why now

Artificial intelligence changed the shape of computing demand, and distributed technologies before it had already begun to. Training and inference at frontier scale consume power in blocks of hundreds of megawatts, arriving over months rather than the steady single-digit growth that utilities spent forty years planning around.

Data centers used roughly 4.4% of US electricity in 2023. Credible forecasts put that between 9% and 17% by 2030, and the forecasters disagree meaningfully about where in that range it lands. What they agree on is more useful: data centers already account for about half of all growth in US electricity demand.

The disagreement is about magnitude. Nobody credible is arguing about direction.

In 2020 the average data center rack drew eight to nine kilowatts. A current-generation AI training rack draws well over a hundred and thirty, roughly fifteen times the density of the installed base, concentrated into the same floor area.

That breaks the assumptions underneath conventional design. Air cooling stops being adequate long before those densities, so liquid cooling becomes structural rather than optional. Power distribution built for a twenty-kilowatt rack needs replacing rather than upgrading. And the substation capacity required now lands on a far smaller footprint, which changes where these facilities can go at all.

Each new chip generation has raised the number. Nothing on the public roadmap reverses it.

Connecting new capacity to the grid takes longer than building the thing that needs connecting. The median project reaching commercial operation in 2025 waited more than five years from its initial request, and the overwhelming majority of capacity that enters an interconnection queue never gets built at all.

This is now the defining constraint of the industry. Semiconductor supply is solvable with capital and time. Construction is solvable with capital and time. Grid access is solvable with neither — which is why it determines who builds and who waits.

Regulators are working on it, and reform will help eventually. It will not help a project that needs power this decade.

Jurisdictions across the country are writing policy for this industry where none existed eighteen months ago, permitting frameworks, cost-allocation rules requiring new load to pay for the infrastructure it needs, water and siting standards, and in some places outright pauses on new approvals.

We think this is broadly correct, and we have built our approach to work within it rather than around it. Developers who treat community consent as an obstacle are producing the litigation and the moratoria that make the next project harder for everyone. We would rather be the counterexample.