Compute-cap and training-run threshold regulation

Advocate a numerical regulatory threshold on the scale of frontier-AI training runs — a compute-quantity cap (FLOPs), a parameter-count ceiling, a hardware-throughput limit, or a licensing regime that binds runs above a stated threshold to prior regulatory approval. The threshold is set to sit above today's already-deployed frontier and below the next generation, converting the industry's own scaling roadmap into a regulatory decision-point at which government (or an international body) can intervene before the next class of model exists. The vehicle is legislation (California SB 1047's compute threshold; the EU AI Act's Article 51 systemic-risk designation at 10^25 FLOPs), plans and proposals that specify the threshold explicitly (ControlAI's A Narrow Path; MIRI-adjacent policy work; academic AI-governance research), and coalition advocacy around the number itself as the load-bearing regulatory unit.

An actor chooses this strategy because compute is the AI industry's single most measurable and choke-pointable input — the frontier of AI capability tracks compute more reliably than any other variable, and the compute supply chain concentrates through a handful of chip vendors and hyperscale cloud providers whose customer records are legible to a regulator. A threshold on compute is the closest a regulator can get to a dial on the frontier's rate of advance without having to define what "AI" is in the statute or predict which models will produce which harms. The strategy also stakes out a middle position between two failing poles: a full moratorium (politically unavailable in most jurisdictions; contested even inside the movement) and per-deployment case-by-case regulation (structurally too slow to keep pace with the frontier). The compute-threshold framing is what makes conditional, quantitative regulation of the frontier legible to legislators who need a specific number to legislate against.

It trades definitional simplicity for arms-race friction. A compute threshold is only durable if it is periodically raised or lowered on transparent criteria, and the political economy of setting those criteria is a permanent open question — set too low, the threshold captures every commercial deployment and collapses; set too high, it applies to nothing. The strategy is vulnerable to industry innovation that reduces the compute cost of frontier capability (a training-efficiency improvement that puts a frontier-class model below any political threshold the movement can defend), to jurisdictional arbitrage (SB 1047's veto pushed frontier training decisions elsewhere), and to strategic obfuscation (a vendor slicing its training runs to sit below a threshold, or moving compute to a jurisdiction without one). And the strategy's success invites its own erosion: a threshold that binds today's frontier will not bind tomorrow's without maintenance the movement must sustain politically for the life of the regime.

Ecology

Distinct from mass protest for AI moratorium by shape: a moratorium is a full pause on frontier development; a compute cap is a numerical dial applied selectively above a stated threshold. The two strategies address the same underlying concern (rate-of-advance of frontier capability) but on different regulatory registers — the moratorium via public mobilisation demanding a categorical halt, the compute cap via inside-game regulatory design producing a conditional, statutable, adjustable instrument. Distinct from advocate for a new AI regulatory body by object: an agency is the institutional vehicle a compute-cap regime could sit inside, but this strategy is the substantive regulatory demand — the number and its enforcement mechanism — not the institutional container. Distinct from mandatory algorithmic impact assessment by object: impact assessment regulates deployment of a specific system; a compute cap regulates development at scale, upstream of any specific deployment.