AI Data Centre Approvals Are the New Bottleneck

Every AI data centre conversation still starts with GPUs and ends with megawatts. That is the old map. The new bottleneck is not silicon, and it is not the grid. It is the permit.

The IEA now puts global data-centre electricity demand on a path to more than 945 TWh by 2030, with AI the main driver. That is roughly Japan's entire power system. In the United States, data centres already take about 4 percent of electricity and are heading toward 7 to 12 percent by 2028. McKinsey puts the American build at $6.7 trillion of data-centre, generation and transmission capex this decade. None of that spends if the site cannot get approved.

Permits now take longer than construction

Construction is the easy part. A large campus can be built in two to four years. Grid interconnection and environmental approval routinely take five to ten. In Virginia, the densest data-centre market on earth, a 2024 law now requires a full environmental impact assessment before any rezoning for a data centre. Loudoun County, the so-called Data Centre Alley, has paused new applications while it rewrites the rules. Prince William County has already rejected a $19 billion campus. Santa Clara, the other US AI hub, is so constrained that new large loads are being told to wait until 2027 or 2028.

That is not a local nuisance. It is a capacity cap. Every rejected rezoning is megawatts that never reach a rack.

Water, noise, and the neighbours

The fight is not abstract. A hyperscale site can pull millions of litres a day for cooling, and closed-loop systems only move the problem; they do not delete it. Communities hear the fans, see the substations, and watch industrial zoning march toward houses. In the UK, Ireland and the Netherlands, water and grid stress have already produced moratoria. Ireland's grid operator has effectively frozen new data-centre connections in Dublin. The Netherlands blocked new hyperscale sites in 2024. These are not activist footnotes. They are state decisions.

Australia is next in line. The National Electricity Market is already tight. A campus that needs 200 to 500 MW does not drop into an existing suburb without a fight over water licences, transmission easements and local amenity. The developers who treat community and environmental approval as a paperwork afterthought will lose years, not weeks.

Power was last year's story

Power still matters. I wrote last month that procurement is the real bottleneck once the chips are ordered. That remains true on the grid side. What has changed is the queue in front of the grid. You cannot sign a PPA for a site that does not exist. You cannot interconnect a campus that has no rezoning. The 401(k) capital I covered last week will fund the build, but it cannot buy a permit: https://brawlersguide.com/post/401k-flood-capital-raising-private-markets.

The operators who win the next five years will be the ones who reverse the sequence. Land, water, community, then power, then steel. Not the other way around. The GPU allocation follows the site that already has a yes.

The new diligence question

If you are allocating to this trade, stop asking whose chips are better. Ask whose file is already in front of the council. Ask whether the water licence exists. Ask whether the transmission easement is real. Ask whether the neighbours have already been heard. The $6.7 trillion does not care about your architecture diagram. It cares whether the site can be switched on.

Approvals are now the critical path. Everything else is inventory.