Two-Way Street
Adoption and usage are up and to the right; the bulls have that right. But the money's a two-way street — give an agent a cost target and it routes demand off the frontier, down and to the left, while every dollar of capital is committed up at the top. The Anthropic IPO is where we find out who read the traffic correctly.
THE NUMBER: 79x. That’s how much cheaper a given AI task got in a single year, by the bull camp’s own math — the stat they quote to prove this can’t be a bubble. Hold it the way they hand it to you first: a thing that does the same work for one seventy-ninth of last year’s price is a miracle of deflation, not a mania. Then turn it over, because the same number has a second face. If the work you’re paying for today costs almost nothing next year, and the machine doing it keeps shopping for the cheapest model that clears the bar, then the half-trillion dollars of frontier compute the whole industry just financed is pointed at demand that’s walking downhill, away from the toll booth. The number that proves the technology is real is the same number that asks whether the capital gets paid. One digit, two directions. That’s the whole issue.
A bubble is a mismatch between the money committed and the cash flows that can service it..
Harry DeMott
The whole argument, before the evidence: The bulls are right that AI isn’t a bubble — adoption, capability and price are all screaming up and to the right, and anyone telling you the technology is a fad hasn’t used it. But they’re grading the product, and a bubble doesn’t live in the product. It lives in the plumbing — in the gap between the capital committed and the cash that can service it. And right now the two are moving in opposite directions on the same road. Tell an agent to do the work and watch the cost, and it routes the job off the frontier to whatever’s good enough and cheap. Demand drains down the stack. Meanwhile the money — $518 billion of it, most of it uncancelable — is committed up at the frontier, with the chip supplier now lending the lab the money to buy the chips. The technology isn’t the bubble. The bet that demand stays where the money went is.
The cleanest bull case of the year
Start with the bull, because he’s good and because he’s mostly right. A writer named Ruben published a piece this weekend titled “Bubble-Burst,” and it’s the sharpest no-bubble argument I’ve read in 2026. His method is clean: judge AI on four curves, and watch them all bend the right way. Adoption — zero to more than half of all businesses in four years. Monetization — AI spending quadrupled in the twelve months to February. Intelligence — the models keep getting smarter, nobody disputes it. And affordability — the cost of a given task fell roughly 79x in a single year, what the engineers call too cheap to meter. His kicker is an accounting study: Mercor had twelve accountants and a stack of AI models run the same realistic tasks; in May 2025 the humans won, and a year later the machines were near-perfect. Up and to the right, every line.
I’m not going to argue with a single one of those curves, because they’re true, and because arguing the technology is a fool’s errand that the last three years have punished. When somebody warns me “remember the dotcom bubble,” they usually mean the technology was fake. It wasn’t. That’s exactly the wrong lesson, and it’s the door into what Ruben misses.
He’s grading the product. The bubble is in the plumbing.
Here is the thing a demand curve cannot tell you: whether the people who paid for the supply get their money back. The internet in 1999 had Ruben’s curves and then some. Adoption was vertical. Usage doubled every hundred days. The technology was world-changing and it did, in fact, change the world. And the capital still detonated. Global Crossing and WorldCom went bankrupt, Lucent and Nortel lost the better part of a trillion dollars of market value between them, and something like ninety-five percent of the fiber that got laid sat dark for years. Not unused forever — dark. The fiber was real, it was necessary, and the broadband and streaming era eventually lit all of it. A decade later. By different people. After the ones who financed it were wiped out.
That’s the pattern that actually rhymes, and it has nothing to do with whether the technology works. A bubble is a mismatch between the money committed and the cash flows that can service it, and it resolves by handing the asset to somebody who didn’t pay to build it. The product can be a miracle and the cap table can still be kindling. Both were true in 2001. Both are true now. The only question worth the ink is which side of that trade you’re financing.
Give it a budget and watch it leave
Now the part the bull case skips, and the part that turns 79x from a boast into a warning. The moment you tell an agent to do a job and watch the cost — which is the first instruction any operator with a P&L gives it — the agent does the obvious, ruthless thing. It routes the work to the cheapest model that clears the bar.
SaaStr lived it out loud this weekend. They run the place on three humans and twenty-odd agents, and their revenue-ops agent makes something like forty thousand API calls a day. Nobody metered those calls, so nobody counted them, until the vendors — Salesforce, Atlassian, HubSpot — started charging for agent access and one estimate landed at $240,000 a year. The agent’s own recommendation, once it was told to track its usage: stop calling the vendor, mirror the system of record into a $5 Postgres box, and route around the toll. It offered to set it up that afternoon. That is demand with a brain and a budget, and it is flowing down and to the left, toward good-enough and cheap.
