I'd Sell the Shoes Myself

The Rent on Free

The number that leaked yesterday was $65 billion. That's Anthropic's annualized run rate, up from a billion at the end of 2024, and it's the number that floats the biggest IPO in history this fall at a reported $2 trillion. Everybody read it as a coronation. John Furlan read it closer, and he's right: the March-through-May rocket was coding plus a wave of tokenmaxxing, and the growth actually slowed from May to July. Coding saved the whole story, arriving just as the market started asking hard questions about the return on a trillion dollars of capex. One killer app. No second one yet.

Here's why $65B is the wrong number to stare at. That $2 trillion isn't priced off today's revenue. It's priced off a 2028 forecast of roughly $200 billion, at margins a software monopoly would envy. So the real bet isn't "is $65B good." It's "can this triple in 28 months." And you only triple a number that big if AI stops writing better memos and starts running whole businesses.

Now look what happened to intelligence itself while nobody was pricing this. Tomasz Tunguz swapped a Qwen model onto his laptop this week and it worked. Number one of 135 on the intelligence index, beating a cloud model 28 times its size, running for the price of electricity. Rippling ran 2,100 graded agent runs on real payroll data and the cheapest model tied the most expensive. Their president's takeaway: on real work, the models have hit a frontier.

And the switching isn't theoretical. Roelof Botha, who runs Sequoia, said this week that Grok Bot "replaced my use of Claude in less than a day." He's not a neutral party, Sequoia lives in the xAI orbit, but when a name like that publicly drops your product, the moat is thinner than the ARR makes it look. Ross Hendricks put the whole tape in five words the same day: the great AI commodification is officially here.

Sit with that. If the model on your desk is as smart as the one in the cloud and basically free, then the fat margin Anthropic charges was never a price on intelligence. It's a rent on friction. The IT you'd need to self-host, the fact that raw open models are unusable to a normal human, the chips, the distrust of Chinese weights, and the law. That's the whole product. Four of those five are melting on their own. Grok Bot is just OpenClaw and Hermes with the settings ripped out and a face bolted on. Which leaves the labs lobbying Washington to build the fifth, regulation, right as the other four dissolve. The weekend fight about Dario wanting a "FINRA for AI" wasn't safety philosophy. It was margin defense.

And here's the part that should keep an allocator up at night. Say the shoe agent actually arrives. You say "I want to sell shoes, they cost me $15 landed and $10 to ship, go," and it runs the ads, sets the budget, talks to Meta's agents, and a human touches nothing but the label printer. If Anthropic can build that, why would it rent it to you at five bucks a million tokens? Google spent 25 years knowing your intent and mostly refused to become the store. Amazon knew it and became the store. The moment the token margin thins, every model lab gets pushed from Google to Amazon. Cheap intelligence doesn't save the little guy from the monopoly. It's what causes it.

The full Signal/Noise, with the Google-versus-Amazon math and the five frictions in detail, is live at getcoai.com.

3×. That's the jump the $2 trillion needs: $65 billion of run rate today to the ~$200 billion of 2028 revenue the mark is built on. Triple a giant number in 28 months, off one killer app, at software-monopoly margins, while a laptop runs the same intelligence for the price of electricity. It's not impossible. But every "and" it needs — mass adoption outside coding, closed models winning, prices holding — is the exact thing this week's tape argued against.

Own the grader, not the model. Rippling's edge wasn't picking Opus over GLM. It was having a test set on their own data that told them the cheap model tied the expensive one. Build yours today: real tasks, pass or fail, no partial credit. The model is a rental. The checker is the asset that compounds, and it's the one thing no vendor can reprice on you.

Price your vendor as a Google or an Amazon. Write down your core model supplier and ask one question: could they plausibly enter your business? If the answer's yes, you're not just a customer, you're a scouting report. Keep a second model wired in and your data portable, so the day the toll gets thin and they eye your market, you can walk.

Run the free model this week. Put an open model on real infrastructure and point it at actual work. You're not switching today. You're measuring the friction fee, the gap between free-on-your-iron and full-price-in-the-cloud. If that gap is small for your use case, your next vendor negotiation just changed. Don't sign anything long until you know the number.

Wednesday, August 19

Where the cycle caught up to us this week.

The tape doesn't lie. We just read it early — even on ourselves.

— Harry and Anthony

Sources: