This usually gets framed as an AI story, and understandably so. But here’s the part that gets missed: the grid constraints and building moratoriums hitting hyperscalers apply to all capacity — while AI is quietly eating the entire available envelope.
Every megawatt an operator can bring online increasingly goes toward GPU clusters and training runs. That’s capacity that historically absorbed the boring, essential stuff — your ERP, your databases, your backups, your general-purpose compute. When new power connections are frozen and AI has first claim on what’s left, ordinary enterprise workloads start competing for scraps.
And the supply is genuinely being capped — from two directions.
Top-down: Lawmakers in 15 U.S. states are weighing moratoriums or bans on new data centre development.
New York passed a one-year moratorium on facilities over 20 MW through the legislature.
Virginia — the world’s largest data centre market — has a bill to block new builds until every interconnection request is fulfilled (or July 2028).
Georgia, Vermont, and Pennsylvania are eyeing multi-year pauses tied to impact studies.
Bottom-up: Erin Brockovich — yes, that Erin Brockovich — has turned her attention to data centres. After asking communities to report concerns, she received nearly 4,000 submissions in a single month and launched a public map of contested projects. Her framing isn’t anti-AI; it’s anti-secrecy — projects announced only after permits are secured, officials under NDA before neighbours even know. That grassroots energy already has teeth: by one estimate, ~75 projects worth roughly $130 billion have been blocked or delayed this year.
And the same pattern is playing out worldwide:
Ireland froze new grid connections in Dublin; operators must now bring their own generation.
The Netherlands confines hyperscale builds to two designated national zones.
Singapore rations capacity behind the region’s strictest green-energy standards.
Germany mandates waste-heat reuse; Portugal approves only centres that deliver measurable economic value.
Put it together — a hard ceiling on new supply from both statehouses and townhalls, and AI consuming most of what already exists — and the implication is bigger than slower model training. It’s that general cloud capacity in constrained regions gets tighter, pricier, and harder to reserve. The next capacity crunch may not announce itself as an AI story at all. It’ll show up as a quota you can’t raise, a region that won’t take your expansion, or a reserved-instance price that keeps climbing.
The takeaway for anyone building on cloud: power and location are now first-class architecture decisions — for every workload, not just the GPU-hungry ones. Region selection, on-site generation, PUE targets, and heat reuse used to be a hyperscaler’s problem. They’re becoming yours.
The winners won’t be whoever builds the most capacity. It’ll be whoever builds the most efficient, grid-aligned capacity — and whoever plans their footprint like power is scarce, because it is.
Are you seeing AI demand crowd out your general-purpose capacity planning yet?
#CloudArchitecture #DataCenters #AIInfrastructure #EnergyPolicy #Hyperscale