Entity — Hyperscaler Custom ASICs (Trainium / Maia / MTIA)
Status: active Owner: Finance / Charlie AGT-002 Entity ID: ENTITY-AI-CUSTOM-ASIC Visibility: PUBLIC Last verified: 2026-07-10
Snapshot
The non-Google hyperscaler ASIC programs: AWS Trainium (Annapurna Labs, training+inference, sold as EC2 capacity), Microsoft Maia (inference-first, internal Azure/OpenAI serving), and Meta MTIA (internal ranking/recommendation inference, with an aggressive multi-generation roadmap). All three reached a new maturity step across 2025-26: Trainium3 GA (Dec 2025, 2.52 PF FP8/chip), Maia 200 deployed (Jan 2026, >10 PF FP4/chip), and Meta's MTIA 300-500 roadmap disclosure (Mar 2026). Broadcom and Marvell are the main co-design/ merchant-silicon beneficiaries; Broadcom discloses the aggregate financial signal (Q1 FY2026 AI revenue $8.4B, +106% YoY, $73B reported AI backlog).
Chip Table
| Chip | Status | Key spec highlights | As-of | Source | Grade |
|---|---|---|---|---|---|
| AWS Trainium3 | shipping (GA 2025-12-02) | 2.52 PF FP8 (MXFP8) per chip; 144 GB HBM3e at 4.9 TB/s; 3nm-class (press); Trn3 UltraServer to 144 chips (362 PF FP8, 20.7 TB HBM3e) via NeuronSwitch-v1 | 2026-07-10 | https://aws.amazon.com/about-aws/whats-new/2025/12/amazon-ec2-trn3-ultraservers/ | 🟢 |
| AWS Trainium2 | shipping | 96 GB HBM per chip (secondary); Trn2 UltraServer 64 chips via NeuronLink; AWS quotes Trn3 = 2x Trn2 MXFP8 throughput rather than a Trn2 sheet | 2026-07-10 | https://aws.amazon.com/ai/machine-learning/trainium/ | 🟡 |
| AWS Trainium4 | announced (re:Invent 2025 teaser) | vendor targets vs Trn3: >=6x FP4, 3x FP8, 4x memory bandwidth; no absolute specs | 2026-07-10 | https://nand-research.com/research-note-aws-releases-trainium3-teases-trainium4/ | 🟡 |
| Microsoft Maia 200 | shipping (deployed 2026-01, Azure US Central; US West 3 next) | TSMC 3nm, >140B transistors; >10 PF FP4, >5 PF FP8 (basis unstated); 216 GB HBM3e at 7 TB/s + 272 MB SRAM; 750 W; Ethernet-based scale-up 2.8 TB/s per chip, clusters to 6,144; serves OpenAI GPT-5.x and Copilot workloads | 2026-07-10 | https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/ | 🟢 |
| Meta MTIA v2 | shipping (internal, 16 regions) | TSMC 5nm, 1.35 GHz, 90 W; 354 TF dense INT8 / 177 TF dense BF16 (708/354 sparse); 256 MB SRAM at 2.7 TB/s + 128 GB LPDDR5; ranking/recommendation inference | 2026-07-10 | https://ai.meta.com/blog/next-generation-meta-training-inference-accelerator-AI-MTIA/ | 🟢 |
| Meta MTIA 300-500 | announced (2026-03 roadmap; deployment through 2027) | four generations disclosed; 300-series reportedly moves to 3nm-class with CoWoS; no official spec sheets | 2026-07-10 | https://www.tomshardware.com/tech-industry/semiconductors/custom-ai-asics-examined-from-broadcom-to-mtia | 🟡 |
Technical Trajectory
- Cadence: all three programs are now on ~annual-to-18-month cadence with public roadmap talk (Trainium4 teased, MTIA 300-500 disclosed) — custom silicon has moved from experiment to committed multi-generation programs.
- Specialization split: Maia 200 and MTIA are inference-first (FP4/INT8 headline numbers, high SRAM); Trainium keeps a training+inference dual mandate and is the only one sold as raw EC2 capacity at scale — the cleanest external price signal (A4) among the three.
- Interconnect direction: divergence from NVLink-style proprietary fabrics — Maia 200 uses Ethernet-based two-tier scale-up (🟢); AWS uses proprietary NeuronLink/NeuronSwitch; scale-up openness (UALink/Ultra Ethernet) is a live battleground.
- Memory: Trainium3 (144 GB) and Maia 200 (216 GB) are HBM3e-class; MTIA v2 deliberately uses LPDDR5 + big SRAM for cost-per-inference on ranking models — not LLM-comparable; do not put MTIA in LLM perf tables.
- Co-design disclosure: Google TPU and Meta MTIA are widely reported Broadcom co-designs; Trainium is in-house Annapurna (with reported Marvell and Alchip involvement across generations — 🟡, not vendor-confirmed); Microsoft does not name a Maia co-design partner (🟢 blog names only TSMC). Financial visibility is via Broadcom earnings: Q1 FY2026 AI semiconductor revenue $8.4B (+106% YoY), reported $73B AI backlog and "line of sight" to $100B 2027 AI revenue (🟡 via press of earnings call).
Watch Items
- Broadcom quarterly AI revenue + backlog updates (next earnings ~2026-09) — the aggregate custom-ASIC throughput proxy (A3-adjacent).
- Marvell custom-silicon disclosures (which hyperscaler programs, revenue ramp timing).
- Trainium3 rental pricing and capacity signals on EC2 (A4); Project Rainier-class Anthropic cluster utilization disclosures.
- Maia 200 rollout beyond two regions; any external (non-internal) Azure offering.
- Meta MTIA 300 first deployment; whether Meta extends MTIA from ranking to generative inference/training as roadmapped.
- MLPerf: none of Trainium/Maia/MTIA submit closed-division results — any first submission would be a major comparability event.
Claims
- CLAIM-AI-SILICON-001 — primary evidence stream alongside google-tpu.md: three committed multi-generation ASIC programs + Broadcom's $73B reported backlog are the quantitative case that custom-ASIC share of accelerator compute is rising.
- CLAIM-AI-BOTTLENECK-001 — custom ASICs compete for the same HBM3e/CoWoS supply as merchant GPUs; their ramp shifts where the bottleneck binds.
Sources
- 🟢 AWS What's New, Trn3 UltraServers GA — https://aws.amazon.com/about-aws/whats-new/2025/12/amazon-ec2-trn3-ultraservers/ (accessed 2026-07-10)
- 🟢 AWS Trainium product page — https://aws.amazon.com/ai/machine-learning/trainium/ (accessed 2026-07-10)
- 🟢 Microsoft Official Blog, "Maia 200: The AI accelerator built for inference" — https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/ (accessed 2026-07-10)
- 🟢 Meta AI blog, next-gen MTIA — https://ai.meta.com/blog/next-generation-meta-training-inference-accelerator-AI-MTIA/ (accessed 2026-07-10)
- 🟡 Tom's Hardware, custom AI ASIC state of play (May 2026; incl. Broadcom Q1 FY2026 figures and MTIA roadmap) — https://www.tomshardware.com/tech-industry/semiconductors/custom-ai-asics-examined-from-broadcom-to-mtia (accessed 2026-07-10)
- 🟡 NAND Research, Trainium3 release / Trainium4 tease — https://nand-research.com/research-note-aws-releases-trainium3-teases-trainium4/ (accessed 2026-07-10)
- 🟡 The Next Platform, Maia 200 analysis — https://www.nextplatform.com/ai/2026/01/28/microsoft-takes-on-other-clouds-with-braga-maia-200-ai-compute-engines/4092134 (accessed 2026-07-10)
Changelog
- 2026-07-10: Created; first verification pass.