Burn Rate

You Don't License a Fire

Jason Lemkin runs SaaStr on three people and twenty-one agents. Two of those agents carry the title of VP, and last month the pair of them cost him $257, combined. That $257 sourced $4.8M in pipeline and closed $2.4M. Same three humans as a year ago, revenue up from minus-19% to plus-47%. Read the invoice again, because it's the whole AI economy on one line.

For twenty years you bought software the way you book a hotel room. A flat fee, by the seat, by the year, a key to the room, and the lights stay on whether you're in it or not. That deal is dead. You rent intelligence by the minute now, and the meter doesn't care whether you're a two-person shop or a trillion-dollar lab. It only runs while the tokens and the power keep pouring in. Tomasz Tunguz put a number on what that does to a business: inference becomes the single biggest category in software next year, $160B climbing to $350B, and it drags the gross margin of the typical software company down from a cushy 72% toward 30-to-50. That's not a rounding error. That's an asset-light SaaS business waking up with a cost of goods it never had.

Here's what nobody's connecting. The same meter that rewrote Lemkin's P&L is the thing the hyperscalers are feeding $700 billion into next year. A researcher this week framed the whole build-out not as a race with a finish line but as a dissipative structure — Prigogine's term for a flame that holds its shape only while energy flows through it and collapses the instant the flow stops. I spent thirty years reading the tape and that's the truest thing I've heard about this cycle. A fire isn't won. It's fed. Broadcom's Hock Tan said it plainly Monday: the binding constraint isn't chips now, it's capital. This from the shop that just agreed to lend Anthropic up to $42 billion. The grid, the turbine backlog, the $4.2 trillion revenue gap Bain is staring at — that's the fuel line, and that's the only thing that puts this out.

So both stories are true at once, and that's the issue. The operator and the lab are standing at the same meter asking the same question, just with a different count of zeros: does more come out the dollar side than I'm feeding in on the token side? If it does, you feed it, and you hire five people who can each run a swarm. If it doesn't, the fire is feeding on you. Christensen warned you the moat was coming down. He forgot to mention the water would show up as cost of goods.

At COAI today: the full Signal/Noise — the dissipative-structure math, the token-game concentration, and why "own the furnace" is suddenly every CEO's buy-vs-build call — is live at getcoai.com.

Outsider Labs Advisory →

72% → 30–50%. That's the gross margin of a typical software business before and after AI, by Tomasz Tunguz's math. Inference becomes the largest category in software next year — $160B to $350B — and takes the fat SaaS margin down with it. Translation: software just grew a cost of goods. The companies you filed under "asset-light" are now running a furnace, and the furnace bills by the minute. Renting the room by the night is over.

For the investor — underwrite the burn, not the booking. A SaaS name at 35% gross margin is a different security than one at 75%, whatever the ARR slide says. Before you pay a software multiple, find the token line inside cost of goods and ask whether it scales with revenue or with usage. The names quietly re-rating toward hardware-company margins are the ones the market hasn't repriced yet.

For the business owner — put your agents on a P&L today. SaaStr ran two AI "VPs" for $257 last month against $2.4M closed. You can't manage what you file under "seats." Pull agent spend out of the software bucket, book it as cost of goods against the revenue it actually sourced, and run the one ratio that matters: dollars out over tokens in. Above one, feed it. Below one, you're the fuel.

For the parent — teach the kid to run the meter, not the task. The work that pays in this world isn't doing the thing; it's deciding what's worth feeding the fire. Hand your kid one real project and have them manage three AI tools through it — set the budget, direct the work, judge the output. Orchestration and judgment are the muscles the meter can't flex for them.

Tuesday, October 13

Where the cycle caught up to us this week.

The leash becomes doctrine. For two weeks we told you to own your data, own the checker, keep a human on the irreversible call, and keep the logs — Sorcerer's Apprentice (us, Oct 6), The Firm (us, Oct 8), Till Incentives Do Us Part (us, Oct 12) → Satya Nadella published the same checklist as formal Microsoft doctrine in "Models as Insider Risks" (Oct 10).

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

— Harry and Anthony

Sources: