The AI Bottleneck Moved From Chips to Power

Industry · Working with AI

For two years the scarce thing was silicon. 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. That constraint has moved. The scarce thing now is electricity, and where you can get it.

The clearest sign showed up as a cancellation. In the same week Gartner reported that global data-center electricity use would hit 565 terawatt-hours in 2026, a 26 percent jump in a year, around 75 data-center projects worth roughly 130 billion dollars were postponed or canceled across the United States. Not for want of money. Not for want of chips. For want of power to run them.

I think this is the most under-covered story in the industry, because it does not look like a software story and most of us are software people. But it reaches all the way down to your cloud bill and your architecture, so it is worth understanding even if you never go near a substation.

How big, and how fast

The numbers are large and they point the same way. 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 use, and reaching about 3 percent of global demand. The split inside that total matters more than the total. Traditional data-center load has plateaued. AI-specific load is compounding: one analysis put ordinary data-center growth at 17 percent in 2025 while AI-specific facilities grew 50 percent. The two curves share a building but move at different speeds.

And the load is not spread evenly. It clusters where the fiber and the land already are, which turns a national trend into a set of local crises. Northern Virginia is the extreme case. 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 one grid region, the grid is what gives.

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 inference now dominates. Training a model is a one-time spike. Serving it to millions of people is a load that runs every hour of every day. Inference is estimated at around 90 percent of AI compute today, and unlike a training run it never finishes. 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 to 60 on cooling and distribution. Grids were planned for twentieth-century load shapes, not for a single campus that wants the output of a mid-sized power station, running around the clock.

That last property, always on, high utilization, predictable, 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. Interconnection queues in the big markets now run past three years. A data center goes up in months. When the wait for a grid connection is longer than the build, you either wait or you bring your own power.

The paradox that keeps finance awake

The part that matters most if you build on this is also the strangest. The price of a token collapsed. The bill went up.

Both halves are true, and both are dramatic. Serving a given level of capability has fallen roughly tenfold in a year, and something like a thousandfold over three: a capability that cost around twenty dollars per million tokens in late 2022 costs well under a dollar now. That is one of the fastest cost declines in the history of computing.

And over roughly the same window, token consumption rose more than a hundredfold. Cheaper tokens did not save money. They made more things worth doing, so we did far more of them. Every surface that was too expensive to put a model behind a year ago now has one. Agents that call the model in a loop, retries, long context, tool calls, always-on assistants: each of those multiplies calls, and the multiplication outran the price drop. The units got cheaper, we bought far more of them, and the bill still rose.

The reason this connects back to power is that the consumption is real electricity, and electricity is not falling tenfold a year. The token price is a software-and-hardware efficiency curve. The energy bill underneath it is physical, and physics is not on the same schedule.

What this actually means if you build on top

You do not run a data center, but four things land on you anyway.

Your cloud bill will not fall the way the token price does. Assume inference gets ten times 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 twice as chatty or needs three retries is not actually cheaper.

Region is becoming a real decision. Power approval timelines of two to three years in the hottest markets mean capacity is genuinely tight in specific places, and electricity prices vary enough between regions to move your unit economics on their own. The lazy default of nearest region is quietly a cost and availability decision now.

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 tail. 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 penny-pinching. With the energy constraint underneath, using less compute for the same result is what keeps your product viable and, increasingly, what regulators and customers ask about. Green computing moved from a pledge to a constraint this year, and the pressure only points one way.

The honest 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 four years has no precedent in modern energy history, and the grid takes years to expand.

None of this is a reason to stop building. It is a reason to build as though compute costs something real, because it does: not just dollars on an invoice, but watts on a strained grid. The developers who treat efficiency as part of the craft, rather than something to optimize later, are the ones whose products still make economic sense when the usage grows into the bill. Which it will.