AI Accelerator Manufacturing Cost

Name: AI Accelerator Manufacturing Cost
Creator: Silicon Analysts

Estimated manufacturing cost per AI accelerator — logic die, HBM, packaging (CoWoS), and gross-margin split. 18 chips from H100 through the 2026 generation via the API; the page walks through 8 flagship parts.

Coverage window: 2026-04-07 → 2026-08-16 (21 rows)
Computed from the backing series at page-render time — never typed by hand.
History honesty: Current-state table; the vintage of each value is its as_of. Time series accrues via the daily ledger and *_history tables, not here.
Cadence: Weekly — cost models recompute when inputs move (Monday refresh fleet, 11:00–12:30 UTC); values are stable between input changes.
Revision policy: Current-state table updated in place with as_of (vintage) and last_validated (confirmation). Prior values persist in the *_history audit tables and the daily ledger.
Known gaps: Every figure here is Silicon Analysts' bottom-up cost model — die, memory, packaging, and margin estimated from public component and foundry data — not a vendor disclosure; vendors do not publish accelerator costs. Each value carries its as_of date. Twenty-one accelerators are modeled, 18 of which are served via the API. The 2026 generation — NVIDIA GB300 and Rubin, AMD MI455X (MI400 series), Google TPU v6e (Trillium), AWS Trainium 3 — was added 2026-08-14 on current HBM contract pricing; within those, Rubin's BF16/FP8 throughput and die size are modeled from vendor NVFP4 figures pending disclosure, and no 2026-generation part carries a vendor-published package TDP — all five are modeled or inferred (GB300's per-GPU 1,400W TGP is the one vendor-stated anchor). The 16 pre-existing models were revalidated 2026-08-14 (no public signal justified moving any street price); on 2026-08-16 the six actively-procured HBM3e models were re-based to current ~$17/GB contract pricing, while legacy HBM3/HBM2e rows keep their vintage. The two HBM4 parts deliberately differ: Rubin is costed at its settled 2026 supply price (reported ~$560 per stack), while MI455X carries a forward ramp-period estimate (~$35.5/GB) — their per-GB HBM costs are not comparable to each other or to the HBM3e rows. B100, MI300X, MI325X, MI355X, and Groq LPU currently have no recent public unit-price signal at all. Google TPU v7 (Ironwood, GA April 2026) is not yet covered; adding it is planned.
Authored 2026-08-16; next review 2026-11-15.
Access: Fully open
API: /api/v1/accelerators
Machine-readable: /api/v1/catalog/ref-ai-chip-costs/contract