By Nathan Benaich (Air Street Capital)Topic sample · January 1 to June 1, 2025
In the topic sample, LLM research is overrepresented among papers using AMD MI300, AMD MI250, Huawei Ascend, and NVIDIA H100/H200. Their LLM shares exceed the overall paper sample’s baseline by 42.5, 34.2, 33.9, and 27.8 percentage points, respectively.
This measures specialization, not which chip appears in the most LLM papers or runs them fastest. The analysis covers 6,356 papers, uses model-assigned topic labels, and requires at least three papers per chip-topic pair. It is a 2025 sample, separate from the 2026 citation estimates.
Citation estimates use inputs through September 1, 2026; fleet counts have an October 1, 2026 cutoff. The research-topic sample covers January 1 to June 1, 2025. See each answer for its period and limitations.
All data sourcesBenaich, Nathan. “Which chips are associated with LLM research?” State of AI Report Compute Index. Web page updated 2026-10-11; data periods as specified above.