High-bandwidth memory, memory chips stacked one on another beside an accelerator. Three makers supply it, two of them Korean, and it is booked years ahead. Chapter 5.
From The AI Compute Race. Read the book
Where the book uses it
- Chapter 5 — The Island in Every Machine
Why every accelerator runs through Taiwan: the foundry, the packaging and the memory that are booked years ahead, and what that does to anyone arriving late with cash.
Where the record uses it
- Scorecard indicator HBM
Sources that use it
- Amin Vahdat, “Our eighth generation TPUs: two chips for the agentic era,” Google, The Keyword, 22 April 2026 (the eighth-generation TPU announced as two parts, TPU 8t for training and TPU 8i for inference, designed with Google DeepMind; TPU 8t nearly 3x compute performance per pod, 9,600 chips and 121 ExaFlops a superpod; TPU 8i 288 GB of HBM with 384 MB of on-chip SRAM; both up to twice the performance per watt of Ironwood and “generally available later this year”; both parts announced as “coming soon”). In the Sources
- Eric Xu, keynote, Huawei Connect 2025, 18 September 2025 (Ascend 950/960/970 roadmap; Atlas 950 and 960 SuperPoDs; “we will offer two proprietary HBMs”; the chips’ planned availability, the 950PR in the first quarter of 2026, the 950DT in the fourth quarter of 2026, the 960 in the fourth quarter of 2027 and the 970 in the fourth quarter of 2028). In the Sources
- Meta, “Four MTIA Chips in Two Years: Scaling AI Experiences for Billions,” 11 March 2026 (“we have deployed hundreds of thousands of MTIA chips in production”; the post names no foundry; the post’s MTIA 300 is the generation that carries HBM; Meta’s post of 10 April 2024 gives the first two generations off-chip LPDDR5 memory, not HBM, and TSMC’s 7- and 5-nanometer processes). In the Sources
- Micron Technology, first-quarter fiscal 2026 prepared remarks, 17 December 2025 (price and volume agreements completed for the entire calendar-2026 HBM supply). In the Sources
- Micron Technology, “Fiscal Q3 2026 Earnings Call Prepared Remarks,” 24 June 2026 (sixteen strategic customer agreements across the memory business, not HBM alone: “we have completed 16 SCAs with customers across the data center, consumer and auto market segments”; “The 16 signed agreements represent roughly 20% of our DRAM volume and a third of our NAND volume over this period”, typically “a five-year term, from calendar 2026 through the end of calendar 2030”; “They provide committed DRAM, including HBM as appropriate, and NAND supply”; for the agreements signed so far, including some after the quarter ended, “RPO is approximately $100 billion”). In the Sources
And 25 more: search the Sources.
Related terms
From the glossary of The AI Compute Race.
How to cite this page
Dryburgh, Lee S. “HBM.” In The AI Compute Race. DRYBURGH. https://dryburgh.com/vol1/glossary/hbm (accessed 2026-11-01).
BibTeX
@misc{dryburgh-hbm-2026,
author = {Dryburgh, Lee S.},
title = {{HBM}},
howpublished = {In \emph{The AI Compute Race}, DRYBURGH},
year = {2026},
url = {https://dryburgh.com/vol1/glossary/hbm},
note = {Accessed 2026-11-01}
}