Energy suck

The con­crete is flowing in Abilene:

That’s the first Star­gate data center, intended for OpenAI’s use, cap­tured by the Sentinel-2 satel­lites between March 7 and April 14: a little over a month of work. The sense of urgency is palpable, even from space!

It’s ironic that, just as engi­neers broke the curve of energy efficiency, pro­ducing chips (like Apple’s M series) that yanked the Pareto fron­tier of power vs. energy into undreamt-of territory, the soft­ware industry found an out­landish new need for raw compute.

It’s tempting to say “this always happens”—humans are always dis­cov­ering new needs, new desires, right? — but I don’t think it does. Dri­vers of fuel-efficient cars (or, indeed, elec­tric cars) don’t sud­denly develop dreams of elab­o­rate road trips. (Here’s a ton of data regarding that sce­nario.) No, this doesn’t always happen; it just hap­pened this time; and it’s an epochal bummer.

By the mid-2010s, a sur­prising pic­ture was coming into focus, driven by the require­ments of mobile phones: a world of super-pow­erful devices con­suming energy in dainty sips. Just like the cam­eras on those phones, there seemed to be some­thing physics-breaking about it — the result of incred­ible inge­nuity and care.

And then came AI.

In fairness: first it was AWS, then GCP and Azure and all the rest. Then, having per­fected the design for this kind of facility — hot and dark, down by a river somewhere, a mil­lion square feet watched over by, at most, a few dozen technicians — adaptation for AI was straightforward … and, oops: those Nvidia chips are energy hogs.

Here’s the cur­rent pic­ture:

The [Department of Energy] finds that data cen­ters con­sumed about 4.4% of total U.S. elec­tricity in 2023 and are expected to con­sume approx­i­mately 6.7 to 12% of total U.S. elec­tricity by 2028.

This is a wild trendline!!

The loads pro­duced by data cen­ters are weird, in the sense that they are steady, day and night; nearly every other load (including indus­trial loads) has some kind of cyclic pattern. This is causing new prob­lems for some power grids.

In your imagination, copy and paste the satel­lite imagery above across the U.S.: Louisiana farmland, Nebraska cornfields, Ari­zona desert … Everywhere, the con­crete is flowing, and soon the server racks will sprout. Another cool bil­lion for Nvidia’s 10-K, and another, and another.

The AI data center buildout seems wasteful to me because it’s duplicative: a bunch of com­pa­nies all racing to acquire exactly the same thing. Imagine another timeline, in which a stable, savvy U.S. gov­ern­ment declared AI research a national pri­ority and orga­nized an ambi­tious project to pursue it: bringing inves­ti­ga­tors together into super-powered labs, funding the con­struc­tion of national data cen­ters, making them avail­able to aca­d­emic researchers — the pio­neers who have, in our timeline, been effec­tively sidelined.

(Remember, the true sub­lime of the Man­hattan Project wasn’t the sci­ence camp at Los Alamos: it was the pop-up ura­nium and plu­to­nium enrich­ment facilities, doing exacting chem­ical work at basi­cally incon­ceiv­able scale. If you are looking for “state capacity”, there it is.)

A clear remedy to the moral problem of a tech­nology built on the commons would be the com­mit­ment of these tools and research, plus their outputs, back to the commons.

This is a sce­nario held in wistful regard by plenty of folks in the AI industry, by the way.

Of course, there’s a response that goes: “Let us cook — after super­in­tel­li­gence is born in one of these data cen­ters, it will invent lim­it­less free energy, and we won’t have to worry about any of this.” I’m open to exploring the unknow­able ben­e­fits of pow­erful AI, but this response is laughable, because we already have the poten­tial for lim­it­less free energy. No fur­ther inno­va­tion is required — only coordination.

It feels totally appro­priate to say to OpenAI, Google, and the rest: sure, build all the data cen­ters you want … just be sure to budget for the solar farms and bat­teries to sup­port them.

There’s room for big con­crete boxes in this world — I’ve done some great work in big con­crete boxes — but I find this par­tic­ular species trou­bling rather than inspiring. There is no get­ting over the tragedy of finding our­selves on the cusp of a new energy regime for computers, only to tumble back into the furnace.

The lap­tops these days run without fans, cool and quiet, faster than ever. Meanwhile, the long cor­ri­dors of the data cen­ters roar, and roar, and roar.