Graphics processing unit, the commonest accelerator, a chip built for the arithmetic of AI. Chapter 4.
From The AI Compute Race. Read the book
Where the book uses it
- Chapter 4 — One Rack
How a datacenter is actually built, from a converted factory in Memphis to a greenfield campus in Texas, and what the clock from first ground to first power really measures.
Where the record uses it
- Campus Stargate Norway
- Campus Stargate UK
Sources that use it
- Chris Miller, Chip War: The Fight for the World’s Most Critical Technology (Scribner, 2022) (the choke-point arithmetic Miller printed in 2022, on 2019 data: Taiwan 37 percent of what he called the world’s new computing power, a share of logic wafer capacity in the report his note cites; two Korean firms 44 percent of memory; ASML 100 percent of EUV lithography — “OPEC’s 40 percent share of world oil production looks unimpressive by comparison”; an advanced logic fab “costs twice as much as an aircraft carrier but will only be cutting-edge for a couple of years”; the foreign-direct-product rule against Huawei, which Miller dates to May 2020 though its reach to any chip sold to Huawei is the rule of August 2020 — any chip made anywhere with US technology requires US permission; a former senior official afterward: “Weaponized interdependence… It’s a beautiful thing”; TSMC founded 1987 with the state providing 48 percent of the capital — “a project of the Taiwanese state”; on EUV: “wasn’t globalized, it was monopolized” (ch. 32, “Lithography Wars”, its last paragraph); on the industry’s paradox, a U.S. semiconductor executive to a White House official: “Our fundamental problem is that our number one customer is our number one competitor” (ch. 49, Everything We’re Competing On); Introduction, of the choke points: “No other facet of the economy is so dependent on so few firms”). Also ch. “The Next Offset”: “One Chinese study has estimated that as many as 95 percent of GPUs in Chinese servers running artificial intelligence workloads are designed by Nvidia” — a cited study’s figure, which the study itself (CAICT, 2021, p. 15) gives as about 95 percent of China’s AI-server market in 2020, a market share and a central figure, not an upper bound, not Miller’s own count (the “in the car with us” line is Matt Turpin’s argument as Miller reports it; the 95 percent figure is a Chinese study Miller cites; Miller’s endnote dates the 37 percent to 2019 Semiconductor Industry Association data and sources the 44 percent to the same report). In the Sources
- AMD, “AAI 2026: AMD Delivers Full-Stack Compute for the Agentic AI Era,” 23 July 2026 (AMD Helios rackscale solutions “now in production to be deployed by leading AI companies at gigawatt scale”, 72 Instinct MI455X GPUs and 18 sixth-generation EPYC “Venice” CPUs to a rack; “OpenAI expects to bring Helios online beginning in the fourth quarter of 2026, with deployments accelerating throughout 2027”; the release names the MI400 Series and the MI455X, and no MI450). In the Sources
- AMD, “AMD and Anthropic Announce Strategic Partnership to Deploy Up to 2 Gigawatts of AMD Instinct MI450 Series GPUs,” 22 July 2026 (up to 2 GW for Anthropic, the first gigawatt beginning in H1 2027; AMD may invest up to $5 billion in Anthropic; “AMD has committed to make a strategic equity investment of up to $5 billion in Anthropic in the future”); OpenAI up to 6 GW and Meta up to 6 GW, each in its own release, at the second and third addresses. In the Sources
- Anthropic, “Higher usage limits for Claude and a compute deal with SpaceX,” 6 May 2026 (the page lists its compute agreements, among them “A 5 GW agreement with Google and Broadcom, which will begin coming online in 2027”, three weeks before the Series H release of 28 May repeated the figure; also “An up to 5 gigawatt (GW) agreement with Amazon, which includes nearly 1 GW of new capacity by the end of 2026”, and the SpaceX agreement for “all of the compute capacity at their Colossus 1 data center”, “more than 300 megawatts of new capacity (over 220,000 NVIDIA GPUs) within the month”; in the record) In the Sources
- Anthropic, “Series H,” 28 May 2026 ($65 billion at $965 billion post-money, MGX a significant investor; “We signed agreements with Amazon for up to five gigawatts of new capacity, with Google and Broadcom for five gigawatts of next-generation TPU capacity, and with SpaceX for access to GPU capacity in Colossus 1 and Colossus 2”). In the Sources
And 156 more: search the Sources.
Related terms
From the glossary of The AI Compute Race.
How to cite this page
Dryburgh, Lee S. “GPU.” In The AI Compute Race. DRYBURGH. https://dryburgh.com/vol1/glossary/gpu (accessed 2026-11-01).
BibTeX
@misc{dryburgh-gpu-2026,
author = {Dryburgh, Lee S.},
title = {{GPU}},
howpublished = {In \emph{The AI Compute Race}, DRYBURGH},
year = {2026},
url = {https://dryburgh.com/vol1/glossary/gpu},
note = {Accessed 2026-11-01}
}