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

The AI buildout, quantified | SemiAnalysis

15 Oct 2026 · 2 min

Watch on YouTube

1:31

In this narrated piece, the speaker reveals the hidden physical and economic layers behind every AI query. Each token produced by a model carries a price, rooted in the massive hardware that powers it. The description walks through the micro‑scale of chips—72 per rack, stacked memory, billions of switches—fabricated on precision wafers in fabs worldwide. These chips are housed in an ever‑growing network of data centers, now numbering around 5,000 globally with over 1,000 in China, each consuming power comparable to a small city and fed by thousands of transformers, turbines, and 25,000 generating units. The cost of this compute infrastructure fluctuates faster than anyone can track, and the narrative even hints at future extensions beyond Earth. Overall, the piece paints a picture of a beautiful yet frightening machine whose scale, energy demand, and financial dynamics are rapidly expanding.

Chapters

Every time you ask a machine something, somewhere, something answers. That answer is made of tokens, and every token has a price. It was born on a chip you'll never see, rented by the hour from a cloud you had to trust. 72 of those chips share a single rack. They breathe heat and they talk in light. Each one is wrapped Trapped in memory, stacked like pages around billions of switches, so small we have to cut them open to see them. They're printed on wafers in fabs scattered across the world by machines of almost impossible precision. Then they come home to buildings that drink power like cities, 5,000 of them and counting, over 1,000 in China. Fed by transformers and turbines, 25,000 generating units humming Through the night. And all of it is paid for by someone at some price. It changes every month, faster than anyone can follow. Some of it may even leave the Earth. It's a beautiful, frightening machine, and we've been watching every part of it.