High-bandwidth memory (HBM), the stacked memory in AI accelerators, now makes up most of their component bill. Epoch AI estimates its share of component spending rose from 52% in early 2024 to 63% in late 2025, averaged across AI chips from Nvidia, AMD, Google and Amazon and weighted by production volume. Source.
The estimate covers four categories (memory, logic dies, advanced packaging and auxiliary parts), not complete servers or the cost of running models, and per-chip costs carry uncertainty. Epoch expects HBM's share to grow in 2026 as memory supply stays tight and prices rise.
