On 18 August, the governor of Pennsylvania signed an executive order making local community approval mandatory for the siting of AI data centres. After a project in India was blocked over water concerns, opposition to these facilities is now reaching US states, this time in regulatory form.
For a month now, we have been tracking the physical constraints of the computing race. Here is the political constraint, and it is taking formal shape.
What this executive order changes
Until now, siting a data centre was largely a matter of negotiation between an operator, a landowner and an electricity supplier. Local authorities had planning tools at their disposal, but no mechanism gave them a specific right of oversight over this type of installation.
Mandatory local approval shifts the balance of power. It turns a technical and economic decision into a local political decision, with all that implies in terms of public debate, electoral timetables and negotiation.
Why these projects are drawing opposition
Three reasons come up time and again, and we have documented them.
Electricity. A large data centre consumes the equivalent of a town. In a region where the grid is constrained, this new demand can weigh on tariffs and on availability for other users.
Water. This is the reason that blocked the Indian project, and we explained it in our article on the environmental cost: the most efficient cooling systems rely on evaporation, and that water does not come back.
The benefit-to-nuisance ratio. This is probably the most decisive argument locally. A data centre represents a considerable investment and creates relatively few permanent jobs relative to its cost. The nuisances are concentrated on nearby residents, while the benefits are spread elsewhere.
These projects bring substantial tax revenue to communities that lack it, construction jobs, and sometimes funding for electrical infrastructure that the region benefits from afterwards. Some municipalities actively seek them out. And there is a sovereignty argument that we developed in relation to computing sovereignty: a country that hosts no capacity depends entirely on that of others. Refusing everywhere would also come at a cost.
What this signals
The trend is consistent with what we have described all week. Demand for infrastructure is exploding, driven by considerable capital and long-term commitments. It is colliding with finite resources and with the people who live there.
Two likely consequences. First, a geographical shift towards the territories least equipped to say no, which raises a question of fairness. We saw the same mechanism in annotation work, where conditions move rather than improve.
Second, a rise in costs. Social acceptability becomes a line item: studies, compensation, negotiations, delays. This mechanically reinforces concentration in the sector, since only the best-funded players can absorb these expenses.
What to take away
We have spent months describing the limits of AI in terms of chips, electricity and memory. Here is the limit that gets the least attention and could weigh just as heavily: the people who live where you want to build, and who now have a way to make themselves heard.
This constraint is hard to get around, because it cannot be resolved by engineering or capital. It is resolved through negotiation, which takes time, and time is precisely what this sector lacks most.