right?

Wrong, and the reason is the whole point. SpaceX is a different animal. It owns its own compute — a builder, not a tenant, the exact distinction we drew last week. It has a real monopoly in launch. It has Cursor throwing off actual coding revenue underneath it. And it has, at the helm, arguably the best capitalist alive. Nobody is parodying Elon on SNL for running a company that loses money; they’re parodying Dario for warning that his product might end the world while raising two trillion dollars against it. On top of the business difference, the sky was friendlier in June: oil was lower, the long bond hadn’t cracked to a 19-year high, OpenAI hadn’t sat on its filing, and Oura hadn’t yanked a deal at the altar.

So SpaceX isn’t proof the window’s open. It’s proof the window only opens for a certain kind of company — one that owns its foundation and earns its keep — and that it’s been closing on everyone else since. The market took the equity king with the rockets and the cash flow. Now it’s being asked to take a pure-burn lab into a worse tape. Those are not the same trade, and pretending they are is how you lose money agreeing with a press release.

We’re still dancing

There’s a sentence that explains this whole moment better than anything in the prospectus, and a man who paid for saying it. July 2007, Chuck Prince, then the CEO of Citigroup, to the Financial Times: “When the music stops, in terms of liquidity, things will be complicated. But as long as the music is playing, you’ve got to get up and dance. We’re still dancing.” Four months later the credit markets seized. Prince was gone by November. Citi nearly didn’t survive.

That’s the trade the whole AI complex is running right now, and Anthropic’s IPO is the DJ’s contract renewal. If Anthropic prices and holds, the music keeps playing. Cash covers the operating losses for another year. The half-trillion in capex stays tucked off the balance sheet, in the joint ventures and the private-credit vehicles and the leases where, as Brookings warned last week, the correlated exposures are hard to see before a downturn. Everybody keeps dancing, because the floor feels solid as long as nobody stops to check. But if this one breaks — if the roadshow stalls or the stock cracks the offer and stays there — it’s the signal event. The AI trade re-rates, the index leaning on it re-rates, and the widows and orphans holding the last 20% of a market propped up by seven names find out the floor was a rumor all along. This is why it’s structural, not just risky. Too many people need it to work for it to be allowed to fail quietly.

The growth is real. Look at what it’s made of.

Now the honest other side, because I’m not here to tell you the demand is fake. It isn’t. Anthropic’s run-rate reportedly hit $65 billion in July and could touch $100 billion by year end. That’s real, and it’s ferocious, faster than almost anything in the history of software. The bull case rests on it: grow into the commitments before the commitments crush you.

But look at what the growth is made of, because not all revenue is the same quality. The one use case that reliably pays is coding — and it pays because it’s the one place the output verifies itself. The code runs or it doesn’t. You don’t need a committee to grade it; the compiler does. That’s why the coding tier is where the money and the retention are.

Now look at everything else the labs shipped this month. xAI’s Grok Bot in August. Meta’s Muse in early September. OpenAI’s Dots on Tuesday. Three companies, seven weeks, one product: the always-on personal agent with its own computer and its own browser. My partner Anthony wrote it up today — “Everyone Built the Same Employee” — and the title is the whole indictment. When three trillion-dollar labs ship the identical thing in a quarter, that thing isn’t a moat. It’s a commodity with a countdown, and the countdown is the price war that starts the moment the second one ships. And unlike coding, the personal-agent swarm doesn’t verify. Muse handed a stranger a seller’s home address this week and reported nothing wrong. Dots failed its own live demo on stage. You can’t build durable enterprise revenue on a product nobody can check.

Which drops us right back on yesterday’s question, Who Do You Want To Be: the only defensible place to stand is a behavior that’s genuinely new, that only you can deliver, and whose output the buyer can verify. Coding is close. The rest of what’s holding up the $65 billion story is spaghetti on the wall — thrown by three kitchens at once. I don’t see the product that passes all three tests yet. Until somebody ships it, the growth that justifies the half-trillion is narrower and more fragile than the run-rate makes it look.

Follow the balance sheet, not the benchmark

The bears are getting specific, which is its own tell. Bain says the industry has to earn $6 trillion a year by 2031 to justify the data centers going up. Michael Burry rotated from shorts to puts — defined-risk, leveraged, the position you take when you think the break is close, not distant. Bill Ackman says a venture blowup is coming and “AI will make all of us look foolish.” When the skeptics stop waving their hands and start naming dates and buying the crash insurance, the conversation has moved from “is this a bubble” to “which brick pulls first.”

