Chips and cash get the headlines, but it's the power grid — turbines, transformers, substations — that's actually slowing new AI data centers down.

This explores why getting enough electricity, rather than chips or money, is becoming what limits how fast new AI data centers can be built. To be upfront: the collection has strong material on how fast AI's energy demand is growing and on what data centers cost. It has very little on the supply side, meaning grid connections, transformers, turbines and power plant build times. The IEA's projection is the only note that names supply chain bottlenecks directly. What the corpus can do is explain why energy turns into a bottleneck at all, and the reason is not what most people expect.

Start with the scale. The IEA's central forecast has global data center electricity use nearly doubling, from 485 TWh in 2025 to 950 TWh in 2030, with AI-focused facilities tripling. Satellite tracking shows the buildout is already underway, and the IEA itself flags supply chain limits and financing conditions as the factors that could slow it down How fast will data center electricity demand actually grow?. Power infrastructure is built on timelines of years to decades. A data center can be built much faster than that, so demand that doubles in five years runs into supply that grows far more slowly.

Here is the counterintuitive part. Electricity is a small share of what an AI data center costs. Epoch AI estimates that a 1-gigawatt facility costs about $8.5B a year to own and run. Servers account for roughly $5B of that and energy for only about $0.6B What actually drives AI data center costs?. That changes what kind of bottleneck energy is. Power is cheap compared with the hardware it runs, so operators will pay nearly any price for it. The constraint is not cost but whether power is available at all, and when. Billions of dollars of servers sitting idle while they wait for a grid connection is far more expensive than paying extra for electricity. Energy becomes a bottleneck of timing and access, not of price.

Why hasn't efficiency fixed this? The IEA describes AI's electricity demand as the net result of three forces that move independently. Energy per task is falling by orders of magnitude each year. Adoption is surging. And usage is shifting toward energy-hungry tasks like video generation and reasoning How do efficiency gains and new AI capabilities balance each other?. Epoch's Denain makes the efficiency case: the cost to reach a given capability level falls roughly 5–10x a year Does inference cost remain a persistent economic burden?. But cheaper capability tends to bring more usage, not less consumption. Emberson and Sevilla estimate that token demand is growing about 10x a year while inference supply grows only about 3.4x. They point to early signs of strain, such as Anthropic reducing usage quotas during peak hours Will AI inference demand outpace supply capacity soon?.

One more pressure moves energy demand to a new place. Research shows that giving a smaller model more compute while it answers can match a larger model on hard problems Can inference compute replace scaling up model size?. As reasoning-style AI spreads, more of the energy load comes from serving users around the clock, not from one-off training runs. That makes power a continuous operating requirement rather than a periodic one. If you want the physical side, such as grid queues, transformer shortages or gas turbine backlogs, the collection doesn't cover it yet. That gap is worth filling.