# Microsoft says its Maia chips will become its main inference platform at multi-gigawatt scale from 2027

The third-generation Maia 300, built on TSMC's N2P process with eight HBM4 stacks, is claimed to deliver 40% more tokens per watt than leading GPU systems. That figure comes from Microsoft and has not been independently measured.

- Published: 2026-10-13T06:13:53.502Z
- Canonical: https://polylog.news/ai/2026-10-13/microsoft-says-its-maia-chips-will-become-its-main-inference
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [Dataquest India](https://www.dqindia.com/news/microsoft-maia-ai-chips-multi-gigawatt-deployments-2027-12669825), [Interesting Engineering](https://interestingengineering.com/ai-robotics/microsoft-maia-300-96-chips-rack), [WindowsForum](https://windowsforum.com/news/microsoft-azure-maia-300-targets-multi-gigawatt-ai-inference-deployments-in-2027.448071/), [RuntimeWire](https://runtimewire.com/article/microsoft-maia-300-ai-chip-production-tsmc)

Microsoft will scale its Maia accelerator family to multi-gigawatt deployments starting in 2027, and Maia will become the primary inference system for Microsoft's own products and services, the Russian-language channel [AI ML Big Data reported](https://t.me/ai_machinelearning_big_data/11118). [Dataquest India](https://www.dqindia.com/news/microsoft-maia-ai-chips-multi-gigawatt-deployments-2027-12669825) confirmed the plan. Microsoft also unveiled Azure Maia 300, its third-generation custom accelerator, on October 12.

The disclosed specifications target serving, not training. According to [WindowsForum's summary of the launch](https://windowsforum.com/news/microsoft-azure-maia-300-targets-multi-gigawatt-ai-inference-deployments-in-2027.448071/), Maia 300 uses a TSMC N2P chiplet design with eight HBM4 stacks delivering 18.0 TB/s of memory bandwidth and improved MXFP4 support. A rack holds 96 accelerators. Microsoft claims up to 7x the rack-scale token throughput of Maia 200. [Interesting Engineering](https://interestingengineering.com/ai-robotics/microsoft-maia-300-96-chips-rack) put the rack at 167 kW. Satya Nadella, Microsoft's chief executive, said the chip should deliver 40% more tokens per watt than leading GPU-based systems. Microsoft has not named the GPU baseline or the workload behind that figure.

Maia 200 is already in production in the United States, Europe and Australia and serves Microsoft Copilot. Microsoft told investors that running OpenAI and internal models on Maia 200 cuts operating costs by 30% to 40% compared with top Nvidia hardware, [RuntimeWire reported](https://runtimewire.com/article/microsoft-maia-300-ai-chip-production-tsmc). The same outlet said Microsoft has sought TSMC capacity for more than 300,000 Maia 300 chips for 2027 delivery and wants capacity for more than one million over time.

Bandwidth and low-precision number formats set decode throughput, and Microsoft optimized for both. Eighteen TB/s per package and native MXFP4 support favor memory-bound token generation for mixture-of-experts (MoE) models with large key-value (KV) caches. That is the workload that dominates Copilot and Azure OpenAI traffic.

## What this means

Inference is the larger and more predictable share of hyperscaler AI spending, and Microsoft has now said publicly that its own silicon will carry most of it from 2027. Nvidia faces the greatest exposure: the risk is to its pricing power on inference fleets, while training demand is not directly affected. Marvell, which designs Maia according to RuntimeWire, and TSMC, which fabricates it, gain. The precedent is Google's Tensor Processing Unit (TPU), which let Google serve its own models with limited Nvidia dependence and later became an external product. Markets already expected hyperscaler custom silicon. What is new here is the "primary platform" commitment and the gigawatt scale. One assumption worth questioning is whether the 40% tokens-per-watt claim should be treated as settled. Earlier custom accelerators often slipped schedules or underdelivered against the next Nvidia generation, which here is Vera Rubin. The desk's read is that Microsoft will keep buying Nvidia for training and for Azure customers who require CUDA, while shifting first-party inference to Maia over 2027 and 2028. That depends on TSMC N2P yields and HBM4 supply. It would be wrong if Maia 300 volume slips into 2028 or if Microsoft's next Nvidia orders for inference-heavy regions do not shrink. Production of both Maia and its rivals is concentrated at TSMC in Taiwan, which remains the single-jurisdiction risk for the whole sector.

## What to watch

- Whether Microsoft publishes or allows third-party benchmarks of Maia 300 tokens per watt against a named Nvidia Rubin configuration. The 40% claim is unverified until it names a baseline.
- Microsoft's capital-spending commentary in its next earnings call, especially any split between Nvidia purchases and Maia deployment. A falling Nvidia share would confirm the shift in practice.
- TSMC N2P and HBM4 supply announcements for 2027. Either one can delay Maia volume regardless of design quality.