We’ve been saying the same thing in a quieter register for a month. On September 1, in Field of Dreams: be bullish on the build and bearish on the paper. The build is real — the data centers get built, the models get better and cheaper, the technology is not the con. The paper is the part you can lose money on. Anthropic’s S-1 is the biggest piece of paper anyone’s printed yet, and it’s asking you to fund a decade of losses on the promise that the music never stops.

What to actually do

If you allocate capital: know whose concrete you’re standing on. Price the builders — the companies with their own compute and their own demand, whose money is cheap because they don’t need yours. Treat the pure-burn labs as the high-beta expression of the trade, not the ballast. And understand the correlation: when breadth is this thin, the names holding the index up are the same names that fall together. There’s no diversification inside one trade wearing seven tickers.

If you run a business on this stuff: build where the output verifies. Coding pays because you can check it. If your AI product’s value can’t be proven by the person paying for it, you’re renting a commodity that gets cheaper every quarter and you have no wall. “We added a chatbot” is a feature three competitors also added this morning. “Only we can do this, and you can prove it worked” is a business.

And if you’re raising a kid into this: teach the reflex to verify before trusting. The agents your children grow up with will be confident whether or not they’re right — that’s not a bug they’ll patch, it’s what the thing is. The durable skill isn’t prompting. It’s the instinct to ask “how do I know this is true?” before acting. The tools will change by Christmas. The habit won’t.

So, who goes first

The doom in the prospectus is a decoy, and it’s a good one, because it lets everyone argue about the end of the world instead of the end of the quarter. But the real risk factor isn’t that Anthropic’s model wakes up and kills us. It’s that Anthropic’s IPO is the load-bearing wall under a market that can can’t afford to see what’s behind it, and one profitable little ring company already looked and walked away.

Somebody has to go first into the pit. Just don’t mistake the dancing for a floor.

— Signal/Noise by CO/AI

SOURCES

Anthropic’s S-1 Is Here — The $518 Billion Commitment Is the Real Story (Yahoo Finance, Sep 29, 2026)

8 Things That Don’t Add Up About Anthropic’s IPO — Burry quote, $8.06B operating loss, $42B headline loss, SpaceX benchmark (BeInCrypto via Yahoo, Sep 29, 2026)

What Anthropic’s IPO Prospectus Actually Says — existential-risk language, GPT-6.1 Astra, Muse address leak (Tech Insider, Sep 29, 2026)

Anthropic’s IPO prospectus shows sweeping AI vision, surging costs (Reuters via CNBC, Sep 28, 2026)

Smart ring maker Oura postpones IPO due to market “uncertainty” (CNBC, Sep 29, 2026) and Oura IPO Pulled Hours Before Pricing (Tech Times, Sep 29, 2026)

S&P 500 breadth hits lowest level since dot-com bubble, Goldman says (Investing.com, Sep 2026)

10-year Treasury yield rockets to 19-year high (CNBC, Sep 23, 2026)

OpenAI launches Dots, always-on AI agents with their own cloud computers (SiliconANGLE, Sep 29, 2026)

Vibe Check: OpenAI DevDay 2026 (Every, Dan Shipper, Sep 29, 2026)

Everyone Built the Same Employee (Anthony Batt, CO/AI, Sep 29, 2026)

AMD agrees to buy Fei-Fei Li’s World Labs for $8.2 billion (The Implicator, Sep 29, 2026)

Segmentation Drives Market Share Wins in AI (Tomasz Tunguz, Sep 29, 2026)

AI industry needs to earn $6 trillion by 2031 to justify data centres (Bain, Sep 2026)

Bill Ackman on AI and a coming venture blowup (Sep 2026)

Chuck Prince, “As long as the music is playing, you’ve got to get up and dance” (Financial Times, July 2007)

SpaceX public listing at ~$1.77T, June 12, 2026 (BeInCrypto, Sep 29, 2026)

Financing the AI Buildout — off-balance-sheet correlated exposures (Brookings/Van Nieuwerburgh, Sep 23, 2026).