The estimate

TrendForce estimates global data center power demand capacity at 161 GW in 2026, up about 31% from 122.9 GW in 2025, with AI servers an estimated 33.4% of that total.

The term is doing specific work. TrendForce measures power demand capacity in gigawatts, not consumption in terawatt hours, and splits it across general servers, AI servers, other IT equipment, and non-IT equipment such as cooling and power supply systems. So the 33.4% is AI servers' share of demand capacity, which isn't the same claim as AI using a third of data center electricity.

Looking out, TrendForce puts 2030 demand capacity at 490.7 GW against 222.6 GW of grid capacity available to data centers, and states the gap as 268 GW. That grid figure assumes all non-data-center demand is met first, which is an assumption worth carrying with the number.

When it starts to bite

TrendForce has supply and demand diverging from 2026, with the gap widening significantly from 2028. TrendForce has the divergence beginning this year, with 2028 the point where it turns sharp.

For a component buyer that timing matters more than the 2030 number. Projects that can't get a grid connection get delayed, re-sited, or built with on-site generation, and each of those changes what gets ordered and when.

What it changes for demand

This is a demand signal rather than a shortage report. Nothing here says a part is short today.

What it tells you is where the pull goes next. Power conversion and distribution content, cooling, and the passives and discretes around them scale with installed megawatts rather than with accelerator count, which is the same pull our mature-node update traced into PMICs and power discretes. That content grows whether the compute sits in a hyperscale hall or a smaller site built closer to available power.

What Buyers Should Do Now

  1. This quarter: If your customers build for data centers, ask where their projects sit in the grid connection queue. A slipped energisation date moves your delivery schedule more than any component lead time.
  2. Next 90 days: Check whether your quoted content scales with megawatts or with server count. Power and thermal content follows the first, and forecasts built on unit volumes will understate it.
  3. For 2027 and beyond: Treat on-site generation and power infrastructure as a demand category in its own right when you plan inventory, rather than as an accessory to the compute forecast.