Data-center bans may change the bargain without slowing AI

A local data-center ban sounds like a vote on AI itself. Arvind Narayanan's argument on Hard Fork is that the vote usually reaches a much smaller lever: it can change where capacity is built and how a community bargains with developers, while doing little to slow the world's overall AI progress.

Narayanan, a Princeton computer science professor and co-author of AI as Normal Technology, returned to Hard Fork to discuss the backlash against data centers. He separated two possible goals. A community may want to protect itself from noise, water use, electricity demand, or opaque deals. A community may also want to slow AI development because it sees the technology moving too quickly. The same moratorium serves those goals very differently. 1

For the second goal, Narayanan gave a deliberately small estimate: a one-year moratorium in a typical U.S. state would slow the industry's efficiency curve by roughly five to ten hours. The estimate describes the time the industry would need to recover the foregone compute through efficiency improvements, aggregated across the AI industry. It is an order-of-magnitude argument, not a forecast for one particular project or state. 1

Training and inference create different bottlenecks. Training frontier models depends on specialized clusters, so stopping a few local projects does little unless the restriction becomes national or global. Inference—the serving of models to users—uses a much broader pool of capacity, which makes a local ban a small subtraction from the total.

Narayanan's arithmetic rests on a second distinction: new buildings are only one source of additional capacity. Software teams learn to serve a given capability with fewer machines and less power. Newer GPUs also do more work per unit of power, and those chips can replace older hardware inside existing data centers. Narayanan's claim is that these software and hardware gains are about an order of magnitude larger than the capacity added by new physical buildings. 1

The result is counterintuitive because companies still behave as if every new cluster matters. They do. The missing distinction is between the industry's total progress and one company's position against its rivals.

Efficiency improvements spread across the industry. If every major lab can adopt a better serving method or a more efficient GPU, the improvement raises the general capacity available for AI. A single lab's physical compute, however, is scarce relative to the compute its competitors control. Narayanan told Kevin Roose and Casey Newton that new capacity can therefore have a large effect on a company's relative competitive position even when it barely changes the aggregate rate of progress. 1

That split changes what a local campaign can plausibly accomplish. A state moratorium may move a project to another state. It may impose delay or raise costs for one developer. Those effects can matter to the parties involved. The national AI curve keeps receiving efficiency gains from software and hardware, so the local action rarely becomes a brake on the whole technology.

Narayanan said data-center opposition can still be useful when it is aimed at the local bargain. Residents may want information about power contracts, water use, noise, land, tax treatment, and the decision process. Those are questions a city or state can put to a developer directly. A community can also seek compensation or public benefits without claiming that the negotiation will halt AI.

This is a more precise theory of change. The campaign is about who bears the cost of a facility and who receives its benefits. The campaign is also about whether elected officials disclose the deal before it is settled. Those objectives survive the fact that another jurisdiction may eventually host the servers.

The episode's broader policy lesson is to target the bottleneck that matches the goal. If the goal is slower AI progress, rules on consequential uses, deployment, or national infrastructure may reach the technology more directly. If the goal is local control, a permitting fight can be meaningful even when the global model-training curve barely notices it.

Earlier in the same episode, Roose and Newton discussed Meta's agreement to pay up to $17.1 billion and change how Facebook and Instagram handle teen accounts. The settlement followed a multistate case involving alleged collection of information from children under 13 without parental permission and product features that encouraged frequent use. The hosts said the agreement includes a default two-hour cumulative limit across Facebook and Instagram, hiding Instagram like counts by default, disabling some extreme makeup filters, and expanded overnight restrictions if TikTok and YouTube adopt comparable terms. 1

The comparison is useful because the enforcement target sits close to the claimed harm. The settlement changes the product teenagers use. The data-center moratorium sits several steps away from aggregate AI progress, while remaining close to local resource and governance questions. Narayanan's argument is therefore less a defense of data centers than a demand for better aim: regulate the point where the desired change can actually occur.

Listen to the original Hard Fork episode.