CONCEPT
Compute as Governance
Jack Clark’s insight that because frontier AI capability tracks the quantity of specialised computation used to produce it, and because computation is a physical and therefore regulable resource, compute is the best available lever for governing AI development in the present moment—the chokepoint where intention becomes enforceable constraint.
Software is famously slippery: ideas escape, algorithms can be copied, models can be downloaded. But the frontier of AI capability, at least for now, is gated by something stubbornly material: enormous concentrations of advanced semiconductors, manufactured by a handful of firms, requiring vast amounts of power and capital to operate. Jack Clark recognised early that this physicality is a governance feature rather than a limitation: physical things can be counted, tracked, and regulated in ways that pure information cannot. The logic begins with
scaling laws—the empirical relationships showing that more compute, applied to larger models and more data, reliably produces more capability—which mean that compute is not merely an input to AI progress but a proxy for it. The amount of computation used to train a frontier system is a measurable, external signal of how capable that system is likely to be, observable without needing to