Hyperscaler Custom Silicon: TPU, Trainium, Maia, MTIA — 2026-10-02

Top developments

Marvell's Google partnership reshapes design-services consolidation

Marvell announced a significant custom AI silicon deal with Google, reinforcing its position as co-designer for multiple hyperscalers—Amazon (Trainium), Microsoft (Maia), and now Google. The deal underscores why Broadcom and Marvell together control ~95% of the custom ASIC co-design market, with Broadcom alone holding 70%+ market share. Google spends approximately $8 billion annually with Broadcom on TPU development, while Marvell projects up to $11 billion in AI ASIC revenue for 2026.

AWS Trainium 3 shipments encounter supply-chain friction

AWS's latest Trainium 3 shipments are running behind expectations due to parts shortages, assembly complexity, and rapid demand growth across the global AI hardware supply chain. Despite delays, Trainium remains central to Anthropic's multi-cloud strategy, complementing its massive Google TPU commitments.

Anthropic's $110 billion AWS and Google multi-gigawatt compute lock-in

Anthropic's draft IPO prospectus reveals approximately $110 billion owed to Amazon Web Services over roughly a decade, making AWS effectively both landlord and shareholder in the AI lab's infrastructure. The company simultaneously committed to up to 1 million Google TPUs, cementing Anthropic's position as the largest single customer validating both custom silicon platforms.

1 sources

constellationr.com

Google Cloud

Custom ASIC market share projected to reach 28% in 2026

Analyst projections place custom AI chips at nearly 28% of total data center accelerator shipments in 2026, marking a structural shift away from GPU-only infrastructure as every major hyperscaler deploys proprietary silicon. Marvell's 2025 custom ASIC revenue reached approximately $1.5 billion, with expansion targets for 2026.

Alibaba accelerates custom silicon roadmap with Zhenwu V900

Alibaba's T-Head subsidiary unveiled the Zhenwu V900 AI accelerator targeting 1Q27 mass production, positioning the Chinese hyperscaler as a significant alternative to Western custom silicon platforms. Alibaba is emerging as a custom silicon powerhouse alongside ByteDance efforts, signaling geographic diversification in hyperscaler chip sovereignty.

Context & numbers

Market concentration: Broadcom and Marvell control ~95% of custom ASIC co-design services. Broadcom holds 70%+ market share; Google alone allocates ~$8B annually to Broadcom for TPU development.

Revenue projections: Marvell targets up to $11 billion in AI ASIC revenue for 2026, up from ~$1.5 billion in 2025.

Market share expansion: Custom AI chips expected to reach 28% of data center accelerator shipments in 2026 versus GPU-dominated infrastructure.

Customer commitments: Anthropic committed ~$110 billion to AWS over ~10 years and up to 1 million Google TPUs by 2027; AWS deployed 500,000+ Trainium chips as of May 2026.

Anthropic operating metrics (from IPO draft): $4.6B revenue in 2025 with $8B operating loss, indicating infrastructure costs remain central to AI lab scaling economics.

On the radar

• AWS Trainium 3 yield recovery timeline: Monitor Q4 2026 supplier updates for resolution of assembly bottlenecks and memory chip shortages affecting Trainium 3 ramp. Supply constraints may create margin expansion opportunity if resolved before competitor chips scale.

• Anthropic IPO prospectus filings: Watch for detailed capex forecasts and unit economics tied to TPU vs. Trainium utilization rates—will signal which custom chip platform leadership favors for inference vs. training workloads.

• Alibaba Zhenwu V900 adoption signals: Track early customer announcements (Alibaba Cloud, ByteDance, Tencent) post-1Q27 production to assess whether Chinese custom silicon gains traction outside domestic markets.

• Broadcom 2027 design-services guidance: October-November earnings reports may reveal pipeline growth or saturation signals for next-generation TPU v8 and Trainium 4 co-design contracts.