Tokens & Transistors

How AI actually runs: the memory, chips and math behind every chatbot reply, animated in 3D with every number sourced.

Each episode follows one piece of AI infrastructure down to the hardware. Why can one long chat fill 43 GB of GPU memory? Why does a GPU serving one person leave more than 99% of its math idle? Why can cached tokens cost a tenth of the price?

For developers and technical people who use AI tools and APIs and want to know what happens on the other side of the request. No linear algebra needed.

Every figure on screen comes from a paper, spec sheet or model config, linked in the video description. Estimates are labelled IDEALIZED. Animated with Remotion and three.js; narrated with an AI voice.

Series: Under the Hood: AI Systems
Coming up: speculative decoding, quantization, mixture-of-experts.
New long-form episode every other week, Shorts in between.

Made by Nasim, founding full-stack engineer at ed-tech startups.