Flexible AI Data Centers

Everyone is racing for speed to power, and the default route runs on gas

In January 2026 I grew obsessed by one question: how do we power the AI boom more sustainably?

Every AI lab is racing to cut its time to first token, and every data center developer is racing for speed to power. Next to speed, generation efficiency and the cost of electricity barely register. The trouble is what the race runs on. Facing 5 to 7 year waits to connect to the grid,¹ developers are kicking their sustainability goals down the road and going off-grid on gas. I wanted to know whether speed and sustainability can be reconciled.

I started in a Stanford class and ended up on the road across 10 cities

I began in Stanford’s Innovation for Climate and Sustainability class. My team and I interviewed 70+ people between January and March, and I went to DistribuTECH in San Diego. That became Part I in April.

I kept going through the spring as an independent study under Steve Westly’s guidance. I added 30+ interviews and took my argument to SF Climate Week and Stanford’s Sustainable Data Centers Symposium, where I learned enough to overturn part of it. That became Part II in June.

Then I spent the summer in the field on Stanford’s EcoSprint program. I packed a cabin bag and booked flights and conference tickets. I went to 3 conferences, visited a utility, a data center operator and 3 power and computing companies, and interviewed 50+ people, 40+ of them in person, across 10 cities. That became Part III in October.

Transmission is the bottleneck, and one new line takes 10 to 15 years to permit

My hypothesis going in was naive. I thought the US needed to build more clean generation. It already has 2,000+ GW of power projects waiting in line to plug into the grid,² close to double everything plugged in today. The constraint is transmission, and permitting a single new line takes 10 to 15 years.³ Utilities cannot build transmission that fast, so they put new data centers in a queue and make them wait.

Developers who cannot wait are building their own power behind the meter

Developers plan to build about 30% of new US data center capacity on power they own rather than power the utility delivers.⁴ Gas turbines are the quickest thing to put there, which is how a race for speed becomes a problem for emissions.

Flexibility is how speed to power and sustainability reconcile

A data center that can adapt its power demand, and cut it in the few hours each year when the grid is strained, can connect years sooner without a gas plant of its own. Researchers at Duke’s Nicholas Institute estimate that the existing grid could take 98 GW of new load if large users cut 0.5% of what they draw in a year.⁵ All 3 papers test that idea and ask who could build a business on it.

Of the 19 startup ideas I evaluated, 2 stood out

In Part III I sort 19 ideas by the stage of a data center project they belong to, and say what I think of each. My top idea is batteries at factories and warehouses, pooled and sold to data centers as capacity. My second is software that lets the small crew at a campus run its on-site power and cooling from one place.

1. Camus Energy, encoord, and Princeton ZERO Lab, Flexible Data Centers: A Faster, More Affordable Path to Power. Back
2. Lawrence Berkeley National Laboratory, Queued Up: 2024 Edition. Back
3. RMI, How Virtual Power Plants Can Help the United States Win the AI Race. Back
4. Cleanview, Bypassing the Grid: How Data Center Developers Are Building Their Own Power Plants. Back
5. Nicholas Institute, Duke University, Rethinking Load Growth. Back