Dylan PatelFounder and Chief Analyst, SemiAnalysis · The semiconductor-AI stack

Chips, power, and a little chaos

8 Oct 2025 · 1 min

Watch on YouTube

0:42

In this short commentary, the speaker marvels at an unprecedented AI infrastructure rollout in Memphis, Tennessee. Within six months, the subject acquired a factory, installed roughly 100,000 GPUs, and launched model training, employing large‑scale liquid cooling, external CAT turbine generators, and mobile substations to supply 200‑300 MW of power. The effort is now scaling toward a gigawatt‑level power footprint at a similarly rapid pace. The speaker expresses awe mixed with a hint of desensitization, likening the experience to a child overwhelmed by candy and no longer appreciating apples.

Chapters

What he was doing in Tennessee, in Memphis, first time was crazy, right? 100K GPUs in 6 months. He bought a factory in like February of '24 and had models training within 6 months. And he did liquid cooling at this scale for AI, all these sorts of crazy firsts, putting generators outside like CAT turbines, all these things for different things to get the power, mobile substations, all these different crazy things. So he does this, it's like, holy crap. And he did for 100K GPUs, you know, 200, 300 megawatts, right? Now he's doing it for a gigawatt scale and he's doing it just as fast. You would think like, this is obviously way more impressive that he did it again. Maybe I'm desensitized, but like, it's like, you know, like you've given the child too much candy and now like the child doesn't like apples.