Microsoft Maia 300 aims to cut Azure's reliance on Nvidia GPUs
Microsoft said Maia 300, its third-generation custom AI accelerator, will become the primary inference system for company products starting with multi-gigawatt deployments in 2027. CNBC framed the move as reducing dependence on Nvidia GPUs for running models at scale.
AI-assisted · US Brief desk · Sources listed below
What happened
Microsoft on Monday said Azure Maia 200 is already in production for workloads including Microsoft Copilot, and that Maia 300 — the third-generation custom accelerator — will begin production deployments in 2027 as Microsoft scales Maia to multi-gigawatt deployments and makes it "the primary inference system for Microsoft products and services," EVP Scott Guthrie wrote.
Why it matters
CNBC framed the announcement as Microsoft leaning harder on homegrown silicon after years of filling Azure with Nvidia GPUs. CFO Amy Hood has previously tied custom chips to margin improvement; Evercore ISI told clients Maia could be about 10% of Microsoft's AI-chip spend by early 2027, CNBC reported.
More context
What Microsoft claims: Guthrie said a full 96-accelerator Maia 300 rack can deliver up to 7x the rack-scale token throughput of Maia 200 and about 40% more tokens per megawatt than leading GPU-based systems. Specs in the post include eight HBM4 stacks at 18.0 TB/s peak memory bandwidth within a 950-watt TDP, and racks packing 96 accelerators in a 167-kilowatt envelope.
Primary excerpt: "We will scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services," Guthrie wrote. Maia 300 is "designed for the new types of AI workloads emerging today: Long-context reasoning, latency-sensitive conversational experiences and sustained, multi-step agentic tasks."
What's still unknown: General manager Alistair Speirs declined to give a firm multi-gigawatt timeline, citing "optionality" as models change, CNBC reported. Unlike Google TPUs or Amazon Trainium, Maia chips are not rented by the hour to outside customers yet. Speirs said training still partners with Nvidia and AMD.
2 listed sources
References listed by US Brief; a source count is not a verification score.
Editorial sourcing notes
Architecture and deployment claims are from Microsoft EVP Scott Guthrie's Oct. 12, 2026 blog post. Competitive framing and Nadella/Hood quotes are from CNBC's same-day report. Efficiency comparisons are Microsoft's claims versus "leading GPU-based systems," not an independent benchmark.
Like US Brief? You can support it with a tip.