ENVIZNThe Power Problem Behind AI
Compute capacity can be ordered faster than grid capacity can be built. That difference in timelines is becoming one of the defining constraints of the AI buildout.
01The mismatch in one comparison
A hyperscale data-center building may move from groundbreaking to completion in roughly two years. Major transmission can take five to ten. Much of the strategy surrounding AI infrastructure follows from the gap between those clocks.
- Buildout timelines: silicon (months) vs. buildings (years) vs. grid (decade)
- Why the gap widened: load growth returned after twenty flat years
02From chips to megawatts
The unit math is unforgiving. Start with an accelerator drawing on the order of a kilowatt, multiply by tens of thousands, then add cooling and electrical overhead. A large training campus can reach hundreds of megawatts before future expansion is counted.
- A worked example: cluster size → IT load → facility load at realistic PUE
- Why each model generation raises the floor rather than the ceiling
- Inference: smaller per site, but everywhere: a different grid problem
03Power-first site selection
Fiber, land, latency, taxes, and customers still influence data-center geography. For the largest AI campuses, however, deliverable electrical capacity can eliminate a site before those advantages matter. The map is increasingly drawn around power.
- The migration toward ERCOT, the Midwest, and the Southeast
- "Powered land" premiums and the speculative market around substations
- Why some announced campuses may miss their energization schedules
04The workarounds
When the grid cannot deliver on the required schedule, buyers look for another path. Behind-the-meter generation, nuclear contracts, batteries, and flexible operation can compress one constraint only by accepting another combination of cost, emissions, complexity, or time.
- Behind-the-meter gas turbines: fast, controversial, increasingly common
- Nuclear PPAs and restarts: long-dated bets on firm power
- Batteries and demand flexibility: shaving the peaks that drive upgrades
05The efficiency counterargument
Efficiency lowers the cost of useful computation, but lower cost can stimulate more demand. The central uncertainty is therefore not whether AI hardware becomes more efficient. It is whether efficiency improves faster than the appetite for intelligence expands.
- Jevons paradox applied to compute: cheaper intelligence → more of it
- What would have to be true for AI power demand to flatten
- Why we treat demand forecasts as scenarios, not predictions