The AI Bottleneck Moved From Chips to Power

Industry · Working with AI

For 2 years silicon was the most in-demand resource. If you wanted to train or serve a big model, you queued for H100s, you paid a premium, and hardware access set the ceiling on what you could build. This has now shifted over to electricity, and where you can get it.

In the same week Gartner reported that global data-center electricity use would hit 565 terawatt-hours in 2026, a 26% jump in a year, around 75 data-center projects worth roughly 130 billion dollars were postponed or canceled across the United States. None of which was due to money or chips, but for the power needed to run them.

How big, and how fast

The International Energy Agency's central case has data-center electricity roughly doubling from about 485 terawatt-hours in 2025 to around 950 by 2030, which would be close to Japan's entire annual usage, and reaching about 3% of global demand. Traditional data-center load has plateaued, while AI-specific load is exponentially growing. One analysis put ordinary data-center growth at 17% in 2025 while AI-specific facilities grew 50%.

The impact isn't spread evenly because it clusters where the fiber and land already are, which makes problem areas more local. Ireland's data centers were already around a fifth of national electricity a couple of years ago. Two planned facilities in Wisconsin will want a combined 3.9 gigawatts between them, the draw of millions of homes. When that much demand lands on 1 region, that specific grid will suffer greatly.

Why a laptop's worth of code became a power-plant problem

Two things turned this from a big-but-manageable trend into a constraint.

The first is that AI has shifted from training-centric to inference-centric. Training a model is a one-time spike. Serving it to millions of people is a load that runs every hour on the hour. Inference is estimated at around 90% of AI processing today, and it never finishes like a single run would. A model that trained once now answers questions forever, and every answer costs watts.

The second is power density. A rack of modern AI accelerators can draw well over 100 kilowatts, several times what a conventional server rack pulls, and for every 100 megawatts of compute in an average building you spend another 50 - 60 on cooling and distribution. Grids were planned for 20th-century load shapes, not for a single campus that wants the output of a mid-sized power station, running 24-7.

That always on, high utilization requirement is why nuclear keeps coming up. A continuously running inference fleet is a near-perfect match for baseload generation, which is why you are seeing hyperscalers sign nuclear deals and, where the grid cannot connect them in time, build their own gas generation on site. A data center can now up in months so when the wait for a grid connection is longer than the build, you either wait or you bring your own power.

You don't run a data center, but 4 things land on you regardless

Your cloud bill will not fall the way the token price does. Assume inference gets 10x cheaper every year and you can stop thinking about cost, and you will be wrong, because your usage grows into the saving and then past it. Track cost per successful outcome rather than cost per token, because a cheaper model that is 2x as chatty or needs 3 retries is not actually cheaper.

Region is becoming a real decision. Power approval timelines of 2 to 3 years in the busiest markets mean capacity is genuinely tight in specific places, and electricity prices vary enough between regions to move your unit economics on their own.

Architecture is your main lever, and it is entirely in your hands. The teams staying ahead of the curve are not waiting for cheaper models. They route routine work to a small model and save the frontier model for the hard jobs. They cache repeated context instead of resending it. They control context size instead of stuffing the window. They batch anything that can run offline. Each of those cuts calls or watts directly, and none of them requires anyone to build a power plant.

Efficiency is now a feature. For years, "make the prompt shorter" or "call the model less" was just being overly cost-cutting and cheap. With the energy constraint underneath, using less compute for the same result is what keeps your product viable and what regulators and customers ask about. One big prompt isn't necessarily the expensive part. Two hours of iterative back-and-forth to reach the same result can have a much larger energy footprint.

The bottom line

The AI boom stopped being only a story about data and models and became a story about electricity and infrastructure. A doubling of demand in under 4 years is pretty much unheard of in the modern history of electricity, and grids will take years to compensate.

The developers who treat efficiency as part of the job, rather than something to optimize later, are the ones whose products still make economic sense when the usage grows into the company's cloud bill.