Series · 5 essays · Chemical engineering
The Science of Making Tokens
A five-part essay series tracing the actual supply chain behind a single AI token: from a quartz mine to one forward pass. Written from a chemical engineer's perspective, because AI is not magic, and treating it like it is makes you bad at deploying it.
Why this exists
I trained as a chemical engineer. I spent years running a chemical vapor deposition rig in an undergrad chemistry lab, growing graphene on copper and silicon and nickel and almost never getting consistent results. That experience left me with one durable conviction: the things we now call AI are not magic, and the people deploying them well are the ones who treat them as physical, quantifiable, and bounded by the real-world constraints of the systems that produce them.
Most of the AI conversation skips this entirely. The discourse talks about the model, sometimes about the training data, and almost never about the substrate. The substrate is the interesting part. The substrate is the part that explains why some things are easy and others are unreachable, why some companies have moats and others do not, and why the next decade of AI is going to be shaped at least as much by ASML and TSMC and the western power grid as by the labs everyone reads about on twitter.
This series is my attempt to write the version I wish existed when I first started talking to clients about AI deployment. Quantitative, grounded, demystified, and useful to anyone who has to make real decisions about what to build, what to buy, and what to ignore.
The series
- 01 · Sand into Silicon. Quartzite to 9N polysilicon via Siemens process. The purest commodity humans make at industrial scale.
- 02 · Pulling the Crystal. Polysilicon to a single-crystal 300mm wafer via Czochralski. Four firms make almost all of them.
- 03 · Photolithography. Eighty rounds of patterning on a single wafer. The most complex object humans make at scale.
- 04 · A City's Worth of Power. Training cluster economics, siting decisions, and what 100 MW actually buys you in compute.
- 05 · The Forward Pass. Inference at the silicon level. Per-token energy, throughput, and a full cost reconciliation.
Who this is for
Operators trying to scope an AI initiative and wanting to know what their vendor is actually buying. Engineers and PMs at the labs who want a quick read on the downstream economics. Investors trying to model the long-tail cost structure. Anyone who has been hand-waved at one too many times by someone who treats AI as either pure software or pure magic, and wants the engineering reality in between.
An interactive companion
The simplified process flow is on the home page as an interactive component. Hover any step for the engineering numbers, click through for the full breakdown.