On DeepSeek and Export Controls
Map
Notes
Library
Why efficiency gains make export controls more urgent, not less.
DeepSeek is impressive engineering.
It is not a unique breakthrough.
A reading of Dario Amodei
on DeepSeek and export controls
Scaling laws. Cost curves. Geopolitical stakes.
Each one required to hold the argument.
— why cheaper per unit means bigger bets
Efficiency gains don't shrink total chip demand. They accelerate traversal of the scaling curve. Cheaper inference costs mean frontier labs can afford larger training runs. Cheaper training means more experiments. More experiments means faster progress toward AGI-class systems. The curve doesn't flatten — it compresses in time.
Every cost reduction in AI infrastructure has historically expanded total spend, not contracted it. GPUs became cheaper; labs built larger clusters. Inference optimisation made deployment cheaper; companies deployed at scale. The pattern holds. DeepSeek's efficiency is real — and it means Chinese labs can now reach frontier scale with fewer chips. Which is precisely why the window for export controls is closing, not opening. The question is not whether China will build capable systems. It is whether they will accumulate the chip density needed by 2026–2027, when AGI-class systems become feasible. Export controls are the only lever that can force a unipolar outcome.
Export controls are the single variable that determines which one we build toward.
Efficiency gains don't shrink total chip demand.
They accelerate the scaling curve.
The DeepSeek moment has produced two opposing reactions. One camp reads it as proof that export controls don't work — that Chinese labs will always find a way. The other reads it as proof that controls must tighten immediately. Amodei's argument cuts through both. DeepSeek-V3 is genuine engineering. Its cost and performance are exactly what historical efficiency trends predict. No miracle. No anomaly. Just the curve continuing on schedule.
The real question is not whether China can build capable systems — they can, and they will. The question is whether they can accumulate the density of chips needed by 2026–2027, when AGI-class systems become feasible. That is a different problem. Efficiency gains accelerate the scaling curve; they don't flatten it. Cheaper per unit means bigger bets. Bigger bets mean faster traversal toward frontier. Which means the window for export controls is closing, not opening. The controls are working — they are forcing China to be efficient. But efficiency is not the same as abundance.
The geopolitical stakes are binary: unipolar (US and allies hold durable lead over authoritarian states) or bipolar (US-China parity in frontier AI). Export controls are the only lever that can force unipolar. Everything else — talent, open-source models, algorithmic breakthroughs — is available to both sides. Chips are not. DeepSeek's chip mix (partly pre-ban, partly smuggled, partly still-allowed H20s) shows controls are working and adapting, not failing. But the margin is tightening. Read this essay before you decide whether controls are worth the cost.
We read every message. No algorithm, no auto-reply — just Sophie and the LOOMUS team, with coffee, sometime this week.
Your note is in the inbox.
We'll write back when the kettle's on.
Saved with your account — part of a study that grows every time you distill. Nothing here expires, and nothing is lost.
Every LOOMUS reader may claim a planet — once, for life. It’s a slow choice, made when you’re ready.
Enter The Choosing →or simply look around your space →
No hurry; the sky holds your seat.