This piece is the first in a new series from the Institute for Progress (IFP), called Compute in America: Building the Next Generation of AI Infrastructure at Home. In this series, we examine the challenges of accelerating the American AI data center buildout. Future pieces will be published

Substack is a platform for independent writers and journalists to publish and monetize their content. Through newsletters, articles, and podcasts, Substack offers insights into a wide range of topics such as politics, technology, culture, and personal development. Readers can subscribe to their favorite writers and receive regular updates, analysis, and commentary on the issues that matter to them. Additionally, Substack provides tools and resources for writers to grow their audience, engage with their readers, and monetize their content effectively.

Substack

The rise of AI will accelerate the trend of building larger and more power-intensive data centers. Finding enough power for these data centers will become increasingly challenging.

How to Build an AI Data Center

I’ll add:
<ol>
<li>
Don’t build a data center. Start in the cloud, then continue in the cloud. Heard at AWS re:invent 2014: friends don’t let friends build data centers.
</li>
<li>
Migrate to a non-Transformer architecture with 100-1000x cheaper training and inference costs. Doesn’t exist yet? Give it a year or two (your datacenter would barely be up and running by then anyway) and thank me later.
</li>
<li>
Rejoice in the knowledge that:
<ul>
<li>you just saved tens or hundreds of millions of dollars</li>
<li>your competition is sitting on thousands of 95% idle 2-year-old GPUs that have depreciated to less than 1/10th what they paid during surge pricing</li>
<li>your cost per GPU-hour in the cloud has already been lowered four times due to competition between cloud vendors</li>
<li>you are on the latest generation of GPUs because the cloud lets you pick and choose what you provision and you didn’t get suckered into a 3-year reserved/prepaid contract</li>
<li>and you are only paying for what you actually use, not your peak predicted training load from the Transformer days</li>
</ul>
</li>
</ol>