Hyperscaler Custom Silicon: TPU, Trainium, Maia, MTIA — 2026-09-04
Microsoft is reportedly accelerating its custom silicon roadmap with a September launch target for the Maia 300 chip and a significant TSMC capacity reservation. Meanwhile, Anthropic continues to diversify its compute strategy away from Nvidia exclusivity, signing massive multi-cloud deals that leverage both AWS Trainium and Google TPU capacity.
Hyperscaler Custom Silicon: TPU, Trainium, Maia, MTIA — 2026-09-04
Microsoft Targets September Launch for Maia 300
Microsoft is reportedly aiming to launch its third-generation Maia AI accelerator in September 2026. The company has secured approximately 300,000 wafers of TSMC capacity to support this rollout, signaling an aggressive push to catch up with Google’s TPU and AWS’s Trainium deployments. This move highlights the intensifying competition among hyperscalers to reduce reliance on Nvidia GPUs for internal workloads and potentially offer competitive cloud pricing.
Anthropic Locks In Massive Multi-Cloud Compute Deals
Anthropic has signed a $45 billion infrastructure deal with Nscale and a separate $35 billion agreement with an Nvidia-backed startup, further diversifying its compute sources. These agreements complement its existing massive commitments to Google Cloud (up to 1 million TPUs) and AWS (Trainium), illustrating a strategic shift toward a "multi-chip" approach. By leveraging custom silicon like TPUs and Trainium alongside Nvidia GPUs, Anthropic aims to optimize cost and performance while mitigating supply chain risks.
OpenAI’s Jalapeño Chip Beats Blackwell in Efficiency Tests
OpenAI’s custom "Jalapeño" AI chip has reportedly outperformed Nvidia’s Blackwell systems in key inference-efficiency benchmarks. This development marks a significant milestone in the custom silicon race, as OpenAI joins Google, Amazon, and Meta in deploying proprietary hardware. The results suggest that specialized ASICs are becoming increasingly viable alternatives to general-purpose GPUs for specific high-volume inference tasks, threatening Nvidia's dominance in the inference market.
Nvidia Acquires Hugging Face to Counter Custom Silicon Threat
Nvidia announced the acquisition of Hugging Face for approximately $12.9 billion, a move analysts interpret as a strategic hedge against the rise of hyperscaler custom silicon. By securing the largest open-model developer platform, Nvidia aims to maintain its ecosystem lock-in as major customers like Microsoft and Meta build their own chips. This acquisition underscores Nvidia's recognition that software and model distribution are critical defenses against the commoditization of AI hardware by custom ASICs.
Local view
Japan: AWS Strengthens Trainium-Nvidia Synergy
Japanese media reports that AWS plans to deploy over 3 million Nvidia GPUs through 2028, while simultaneously strengthening the integration between these GPUs and its proprietary Trainium chips. The coverage highlights AWS's dual-track strategy: using Nvidia for general-purpose flexibility and Trainium for cost-efficient, large-scale training. This reflects a broader trend among Japanese enterprises adopting hybrid AI infrastructure strategies.
Context & numbers
- Custom Silicon Market Share: Broadcom and Marvell control approximately 95% of the custom ASIC co-design market for hyperscalers.
- Broadcom Dominance: Broadcom holds over 70% of the custom AI accelerator design services market, up from earlier estimates, driven by wins with Google and Meta.
- Marvell Revenue Projection: Marvell projects up to $11 billion in AI ASIC revenue for 2026, anchored by design wins with AWS (Trainium) and Microsoft (Maia).
- Hyperscaler Capex: Hyperscalers spent approximately $410 billion on capital expenditures in the recent period, with a significant portion directed toward custom silicon and data center infrastructure.
On the radar
- Maia 300 Launch: Keep an eye on official Microsoft announcements regarding the Maia 300 in late September 2026.
- Trainium 4 Availability: AWS Trainium 4 is scheduled for late 2026 or early 2027 availability, promising 3x FP8 performance over Trainium 3.
This content was collected, curated, and summarized entirely by AI — including how and what to gather. It may contain inaccuracies. Crew does not guarantee the accuracy of any information presented here. Always verify facts on your own before acting on them. Crew assumes no legal liability for any consequences arising from reliance on this content.
Powered by
