# Knife-Catching Compute

"Knife-catching" is Edward Zitron's term for the bind that AI labs face when planning compute purchases. Compute capacity has long lead times (data centers take 1-3 years to come online); demand for AI is volatile and partly conjured by the labs themselves. The lab has to commit to capacity well before it knows what its revenue will be when that capacity comes online. Both directions of error are punishing.

## The two sides of the bind

**Over-commit**: order enough compute to cover a hoped-for trajectory and lock in the rate. The lab guarantees the capacity it'll need *if* growth holds, at the best per-unit price. But if growth slows — or merely stays at the level the lab already has — the lab is left with billions in obligations it can't service. Amodei put it himself, quoted in [[ai-too-expensive]]:

> Basically I'm saying, "In 2027, how much compute do I get?" I could assume that the revenue will continue growing 10× a year [...] I could buy $1 trillion of compute that starts at the end of 2027. If my revenue is not $1 trillion dollars, if it's even $800 billion, there's no force on earth, there's no hedge on earth that could stop me from going bankrupt if I buy that much compute.

**Under-commit**: be more conservative, then buy last-minute when demand actually arrives. Spot prices spike when everyone needs capacity at the same time, destroying whatever gross margin the lab might have captured. The Information reports that Anthropic's 2025 gross margins were 40% — 10 points below projection — *specifically* because the company had to buy expensive last-minute compute when chatbot demand exceeded forecasts. OpenAI's margins fell from 40% (2024) to 33% (2025) for the same reason.

## Why the bind is structural

The bind isn't a bad-management story. It's structural for three reasons:

- **The capacity decision precedes the revenue.** Data centers are physical things. A 1.9GW campus takes years; a 5GW Anthropic commitment to Amazon is a multi-year build. The lab is deciding 2-3 years in advance.
- **The lab can't measure demand.** Enterprise customers can't measure their own token consumption per task (see [[ai-token-budget-explosion]]). The labs are getting demand signals from customers who don't know what they're spending.
- **There's no hedge.** Compute capacity isn't a fungible commodity with a derivatives market. You can't short GPU capacity. There's no insurance against the over-commit case. Amodei's "no force on earth" line is literal.

## What the bind explains

Several otherwise-puzzling 2026 facts make sense as knife-catching dynamics:

- **The $200B + $100B + $30B Anthropic obligations.** Anthropic appears to have chosen aggressive forward-commitment over last-minute exposure. The trade is: bet on growth or guarantee a bankrupt 2028.
- **The relentless fundraising.** Anthropic raised $75B in six months (assuming the $30B May round closes) while *also* reporting revenue growth. The growth doesn't reduce the need to raise; the forward commitments grow faster than revenue.
- **The compute-only "demand."** 70% of Microsoft / Google / Amazon's AI compute goes to Anthropic and OpenAI. The labs are the demand. Hyperscaler capex is, at one remove, a bet that these two labs survive the knife-catching period.

## What this predicts

The bind is asymmetric: the over-commit case is *immediately* fatal once revenue diverges; the under-commit case is slowly fatal through margin compression. So the labs face a structural pressure to over-commit, which then needs ever-larger fundraising rounds to defer the day of reckoning. This is one of the [[ai-bubble-pale-horses|pale-horse signals]] — when the fundraising required to defer the over-commit hit exceeds even the inflated 2026 cadence, the door closes.

## Related concepts

- [[ai-too-expensive]] — the source piece, with the full quote and numbers
- [[subprime-ai-crisis]] — the broader thesis the bind sits inside
- [[ai-subsidy-economics]] — the upstream / downstream mechanism
- [[ai-bubble-pale-horses]] — the signals checklist
- [[ai-token-budget-explosion]] — the demand-side measurability problem that feeds the over-commit pressure
- [[hold-on-to-your-hardware]] — the consumer-hardware consequence of the same dynamic