You Go First

Asps. Very Dangerous.

Everybody's quoting the scary part. Anthropic's S-1 leaked Monday, and by Tuesday the whole internet had the same screenshot: eighty pages of risk factors, the phrase "catastrophic or existential risks to humanity," a safety researcher putting the odds AI wipes us out inside a decade north of ten percent. Michael Burry read it and called it what it is. "We are so awesome it could become dangerous" is hype and puffery. He's right. The doom is a humble-brag. We might be so smart we kill you; we might be so successful we devour the sun. Skip it.

Read the math instead. Four billion in revenue. An eight-billion-dollar operating loss, and forget the accounting noise from the warrants, because the business itself lost two dollars for every one it took in. Then five hundred billion in compute already promised to Amazon, Google, Microsoft and Broadcom, most of it non-cancelable. Half a trillion in commitments against an eight-billion-dollar hole. Under any normal sky that is unfinanceable, and every banker underwriting it knows it.

So why does it get done? Because it has to. This isn't a company going public. It's a load-bearing wall. Look at the tape around it. OpenAI shelved its next model and sat on its own IPO: too dangerous, not ready, you first. Oura, a real company with a real product, pulled its two-billion-dollar deal hours before pricing on Tuesday, blaming "market uncertainty." The 10-year Treasury just hit a nineteen-year high, and Goldman says the S&P's breadth is the thinnest since the dot-com top, a handful of AI names holding the whole index off the floor. Somebody has to go into the pit first and prove there's a bottom. Snakes. Why'd it have to be snakes. Asps, very dangerous. You go first.

If Anthropic prices and holds, the music keeps playing. Cash covers the losses, the capex stays off the balance sheet, and everybody gets up and dances. Chuck Prince said exactly that in July 2007, four months before the floor gave way under Citi. If it breaks instead, it's the signal event: the AI trade cracks, the index leaning on it cracks, and whoever's holding the last twenty percent finds out the floor was a rumor.

And the growth that's supposed to justify all of it? Real. A sixty-five-billion-dollar run-rate, enormous. But built mostly on coding, the one job AI does that actually pays and actually verifies, because the code runs or it doesn't. Everything else is Grok Bot, Muse and Dots: three companies shipping the identical agent in case one of them sticks. Three copies of the same product don't make a profit pool. They make a price war with a countdown.

At COAI today: the full Signal/Noise — why SpaceX was a different animal, and what "you go first" costs the one who jumps — is live at getcoai.com.

Outsider Labs Advisory →

$518 billion. That's the compute Anthropic has committed to buy and can't cancel, against four billion in revenue and an eight-billion-dollar operating loss. Two dollars out for every dollar in, and half a trillion more already promised to the same clouds that fund it. No normal company finances that. This one might, because the market needs it to: if Anthropic can't go public and stay there, the money stops and the whole AI trade finds out there was never a floor.

For the investor — own builders, rent nothing from tenants. Anthropic's own filing draws the line: four billion in, eight billion lost, half a trillion owed to the clouds that also fund it. That's a tenant on somebody else's balance sheet. Price the concrete owners with their own compute and their own cash flow; treat the pure-burn labs as the beta, not the ballast. When breadth is this thin, the fall is correlated.

For the business owner — sell the thing that verifies. Coding pays and the personal-agent swarm doesn't, for one reason: you can check the output. The code runs or it doesn't. If you're building on AI, build where the result is checkable and the behavior is new. "We added a chatbot" is spaghetti. "Only we can do this, and you can prove it worked" is a business. Yesterday's test still holds: unique, new, verifiable.

For the parent — raise a checker, not a believer. The agents your kid grows up with are confident whether or not they're right. Meta's Muse handed a stranger a seller's home address this week and reported nothing wrong. The durable skill isn't prompting; it's the reflex to ask "how do I know this is true?" before acting on it. Teach the question, not the tool. The tool will change by Christmas.

Wednesday, September 30

— Harry and Anthony

Sources: