The Custom AI ASIC State of Play (May 2026) — Broadcom Deals, Google TPUs, Meta MTIA & Beyond
May-2026 ASIC state-of-play: Broadcom ~60% / Marvell ~35% of the custom-AI-ASIC co-design market; Marvell ~$11B AI ASIC revenue 2026; programs — TPU v7 Ironwood (4,614 FP8 TFLOPS, 192GB HBM3E, 7.37 TB/s), Maia 200, Trainium3 (Q2 2026), MTIA, OpenAI Titan.
The Custom AI ASIC State of Play (May 2026) — Broadcom to MTIA
Source: Tom's Hardware — semiconductor industry deep-dive Published: 2026-05-26 | Ingested: 2026-06-24 Authority: analysis (established hardware-industry publication, analyst-style survey). Moderate.
fetch_status: summary-derived. The article body did not render on WebFetch (JS/membership wall); the URL is preserved and the claims below are drawn from the search-result summary of this exact article plus the TrendForce/SemiAnalysis cross-sources already in the KB. Figures should be treated as moderate (single-source-summary unless corroborated).
Summary
A consolidated map of the hyperscaler custom-AI-ASIC landscape as of mid-2026 — who designs the chips, who the merchant co-design partners are (Broadcom, Marvell), and the specs/timing of each program. This is the breadth source that fleshes out the named programs the KB's custom-silicon-vs-gpu page lists in its table.
Key Claims (summary-derived)
The co-design duopoly
- Broadcom ~60% of the custom-AI-ASIC co-design market — anchors Google TPU, Meta MTIA, Microsoft Maia, and OpenAI's "Titan" program. Evidence: moderate (single-source summary)
- Marvell ~35% of the co-design market — partners with Amazon (Trainium) and Microsoft (Maia). Evidence: moderate
- Marvell projects up to ~$11B AI ASIC revenue for 2026. Evidence: moderate
- The investment thesis: Broadcom + Marvell are the "real winners" of the ASIC war — they capture custom-silicon value regardless of which hyperscaler wins. Evidence: weak (analyst framing)
Named hyperscaler programs
| Chip | Company | Notable specs / status (2026) |
|---|---|---|
| TPU v7 "Ironwood" | 4,614 FP8 TFLOPS/chip, 192 GB HBM3E, 7.37 TB/s bandwidth; sold externally (Anthropic) | |
| Maia 200 | Microsoft | 2nd-gen custom accelerator, arrived early 2026 |
| Trainium3 | AWS | Ramps production from Q2 2026 (follows Trainium2 / 2.5) |
| MTIA | Meta | Inference accelerator, Broadcom-partnered |
| "Titan" | OpenAI | Custom program, Broadcom-partnered |
All specs evidence: moderate (single-source summary).
Market context
- Custom ASICs growing at 44.6% CAGR (vs ~16% merchant GPU), targeting the two-thirds of AI compute that is now inference. Consistent with the SemiAnalysis and TrendForce sources already in the KB. Evidence: moderate (multi-source)
Significance for the KB
- First source to put concrete co-design market shares (Broadcom 60% / Marvell 35%) on the named-but-undetailed Broadcom/Marvell partnership the TSMC entity page already references.
- Supplies Ironwood's chip-level specs — previously the TPU v7 row in the concept table had a focus label but no numbers.
- Adds OpenAI Titan as a new named program not previously in the KB.
- Justifies creating entity stubs for Broadcom and Marvell as the merchant ASIC layer.
Caveats
- Body not directly fetched — figures are summary-derived; corroborate the 60/35 split and the $11B figure against Marvell/Broadcom investor materials before using in a flagship piece.
- "Ranked share" of a co-design market is an estimate; methodologies differ across analysts.