Interstellar

The risk factor.

Every S-1 has the paragraph, under Risk Factors: "We derive a substantial portion of our revenue from a limited number of customers." I read those for a living once. You skim it when it's 20 percent from ten names. You stop when it's 80 from one.

Ramp put the number on the AI labs yesterday: 80 percent of OpenAI's and Anthropic's business revenue comes from 1 percent of their customers, "a level of concentration risk unseen in any other software category we track." And the 1 percent? Tech companies and AI companies. Giovanni Cattani did the taxonomy a day earlier: the frontier's demand is three jobs, AI research, writing software, and trading, roughly half of all frontier inference revenue, and every one of them reflexive. Spend tokens, make money, raise capital, spend more tokens. The frontier's best customer is the frontier.

That's a loop, not a market, and every turn costs more. We wrote in Canary in a Coal Mine that intelligence is engineered abundant, so it depreciates. Add this: each block is harder to mine. Fable 5.1 cost more to train than Fable 5, sells to the same ring of buyers, and Anthropic's first move on launch day was cutting cache prices 75 percent to get anyone else in the door. Ramp's own data had Fable 5 at 6 percent of Anthropic's tokens. The 1 percent needs the next model. Nobody else is paying for it.

A star works the same way. Each stage burns hotter and shorter until it reaches iron, the element that costs more to fuse than it gives back. Then it collapses, and what it throws off is what everything else gets built from.

Which is where you live. Bounded work: taxes, tickets, the renewal deck, the junior associate's report. Tunguz measured five years of his own writing this week. AI didn't make him faster, his line edits held at 136 a post. It deleted his bad drafts: the 10th-percentile piece improved twice as much as the 90th. That's the ROI for most of the world, doing what you did badly somewhat better, and there's a ceiling on it, which is why it goes to the cheapest model that clears the bar. The more the world adopts AI, the cheaper tokens get. Ninety-seven cents on Monday.

So the labs have two exits: find an industry that needs the frontier, or become one. Watch the second. This week OpenAI wired ChatGPT into Epic's medical records, crossed a $1 billion ad run rate, and told Cursor its models stop working November 12 because SpaceX bought Cursor. The labs' customers are finding out what "customer" means to a lab.

The bill for the star is already on your desk. Electricity up 4.2 percent, computers 5.1, Tim Cook passing the memory tab to you. Bessent says disinflation in six months. The 1 percent is burning the tokens. You're paying the light bill.

In Interstellar, an hour on Miller's planet is seven years back home. The frontier runs on that clock. Don't buy the star. Build with the iron.

At COAI today: the full Signal/Noise, with Cattani's 2×2, the Ramp spend-per-employee table, and the Lucent vendor-financing rhyme, is live at getcoai.com.

Outsider Labs Advisory →

80% / 1%. That's the share of OpenAI's and Anthropic's business revenue that comes from 1 percent of their customers, per Ramp, "unseen in any other software category we track." The 1 percent is other AI companies: labs training the next model, startups writing the code, quants trading the tape. Each model costs more to make and sells to the same ring. Everyone else does bounded work on the cheapest model that works. The frontier is a star, not a business.

Don't sign a multi-year frontier deal this month. Anthropic cut cache reads 75 percent on launch day; Google shipped its third Flash in six weeks at the same price and hasn't shipped a Pro all year; the token index hit 97 cents. The frontier's price is coming down to find you, because the 1 percent can't carry the next model alone. Today: if a Fable or GPT renewal is on your desk, push it to month-to-month and let them chase.

Point the agent at the work you do badly, not the work you do well. Tunguz's 10th-percentile posts improved twice as much as his 90th. SaaStr's renewal agent built decks for the Silver sponsors who used to get a template, and Silvers replied at a higher rate than Diamonds. Today: list the twenty accounts, reports, or customers that get the form letter. That's the queue. Cheap model, real ROI.

Run the risk factor on your own book. If 80 percent of your revenue comes from customers who are themselves funded by the AI trade, you own the loop whether you meant to or not. Today: sort your top ten accounts by who pays them. Then look at what you're allocated to. The winners here sell into bounded work and the general growth of AI, not into 1 percent of buyers.

Thursday, September 3

Ramp: 80 percent of the labs' business revenue comes from 1 percent of customers, and they're AI companies. Kharazian: "What happens in a market correction? All these companies are highly correlated." The IPO roadshow just got a risk factor. (Full analysis above.)

Google shipped its third Flash in six weeks. Zero Pros this year. Gemini 3.8 Flash: same $0.75/$3.75 price as 3.7, 73.7 percent on DeepSWE. Ars says we may never see 3.5 Pro. That's a verdict on who pays for the frontier, not a delay. (Full analysis above.)

OpenAI cut off Cursor. Then plugged into Epic. Models stop working inside Cursor November 12, after SpaceX bought it. Same week: ChatGPT inside Epic's hospital records and a $1 billion ad run rate in under 200 days. Palmer: "Now AI is eating the software." Including the software that was its customer. (Full analysis above.)

Mythos found zero bugs in curl. A startup's harness found six. Daniel Stenberg posted the frontier models' empty list August 24. AISLE filed 29 reports; six became CVEs in curl 8.22.0. Kroah-Hartman: "I'm seeing the same for Linux." The harness beat the model.

Three sites that didn't exist two years ago wrote 215,128 "best software" pages. Perplexity cites them ahead of Gartner. Their homepages are titled "Facts & Grounding Page," addressed to the machine, not you. A separate audit: a third of Perplexity's numeric citations don't contain the number. Check what the model says about your company. Someone minted it.

Where the cycle caught up to us this week.

Sep 2 → Aug 25. Frontier-only "was not a durable era" and Fable 5 stuck at 11 percent of spend (us, Aug 25) → AI Secret, "A Leader's Curse": "the best model in the world, priced 6× the alternative, and the whole launch was spent defending a premium the market already walked away from."

Sep 1 → Aug 26. Split the token bill: commodity work to the cheap model, the frontier meter only for the judgment-dense 20 percent (us, Aug 26) → Cattani: "demand for AI is ultimately demand for labor"; bounded tasks "converge on the cheapest possible option… non-frontier models will dominate."

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

— Harry and Anthony

Sources: