Thursday Jul 9
HARDWAREOther

Grid Power, Not GPUs, Caps AI Now

9JUL
GPUS READYGRID: NO

The bottleneck moved. Chips are sitting ready while the wiring to run them isn't. Analysts see a third of planned buildouts stalling on wait times that now beat any GPU order.

Across 84 tracked US data-center projects, planned AI capacity hits nearly 43 gigawatts. The number that decides how fast that comes online is no longer GPU supply.

Gartner expects power shortages to restrict 40% of AI data centers by 2027. Sightline Climate counted 12 gigawatts announced for 2026, but only 5 are under construction. Grid queues in Virginia, Phoenix, and Dallas now run four to seven years.

High-voltage transformer lead times stretched from under 2 years to as long as 5. Some operators are now building their own power plants just to skip the wait.

full brief & sources

⚡ Why this matters

  • Every AI roadmap assumes compute keeps scaling. Power is now the thing that says no.
  • A 4-7 year grid queue is longer than most companies' entire AI product roadmap.
  • "Bring your own power" is quietly becoming a real strategy, not a fringe idea.

🔍 What happened

  • 84 tracked US AI data-center facilities now plan for nearly 43 gigawatts of combined capacity.
  • Gartner forecasts power shortages will restrict 40% of AI data centers by 2027.
  • Sightline Climate tracked 12 gigawatts of announced 2026 US capacity; only 5 gigawatts are under construction.
  • Grid interconnection queues in Northern Virginia, Phoenix, and Dallas now run 4 to 7 years.
  • High-voltage transformer lead times stretched from roughly 2 years before 2020 to as long as 5 years now.
  • HSBC flags "bring your own power", meaning on-site generation, as a growing workaround for the grid bottleneck.

💬 Smart takes

  • Gartner: power shortages will restrict 40% of AI data centers by 2027, reframing GPU scarcity as the smaller problem.
  • Skeptic: announced gigawatts are cheap to promise and easy to cancel; the 12-vs-5 gigawatt gap from Sightline Climate shows how much of this capacity may never get built.

🧭 Where this goes

  1. Likelymore hyperscalers sign direct power-generation deals in gas, nuclear, or on-site instead of waiting on grid queues.
  2. Likelydata-center site selection starts following power availability more than fiber or tax breaks.
  3. Possibletransformer and switchgear manufacturers become as loud a bottleneck story as Nvidia GPUs were in 2023-24.
  4. Wild Carda major AI lab publicly delays a model or product launch and cites power, not compute, as the cause.

🥄 The Spoon Take

For two years the AI story was "buy more GPUs." The story now is "good luck getting the power to run them." The bottleneck moved from a chip Nvidia can ship in months to a transformer and grid connection that take years. Roadmaps built on compute scaling alone are betting against physics and utility regulators at the same time.

🤔 Pushback

Every power-crunch story assumes current AI demand projections hold. If model efficiency keeps improving as fast as the last two years, some of this projected capacity may simply not be needed.

·Help Net Security - Power slows AI buildout·Silicon Report - Power is the bottleneck·PV Magazine USA - Transformer lead times stretch to 4 years