Hyperscaler Custom Silicon: TPU, Trainium, Maia, MTIA — 2026-10-14

Top developments

Microsoft Reveals Maia 300 with 2nm Chiplets and HBM4

On October 12, Microsoft announced the Maia 300, its next-generation custom AI inference chip designed to reduce reliance on Nvidia and AMD GPUs. The chip utilizes TSMC’s N2P (2nm) process and integrates 288GB of HBM4 memory across chiplets, aiming for production rollout in 2027. Microsoft claims the Maia 300 will deliver 40% more tokens per megawatt compared to current GPU-based systems, significantly improving power efficiency for large-scale inference workloads. This move underscores Microsoft's strategy to expand margins by shifting its massive Azure AI workload toward proprietary silicon optimized specifically for inference rather than raw training compute.

Amazon in Talks to Acquire Decart for ~$7 Billion to Boost Trainium

Amazon is reportedly in advanced negotiations to acquire Decart, an Israeli AI startup, for approximately $7 billion, as reported on October 13. Decart specializes in software that optimizes AI models to run efficiently across different hardware architectures, including Amazon’s Trainium chips, addressing the "CUDA moat" that has historically favored Nvidia GPUs. By acquiring this chip-neutral optimization layer, Amazon aims to make its Trainium accelerators more attractive to external customers who currently prefer Nvidia due to easier developer tooling and portability.

Hyperscalers Pivot to Inference-Optimized ASICs

The industry trend is increasingly shifting from general-purpose training GPUs to specialized ASICs for inference, driven by cost efficiency demands. Business Standard noted that Microsoft’s Maia 300 joins Google’s TPU, Amazon’s Trainium, and OpenAI’s custom silicon efforts as firms seek to control infrastructure costs. Axios highlighted that every major Nvidia customer is now effectively a competitor, with Big Tech leveraging custom chips to challenge Nvidia’s dominance in the data center. This shift is critical as the AI market matures, with more spending moving toward the high-volume, low-latency requirements of inference workloads.

Local view

Japanese Media Focus on Power Efficiency and Acquisition Strategy

Japanese tech outlets have closely covered the Maia 300 announcement, highlighting the specific claim of a 40% improvement in power efficiency compared to GPU systems. Unite.ai detailed how Microsoft plans to scale Azure to multi-gigawatt levels using these new chips, noting that Maia 300 builds on the efficiency gains of the previous Maia 200 generation. Regarding Amazon, Zaikei reported on the potential $7 billion Decart acquisition, emphasizing that Decart’s technology is key to optimizing Trainium for diverse AI models. XenoSpectrum further analyzed the deal, noting that Decart’s ability to optimize video AI and other models for Trainium could be a decisive factor in AWS’s competitive strategy against Nvidia.

Context & numbers

• Maia 300 Specs: Built on TSMC N2P (2nm), features 288GB HBM4, and targets production in 2027.
• Efficiency Gains: Microsoft claims 40% more tokens per megawatt vs. GPUs.
• Decart Deal Value: Reported at approximately $7 billion.
• ASIC Market Share: Broadcom holds an estimated 70%+ of the custom AI accelerator design services market, while Marvell holds roughly 20-25%.
• Capex Constraints: Broadcom CEO Hock Tan warned that capital, not chips, is the next bottleneck for AI infrastructure, with hyperscalers facing significant cash flow pressures.

On the radar

• Maia 300 Production: Monitor supply chain updates for the 2027 production start, particularly regarding TSMC N2P capacity allocation.
• Amazon-Decart Closing: Watch for official confirmation or adjustment of the $7 billion acquisition price following the WSJ report.
• Intel GPU D270: Intel recently named its new inference chip the GPU D270 (formerly Crescent Island) with 480GB memory, adding another competitor to the inference-specific ASIC market.