Compute Governance & Export Controls

Of all the inputs to frontier AI, compute has unique governance properties: it's countable (chips are physical objects), excludable (you can't train a frontier model on smuggled enthusiasm), and concentrated (the supply chain bottlenecks through TSMC's fabs, ASML's EUV machines, and NVIDIA's datacenter GPUs).

Compare to the alternatives: algorithms diffuse as papers and talent; training data is everywhere; weights fit on a drive. Only compute sits behind a chokepoint that export controls, reporting regimes, and procurement rules can actually grip. Compute-based governance is an imperfect proxy for capability — algorithmic efficiency keeps shrinking what a FLOP buys — but it's the only input governments can currently meter.