Why does AI need so much memory?
Billions of learned numbers. Somewhere to hold them. Room to think.
Before an answer, there is a place for the numbers.
A luminous assembly opens up the difference between storage, working memory, and high-bandwidth memory. Green holds the model; cyan follows the work in progress.
Blocks and light form a schematic computer memory landscape. The moving sculpture carries the explanation continuously, with narration and a music bed throughout.
Look closer: the mathematics
Seven billion weights at two bytes each occupy 14 GB in decimal units. This 16-bit example counts model weights only; conversation state, intermediate calculations, and runtime overhead need additional memory. Lower precision changes the requirement. The sculpture is schematic, not a literal chip layout. The film quotes Micron’s September 30, 2026 results and outlook. Its 2027–2028 supply outlook is a company forecast, not a guaranteed outcome.
Sources
- Micron earnings release · September 30, 2026
- Micron prepared remarks · market outlook, page 5
- NVIDIA · model memory and inference
- Made
- Oct 7, 2026
- Study
- Micron news edit
- Sound
- Dagslända — Ave Air
Ave Air / Dagslända / courtesy of Epidemic Sound