Skip to content
VST—no feedCEG—no feedNRG—no feedTLN—no feedNEE—no feedSO—no feedDUK—no feedAEP—no feedD—no feedEXC—no feedPEG—no feedXEL—no feedGEV—no feedETN—no feedPWR—no feedVRT—no feedCMI—no feedBE—no feedNVDA—no feedAMD—no feedAVGO—no feedMSFT—no feedAMZN—no feedGOOGL—no feedVST—no feedCEG—no feedNRG—no feedTLN—no feedNEE—no feedSO—no feedDUK—no feedAEP—no feedD—no feedEXC—no feedPEG—no feedXEL—no feedGEV—no feedETN—no feedPWR—no feedVRT—no feedCMI—no feedBE—no feedNVDA—no feedAMD—no feedAVGO—no feedMSFT—no feedAMZN—no feedGOOGL—no feed
Join the waitlist
The claim

Everything is compute.

Every model you use is electricity turned into an answer. Not metaphorically — that is the actual conversion. You type a question, a building somewhere draws power, and the power comes back to you shaped like a sentence.

For most of the last decade the interesting question about that building was which chips were inside it. That was a reasonable question. The chips were the scarce thing, the expensive thing, the thing with a waiting list. Every story about artificial intelligence was, underneath, a story about silicon.

That framing has quietly stopped being the whole picture, and the reason is boring in the way that important things usually are.

Follow the input

A chip is a device for converting electricity into arithmetic. It is very good at this and gets better every generation. But it does not create the electricity. Something upstream has to supply it, continuously, at industrial scale, in a specific physical place, with no meaningful ability to store it.

So when you ask what an AI company is buying, the honest answer has two layers. It is buying accelerators. And it is buying the right to run them.

Every AI story is a compute story.

That is the first step, and it is the uncontroversial one. The next two are where it gets interesting.