Microsoft bets on its own silicon: Azure Maia to scale to multi-gigawatt deployments from 2027 as its primary inference system; Maia 300 claims 40% more tokens per megawatt than leading GPUs

Why it matters

【Key points】

On Oct 12 Microsoft said it will scale its in-house Azure Maia chips to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for its own products and services. Maia 200 already serves Microsoft Copilot and the MAI team, with GitHub Copilot and Microsoft Foundry to follow by year end, running OpenAI, MAI and open-source models. Maia 300, unveiled the same day, enters production in 2027 and targets long-context reasoning, low-latency chat and long multi-step agent work. Satya Nadella called it Microsoft's biggest bet yet on its own silicon.

【Verifiable】

Blog (Scott Guthrie): https://blogs.microsoft.com/blog/2026/10/12/microsoft-to-expand-azure-maia-system-to-multi-gigawatt-deployments/

Nadella: https://x.com/satyanadella/status/2109631106404487523

Maia 300 per the blog: eight HBM4 stacks, 18.0 TB/s peak bandwidth at 950 W; compute chiplets on TSMC N2P; enhanced MXFP4; 96 accelerators per 167 kW rack with 27.6 TB HBM. Microsoft expects 40% more tokens per megawatt than leading GPU systems and up to 7x Maia 200's rack throughput; Maia 200 gives up to 1.7x tokens per watt versus leading GPUs on built-in Teams and Copilot features.

【Our read】

1. The news is the role, not the specs: one of the largest AI service providers is moving the bulk of its own inference off GPUs, already running OpenAI models. The post covers inference only, not training.

2. The 40% and 7x figures are Microsoft's own projections with no named GPU or workload; wait for the technical paper and independent numbers.

3. Google's TPUs, Amazon's Trainium and now Maia bet on the same thing: inference is becoming the main AI cost, and more tokens per megawatt means lower prices.