"Capping Compute: A Concrete Proposal to Slow the Race to Superintelligence"

The conversation about artificial superintelligence usually lives in abstractions — alignment, scaling laws, existential risk — which makes it easy to nod along and hard to know what to do. Albert Wenger's follow-up to the Statement on Superintelligence is valuable precisely because it refuses to stay abstract. It proposes a specific, countable lever: a negotiated limit on how many additional gigawatts of datacenter capacity can be built, agreed between the world's two largest AI powers. Whether or not the diplomacy ever happens, the proposal is worth engaging with on its technical merits, because it rests on a claim that is both simple and — if true — quietly profound.

That claim is that the next stage of AI progress, recursive self-improvement, is gated less by cleverness than by raw capacity. The argument runs roughly like this: whatever "the next step in the race" turns out to be, it will not come cheap in computation. If that's right, then even a coarse cap on physical datacenter buildout buys meaningful time, because you cannot recursively improve a system you cannot power. The elegance here is that the control point is not a model, an algorithm, or a technique — it is steel, concrete, and transmission lines.

There's a deeper reason the "gigawatt cap" is a genuinely interesting unit, and it goes beyond the original post. Compute, unlike many candidate objects of AI regulation, has a physical footprint that is remarkably hard to hide. Model weights are bytes — trivially copyable and almost impossible to track. Training runs are ephemeral. But a datacenter is a building: it needs land, grid connections, transformers, and cooling water. Satellite imagery can see it, electricity markets can feel it, and permitting processes can count it. That makes compute one of the few AI-relevant quantities that is remotely verifiable, which is the first requirement of any workable arms-control arrangement.

The obvious objection — raised in the piece itself — is verification: two rivals that don't trust each other can't easily police an agreement. The rebuttal is the nuclear precedent. The United States and the Soviet Union trusted each other even less, yet built a verification regime around the most dangerous technology of the twentieth century that has held for more than fifty years, grounded in the Nuclear Non-Proliferation Treaty's three pillars. It's not a perfect analogy — nuclear material is dual-use in a much narrower sense than electricity — but the history matters because it proves that verification between adversaries is possible when the stakes are high enough.

Here's an insight the analogy actually strengthens rather than weakens: in some ways compute may be easier to verify than fissile material. Enrichment happens in a few secret rooms; gigawatt-scale power draw shows up on grid operators' dashboards and in heat signatures. The catch is that compute is far more broadly legitimate — every hospital, factory, and streaming service runs on it — so the monitoring burden is enormous and the false-positive problem is real. A workable regime would need to distinguish "lots of ordinary compute" from "the specific class of frontier-scale clusters" without strangling the economy. That's a hard, but not obviously impossible, engineering problem.

A second angle the post gestures at but doesn't fully develop is that coarse limits are the norm in arms control for a reason. Treaties historically capped things you could count from a distance — warheads, launchers, ships' tonnage — rather than trying to bound "military effectiveness" in the abstract. A gigawatt cap is squarely in that tradition: you're not banning the science of intelligence, you're bounding the substrate it runs on. That framing is likely to age well, because it sidesteps the intractable question of policing a model's internal capabilities and focuses on the one thing you can observe.

The design details would matter enormously, and they're where the idea gets genuinely difficult. A flat cap without baselines would freeze in place whatever compute advantage already exists, effectively locking incumbents' lead into the treaty text — a nonstarter for the party that's behind. Then there's the covert-build problem: if the cap is on new datacenter capacity, the incentive to over-provision quietly, to distribute compute across many under-the-radar sites, or to squeeze more out of existing silicon all rise. Any real agreement would need to negotiate those dynamics, not just the headline number.

There's also a subtle timing argument embedded in the proposal. Capacity is being added fast, and the window in which a slowdown is even meaningful — before recursive self-improvement is underway — may not stay open indefinitely. In that sense the post is less a finished treaty and more an argument that the physical infrastructure race is the place to intervene while the problem is still tractable. It reframes AI safety from a purely technical field into a question of energy policy, grid planning, and infrastructure governance.

That reframing has a useful side effect worth naming: it connects the AI debate to a fight that is already happening. Communities across many countries are pushing back on datacenter construction over power, water, and land use, and the post notes this opposition is one of the rare issues that cuts across partisan lines. A cap on gigawatts would, in effect, give the "not in my backyard" energy a constructive outlet — turning scattered local resistance into a structured national limit. That's a genuinely different kind of coalition than AI safety typically mobilizes.

None of this makes the proposal an easy sell. Critics will fairly point out that a slowdown risks ceding an edge to whichever party cheats first, and that verification regimes for nuclear weapons took decades to build under conditions we don't yet have for AI. The honest reading is that the plan is best understood not as a finished document but as a forcing function — an attempt to move the discussion from "we should be careful" toward "here is a measurable thing we could actually cap." Even skeptics of the diplomacy can learn something from the direction of the thinking.

What the proposal ultimately gets right is a point that's easy to overlook in a field obsessed with models: superintelligence, if and when it arrives, will be as much an energy and infrastructure phenomenon as a software one. Every flop of progress ultimately flows through a physical facility that must be built, permitted, connected, and cooled. If that's true, then the levers that matter for slowing the race aren't only in the lab — they're in the zoning board, the grid operator, and the treaty negotiation. Intelligence may be invisible, but the infrastructure that produces it is not, and that visibility is the opening this proposal is betting on.

Sources

• Albert Wenger — Slowing Down the Race to Superintelligence · Treaty on the Non-Proliferation of Nuclear Weapons (Wikipedia) · Superintelligence (Wikipedia)