We did the math on AI’s energy footprint. Here’s the story you haven’t heard. | MIT Technology Review

Good article attempting to get to the bottom of the electricity requirements of AI capabilities. The short answer by the way is it is super complicated because every task is unique, and the required GPU capabilities to serve them are all unique.

Powering data centers is a huge amount of energy overall. Not just for AI, but for everything. In 2014, 1.8% of all power was used by data centers and it could be 6-12% by 2028. Some of this is AI, some is crypto, some is cat pictures, some is that killer Instagram post. It is a lot of power.

Data centers are also amenable to variable power as they have variable load. Overall a smart grid working in concert with smart datacenter can be very dynamic. And, for all that electricity there is a lot of value created.

We need to understand and manage these loads, but I don't agree with calls that electricity usage are reasons to stop or limit AI use. I also didn’t agree with that as an argument against Bitcoin and crypto.

Simon Willison did a great writeup on this article too. This article is part of an AI Energy Package if you want to read more.