The AI accelerator

Assembled from 19 blocks · 1 evidence · 29 related

1. Story
2. The specialised processor — graphics processing units, and the tensor-oriented designs that followed them — on which essentially all frontier training runs. Deep learning became practical partly because the linear algebra it needs happened to match hardware built for real-time graphics, an accident of fit that the industry has spent a decade converting into deliberate design. Two properties of the supply chain matter more to this Journey than any specification. Fabrication of the leading-edge nodes these parts require is concentrated in a very small number of facilities, most of them in Taiwan, with capacity at the leading nodes reported by trade press as fully committed through 2026; and the parts are subject to export controls that make their distribution a matter of policy rather than of markets. A capability that depends on a device made in one place and licensed by one government is not distributed the way software is.
3. Knowledge
4. Taiwan
5. The AI accelerator
6. Connections
7. Deep learning
8. The transformer architecture
9. Training compute
10. Taiwan
11. A deep convolutional network wins the ImageNet challenge
12. The Luzon Strait
13. Austronesian
14. The Austronesian expansion
15. Blust's Austronesian subgrouping places the deepest divisions in Taiwan
16. Austronesian speakers move out of Taiwan
17. The AI accelerator
18. Evidence
19. Training compute for frontier AI models grew by roughly 4–5× per year from 2010 to 2024, with frontier language models growing at about 5× per year since 2020; Epoch also maintains a database of large AI data centres built from satellite imagery and permit filings, with concentrations of covered sites in Texas, Virginia, Ohio, Nebraska and Iowa. V55 verification basis: epoch.ai was not fetched. The 4–5× and 5× growth rates were carried in search results attributing them to Epoch's own published trends analysis of that title, and the data-centre database and its state-level site concentrations were carried with attribution to Epoch. Forward projections of compute growth, coverage percentages for the data-centre database, and GPU-equivalent counts appearing in secondary summaries were NOT confirmed and are excluded from this pack.