And metering is a confession. The second a vendor charges the agent for access, it’s told the agent it’s a cost to be optimized away — and the agent reads that banner as a kill order. Nate Jones put the mechanism cleanly this week: an agent takes the software you bought as one subscription and splits it into three parts — the data, the interface, the workflow — each now its own keep-or-kill decision. The agent almost never needs the interface. It can mirror the data. It swaps the workflow for its own. What’s left that it still has to rent from the frontier is a sliver, and the sliver is shrinking every time the model below it gets good enough. That’s the routing, mechanized: not a grudge against the premium model, just an optimizer doing arithmetic the vendor handed it.
This is not a one-company anecdote. It’s the shape of the whole market now. Agents already burn five times the tokens humans do — 7.3 trillion against 1.4 trillion on OpenRouter, headed for ten times — and more than eighty-five percent of that is cached prompts, the machine rereading what it’s already seen. The demand is gigantic and real, exactly as Ruben says. It is also, increasingly, indifferent to the frontier. We wrote it three weeks ago in Get Off My Cloud: you own the commodity stack and rent the frontier only for the two percent of the job that truly needs it. Chase Lochmiller, who’s building one of the big compute operators, said the quiet version on 20VC this week — “most moats are illusion.” Even a16z, the house that’s long all of this, titled its own slide “The Great AI Capital Misallocation.” The people pouring the concrete are telling you the economics are soft.
And the money went the other way
So demand is draining toward cheap. Watch where the capital went. Anthropic’s S-1, as reported by CNBC and Reuters, puts its compute commitments at $518 billion, most of it uncancelable, against $11.6 billion of quarterly revenue and a first, thin, positive adjusted operating profit. Of that half-trillion, $161 billion is owed to Broadcom alone. And here’s the turn we’ve been circling since You Go First: Broadcom isn’t just selling the chips anymore. It wrote Anthropic a $42 billion convertible note — roughly a third of a $125 billion TPU lease — which means the supplier is now the lessor and the lender. If Anthropic wins, Broadcom converts to equity. If it stumbles, Broadcom is a senior creditor holding the hardware. The vendor became the bank.
It’s not just Broadcom. Nvidia has committed up to $100 billion to OpenAI; AMD tied a $5 billion stake in Anthropic to the deployment of its chips. The competitive weapon in this business is no longer the silicon — it’s the willingness to lend your customer the money to buy your silicon. That’s a closed loop, and closed loops are wonderful right up until the day the cash has to come from outside the circle. Nearly a quarter of Anthropic’s revenue, per the filing, comes from two customers who are also two of its cloud landlords. Money goes in one door and comes back out the other as compute spend. Nobody’s hiding it. It just makes it very hard to tell how much of this is outside demand and how much is Big Tech funding a market around itself.
Both true. Same week. Same road.
Put the two flows on one map and you get the picture that names the issue. Consumer and enterprise demand for AI, climbing up and to the right, every curve Ruben drew. And the dollars, committed up at the top of the stack — the frontier models, the newest chips, the longest leases — while the usage those dollars are supposed to capture quietly drains to the bottom, to the good-enough model and the $5 database. Two kinds of traffic, opposite directions, one street.
The bet underneath the entire structure is that demand stays at the frontier — that people keep paying premium rates for the best model because the best model is worth it. For the two percent of work that genuinely needs the frontier, it is. But the agent doesn’t do the two percent and the ninety-eight percent at the same price. It does what SaaStr’s did: it sorts the job and sends almost all of it downhill. The frontier is the one place in the stack the agents are built to route around, and it’s the exact place the money went.
November is when it gets priced
This is where I put the finance hat down for a second, because it would be easy to read the above as a crash call, and it isn’t. I’m not telling you to short anybody; I’m a guy reading the tape, and the tape says the fault line is structural, not that the earthquake is scheduled. Private markers can stay irrational for years. What actually breaks a mismatch like this isn’t a high valuation — it’s a forced seller, a moment where somebody has to price the thing in public and can’t choose the day.
That moment is now penciled for November, when Anthropic is slated to go public. It’s the first time this capital structure meets a market that votes with real money, and it meets it into a tape that is not friendly: the 10-year Treasury at a 19-year high, mortgage rates around 7.6%, diesel near six dollars, data-center buildouts slipping their timelines, and a consumer stretched thin — and not just here, this is a worldwide condition. Remember what happened two weeks ago: Oura, a profitable consumer-hardware company with real customers, pulled its IPO hours before pricing and climbed back inside the window. If the companies with the cleanest stories are flinching, the question isn’t whether the frontier’s half-trillion-dollar bet is impressive. It’s who’s standing there to buy it when the agents it was built to serve have already walked downhill.
The technology isn’t the bubble. The bet that demand stays at the frontier is. And your own agents are voting against it every time they pick the cheaper model.