ML hardware catalog
Epoch AI ML-hardware catalog (CC-BY 4.0) — accelerators (GPU / TPU / NPU / ASIC …) with manufacturer, release date, TDP, memory, process size, and FP16/FP8 throughput. Rows sort by release date descending. Requires a plan with AI compute & models access (can_use_epoch_ai) — see Pricing for the plans that include it. Response echoes applied_filters.
Authorizations
GDELT Cloud API key. Send as Authorization: Bearer gdelt_sk_....
Query Parameters
Case-insensitive exact match on the accelerator class as recorded by Epoch; the enum lists every stored type. Some tokens are composite, e.g. DCU (GPGPU). An unknown value returns 400 INVALID_ENUM. Full value list: https://docs.gdeltcloud.com/reference/enums#epoch_hardware_type
GPU, TPU, XPU-R, NPU, GPGPU, ASIC, XPU, LPU, Hybrid CPU, DCU (GPGPU), Other Case-insensitive substring match on manufacturer (e.g. Nvidia, AMD, Google, Huawei, Cambricon). Alias org is accepted.
Case-insensitive substring match on the fabricator. Observed values: TSMC, SMIC, Samsung, Intel (plus dual-fab SMIC,TSMC parts, which either SMIC or TSMC returns). Echoed in applied_filters. Not strictly validated — an unmatched value returns 0 rows rather than 400. Observed vocabulary, measured 2026-08-10 (5 distinct values). Values outside it are accepted, not rejected. Discover current values: GET /api/v2/epoch/hardware?limit=1 — readfoundry``. The fab that manufactured the accelerator, as Epoch publishes it. Matched as a case-insensitive SUBSTRING with no validation, so a typo returns an empty 200 rather than a 400. Note some rows carry a comma-joined pair (SMIC,TSMC). Values seen in the corpus: https://docs.gdeltcloud.com/reference/enums#epoch_foundry
"SMIC"
Filter to hardware with release_date on or after this date (YYYY-MM-DD). A malformed date is ignored rather than erroring.
"2023-01-01"
Filter to hardware with release_date on or before this date (YYYY-MM-DD). Combine with date_start for a release-date window.
"2024-12-31"
Rows per page (1–200). Defaults to 25.
1 <= x <= 200Canonical page token — pass the pagination.next_cursor from the previous response here to fetch the next page (uniform across all list endpoints).
Row offset for pagination. Prefer the canonical cursor token; both are supported.
x >= 0