Hyperscaler Custom Silicon: TPU, Trainium, Maia, MTIA — 2026-10-23

Top developments

Google and Amazon Open PyTorch to Custom Chips

Google released TorchTPU, an open-source PyTorch backend that allows developers to migrate AI models from Nvidia GPUs to Cloud TPUs with minimal code changes, achieving 52% hardware efficiency across 8,192 chips. Amazon simultaneously launched TorchNeuron for its Trainium 2 and 3 chips, aiming to make its silicon viable for research training beyond just inference. These moves directly attack Nvidia’s software dominance by lowering the barrier to entry for hyperscaler custom silicon.

ByteDance Negotiates Custom ASIC Deal with MediaTek

ByteDance is in advanced negotiations to purchase custom AI accelerators from MediaTek for its data centers, potentially becoming MediaTek’s second major cloud service provider client after Google. If finalized, this deal would expand the ASIC ecosystem beyond the traditional Broadcom-Marvell duet, driven by ByteDance’s need to reduce reliance on Nvidia GPUs amid supply constraints.

Broadcom Secures $50 Billion for OpenAI Custom Chips

Broadcom has arranged over $50 billion in financing to support the production of custom AI chips for OpenAI, marking a significant expansion of ASIC demand beyond traditional hyperscalers. This deal underscores the growing trend of AI labs designing their own silicon to optimize cost and performance, further diversifying the AI chip supply chain.

Microsoft Maia Chips Focus on Inference

Microsoft plans to primarily use its Maia 300 chips for AI inference workloads to expand margins and reduce dependence on Nvidia. While analysts note that Maia chips are not yet performance-competitive with Google’s TPUs or Nvidia’s latest GPUs, Microsoft aims to scale Maia deployments to multi-gigawatt levels starting next calendar year.

Local view

In China, media reports highlight a divergence between nationalistic "domestic substitution" rhetoric and actual purchasing behavior. Despite government pushes for Huawei Ascend chips, ByteDance, Alibaba, and Tencent have reportedly ordered over $35 billion worth of Nvidia RTX PRO 5500 chips, indicating that performance and availability still outweigh policy mandates. Meanwhile, ByteDance’s parallel move to negotiate with MediaTek for ASICs shows a dual strategy of buying Nvidia while building custom alternatives.

Context & numbers

• Custom ASIC Market Share: Custom ASICs are projected to account for 27.8% of AI server shipments in 2026, with shipment growth at 44.6% year-over-year, far outpacing the 16.1% growth of merchant GPUs.
• Broadcom Revenue: Broadcom’s AI semiconductor revenue reached $16.7 billion in fiscal Q3 2026, up 221% year-over-year, with an estimated $73 billion backlog in AI-related contracts.
• Marvell Projections: Marvell, which partners with Amazon on Trainium and Microsoft on Maia, projects up to $11 billion in AI ASIC revenue for 2026.
• Nvidia Orders in China: ByteDance, Alibaba, and Tencent have ordered over 3 million Nvidia RTX PRO 5500 chips, valued at more than $35 billion.

On the radar

• Alphabet Q3 Earnings: Investors are watching Alphabet’s upcoming earnings for details on the $100 billion TPU opportunity and cloud revenue growth, as custom silicon becomes a key competitive lever.
• Anthropic’s Multi-Cloud Strategy: Anthropic continues to balance massive commitments to Google TPUs (up to 1 million chips) with AWS Trainium deployments, signaling a long-term multi-vendor approach to custom silicon.