Two-Way Street

Up and to the Right, Down and to the Left

A guy named Ruben published the cleanest bull case I've read all year, and he's right about everything except the part that pays. His argument: AI can't be a bubble, just look at the curves. Adoption went from zero to half of all businesses in four years. Spending quadrupled in twelve months. The models keep getting smarter. And the cost of a given task fell something like 79x in a single year — too cheap to meter. Every line up and to the right. Case closed.

Every one of those curves is real. Not one of them is where a bubble lives. He's grading the product. The bubble, if there is one, is in the plumbing.

We've watched this movie. The internet wasn't a bubble in 2000 either — the technology was real, the adoption was real, and the telecom cap tables still went to zero, because the money got committed faster than the cash flows could ever service it. Useful technology and a capital bubble ride the same road all the time. They're on it right now, going opposite directions.

Tell an agent to do a job and watch the cost — the first thing any operator with a budget tells it — and it does the obvious thing. It routes the work off the frontier model to whatever's good enough at a tenth the price. Handed a $240,000 bill for agent API access, SaaStr's agent came back and proposed a $5 Postgres box instead, and offered to wire it up that afternoon. That's demand, and it's flowing down and to the left, toward cheap and good-enough.

Now look at where the money went. Anthropic has signed for $518 billion of compute it mostly can't cancel. Broadcom just handed it a $42 billion loan to buy the chips — the supplier is now the bank, and takes the hardware if a payment slips. Every dollar of that is a bet on the frontier. The frontier is the one place the agents are built to route around.

That's the two-way street. Consumer demand for AI, up and to the right, exactly like Ruben says. The dollars, committed up at the top of the stack while the usage drains to the bottom. Both true, same week, same road.

Which is why the Anthropic IPO, now penciled for November, is bigger than one company's payday. It's the first time this gets marked to market — priced in public, into a tape carrying a 19-year high in the 10-year, 7.6% mortgages, diesel near six bucks, data centers slipping, and a stretched consumer, and not just here. The technology isn't the bubble. The bet that demand stays where the money went is. Your own agents are voting against it every time they pick the cheaper model.

At COAI today: the full Signal/Noise — the four bull curves, the routing mechanism, Broadcom-as-the-bank, and why November is the moment it gets priced — is live at getcoai.com.

Outsider Labs Advisory →

79x. That's how much cheaper a given AI task got in one year — the number the bulls quote to prove there's no bubble. Flip it over. If today's work costs a 79th as much next year, and your agents keep routing to the cheapest model that clears the bar, then the half-trillion in frontier compute everyone just financed is pointed at demand walking downhill. The same number that says the technology is real says the capital might not get paid.

1. For the investor — underwrite the frontier's balance sheet, not its benchmark. Anthropic signed for $518 billion in compute and Broadcom lent it $42 billion to buy the chips; the supplier is now the creditor. Before you touch the November IPO, ask the question a bull market hates: who gets paid, and in what order, if revenue grows slower than the lease? The technology can be real and the equity still be the tranche that eats the loss.

2. For the business owner — give your agents a budget today and watch where they go. The operators already doing it are routing off the premium models and dropping the vendors that meter them. Run the test on your own stack today: cap the spend, tell the agent to hit the job anyway, and count how little frontier it actually needs. Most of you are paying for Einstein to summarize meeting notes.

3. For the parent — teach the kid to price things, not just use them. The agent will out-execute them on the task and keep getting cheaper at it. What it can't do is tell whether a thing is worth what it costs, or who gets paid when the music stops. Raise someone who asks where the money really goes. That skill got rarer this cycle, not cheaper.

Monday, October 5

— Harry and Anthony

Sources: