Concentration

The Backorder Is the Moat

I went into the city last night. I run out of New Canaan, Anthony's in New York for the month, and we did the thing you can only do live: dinner, then the Rangers, then three hours arguing about the one company everyone's about to be able to buy. By the end I'd put a hundred bucks on the table. Anthropic goes public, buy it, hold a year. Anthony took the over, I took the under. The whole bet turns on one word, and it's tonight's headline.

Here's what we both agree on. Meta cut its Claude Code users roughly in half, from about 60,000 to 30,000. Microsoft took a knife to a projected billion-dollar internal Claude bill, a third of it gone. The Pentagon dropped Claude outright. Three of the most sophisticated, best-capitalized customers on earth, same week, all walking for the door. And the leaked S-1 says two customers were about a quarter of Anthropic's 2025 revenue. The people who pay the freight are the people who can leave.

The bull answer is Anthony's, and it's a good one: they can't actually leave. You want to run your own models? Go buy the racks. You can't. Buy the memory. Sold out. Fast interconnects, a data center, the power to light it up? Backordered, backordered, spoken for. The switching cost was never the software. It's the supply chain.

Which is the whole problem. That's not a moat, it's a backorder. Ben Evans said it flat this week: no sign of winner-take-all, and no reason the frontier labs have pricing power beyond a shortage of compute. Gemini 4 just walked back to the top of the boards, OpenAI's running near a $70B clip, and Anthropic's early coding lead is getting sanded down. John Armitage, thirty years at Egerton, has one test for a great business: can you raise price despite competition? Today's tape says no. These companies have a sold-out quarter wearing a franchise's costume.

And the same shortage that holds the customer is the thing that eats the margin. To grow into a two-trillion-dollar price you need something like $400 billion in revenue and $100 billion in free cash flow, six times where the run-rate sits now. Every dollar of that growth gets bought with compute they RENT, at shortage prices, against something like half a trillion in commitments they can't unwind. The scarcity props the valuation up and caves the margin in at the same moment.

So I don't need the backorder to clear inside my year. Markets look forward. They just have to recognize the thing for what it is, a supply constraint and not a model moat, and the multiple re-rates on sight. I love the product. I love the vision. I'm short the multiple.

At COAI today: the full Signal/Noise, with the S-1 math, the Armitage test, and why "bullish on AI, bearish on the comp" isn't a contradiction, is live at getcoai.com.

Outsider Labs Advisory →

Two customers were about a quarter of Anthropic's 2025 revenue. The company will tell you it's now 6,000 accounts over $100k, and that's true, the concentration is receding. But the whales who pay the freight are the same whales with the cash, the engineers, and the will to run it themselves the day the compute frees up. A $2 trillion debut rests on them staying put. That's not a moat. That's a backorder with a bow on it.

For the investor — price the moat, not the model. Before you touch the Anthropic or OpenAI IPO, ask Armitage's question: can they raise price while Gemini 4 sits at parity and their biggest customers insource? If the "moat" is really a compute shortage, you're underwriting a backorder at a franchise price. Keep your AI bull case and your AI-valuation bet in separate buckets. Today they point opposite ways.

For the business owner — find out if your lock-in is real. The reason Meta can't fully drop Claude is physical, not loyalty: it can't buy the hardware. Run that same audit on your own vendors. What are you tied to because it's genuinely better, versus what's just sold out everywhere else? The second kind reprices the day supply loosens, and that bill lands on you. Budget for it now.

For the parent — teach the kid to spot the real constraint. The story isn't "AI is magic." It's that the whole thing runs today on a shortage of chips, memory, and power, physical stuff that gets dug and built. Help a kid tell the difference between a thing that's valuable because it's great and a thing that's valuable because it's scarce. That muscle outlasts every hype cycle they'll live through.

Wednesday, October 7

Where the cycle caught up to us this week.

The 1% pays the bills. "80% of OpenAI's and Anthropic's enterprise revenue comes from 1% of customers" (us, Interstellar, Sep 3) → a16z's new consumer data lands on the same spine, the top 1% of payers driving 19.5% of consumer AI revenue, and aisecret ran it under the exact headline "The 1% Pays AI's Bills" (Oct 6). We called it in the enterprise. The consumer tape just printed it.

The tape doesn't lie. We just read it early.

— Harry and Anthony

Sources: