America’s AI data-centre boom may hit a power wall before it hits its targets

TrendForce projects that US data centres could face a gap of more than 170GW between demand and grid-deliverable power by 2030, with serious bottlenecks emerging from around 2028. The useful point is less another giant number than the constraint it exposes: AI expansion may be limited by electricity infrastructure before it is limited by appetite for servers.

Watch Desk analysis

What happened

A TrendForce analysis published on 22 September says global data-centre power demand capacity could reach 211.1GW by 2027, driven largely by the rapid growth of AI servers. In the United States, the analysis says grid capacity is not keeping pace with the demand being created by planned data-centre infrastructure.

TrendForce projects that the US gap between supply-chain demand and power the grid can actually deliver will exceed 170GW by 2030. It expects bottlenecks to start appearing around 2028, turning electricity access into a likely timetable risk for new AI facilities.

Why it matters

The immediate consequence is that having enough chips, buildings and money may not be enough to bring more AI capacity online. Operators may face longer connection queues, tougher competition for generation and greater pressure to locate facilities where electricity can be secured rather than where customers or existing infrastructure make expansion convenient.

The projection also shifts the argument around AI infrastructure. The limiting factor may not be whether companies want more servers, but whether power networks can deliver what those servers need without forcing households, industry or other grid users to wait their turn. The grid, as ever, has declined to be impressed by a pitch deck.

Our read

This is a serious infrastructure warning, not proof that a 170GW shortfall will arrive exactly as forecast. TrendForce’s figures are projections, and the analysis does not establish which projects will be cancelled, delayed, relocated or supplied through new generation.

But the direction is hard to dismiss. AI demand is becoming a physical planning problem involving transmission, generation, cooling and connection queues. Companies building models should be judged partly on whether they can secure dependable power, not only on benchmark scores and fundraising totals.

What to watch

• Connection queues:
  whether US utilities publish longer waits or revised limits for large data centres.
• New generation:
  whether operators fund dedicated, low-carbon power rather than relying on existing grid capacity.
• Project delays:
  which announced AI facilities move their opening dates or change location.
• Demand flexibility:
  whether data centres can shift workloads or reduce consumption during periods of grid stress.

Discussion spark: Should AI data-centre developers pay the full cost of the new power and grid capacity they require, or would that slow infrastructure investment that economies increasingly need?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.