Code for estimating quantities of AI chips
For more information, see: https://epoch.ai/data/ai-chip-sales https://epoch.ai/data/ai-chip-owners
nvidia_estimates, tpu_estimates, amd_estimates, etc generate estimates for chip sales by designer
nvidia_owners allocates Nvidia chips to hyperscaler and official Chinese owners using a revenue-based model. Other Nvidia owners are modeled separately in nvidia_owners_other, with the exception of smuggled Chinese chips, which are handled separately in v_diversion_and_resale.
Chip-family models are being moved out of notebooks into <family>_model.py modules
(TPU so far). For those families:
python run_chip_model.py tpu(orall) reruns the model, rewrites thecsv_export/andowners_csv_export/files, and validates the exports against the published dataset, ending with a per-family PASS/REVIEW summary.<family>_estimates.ipynbis the research notebook: it imports the model module and renders tables, charts, and sensitivity checks without owning the model logic.python validate_chip_sales.py --designer all(or thevalidate_chip_sales.ipynbnotebook) compares the local CSVs against the currently published dataset on its own.
Families without a model module yet (Nvidia, AMD, Amazon) still run via their notebooks.