Microsoft unveils Maia 300 AI chip, its main inference system from 2027
Microsoft's third-generation Azure Maia chip promises 40% more tokens per megawatt than leading GPU systems and up to 7x Maia 200's rack throughput.
Microsoft has announced Maia 300, the third generation of its in-house Azure Maia AI accelerator, and says the Maia family will become the primary inference system for Microsoft products and services, scaling to multi-gigawatt deployments starting in 2027.
What Microsoft is promising
The announcement came from Scott Guthrie, who runs Microsoft's Cloud + AI group. The key claims:
- 40% more tokens per megawatt than "leading GPU-based systems". Microsoft does not name a competitor, but the comparison is clearly aimed at Nvidia-based servers.
- Up to 7x the rack-scale token throughput of Maia 200 in a full 96-accelerator configuration.
- An inference-first design aimed at long-context reasoning, low-latency chat and long multi-step agent tasks.
These are Microsoft's own projections. The chip is not shipping yet, and no independent benchmarks exist.
How it is built
Microsoft describes Maia 300 as a complete system rather than just a chip: silicon, software, racks, networking and data centres are designed together. According to SiliconANGLE's report on the technical blog post, accelerators sit four to a tray in liquid-cooled racks of 96 chips. Some outlets also cite 288 GB of HBM4 memory per chip.
Maia 200 is already working
The previous chip, Maia 200, was announced in January and is now in production. Microsoft says it already runs Copilot features, hosts OpenAI models and Microsoft's own MAI models, and delivers up to 1.7x more tokens per watt for those workloads.
Why it matters
Google, Amazon and Meta already design their own AI chips, and Microsoft is now moving in the same direction at much larger scale. That does not mean Nvidia is out: Microsoft still plans to be among the first to deploy Nvidia's next-generation Vera Rubin systems. But a big customer that can serve more of its own AI traffic gains leverage, and if the efficiency numbers hold up, it could make services like Copilot and GitHub cheaper to run.
Microsoft says it will share more at Ignite 2026 in San Francisco next month.
Keep reading
Google opens Gemini 4 Argon to everyone with a 1M-token limit
Google’s frontier Gemini 4 Argon model is available today via API, Antigravity, the Gemini app and AI Mode in Search for Pro and Ultra subscribers.
Anthropic: small AI models may soon self-replicate on IoT devices
A 12 October research post shows a 26B model running across 15 cheap Android TV boxes and projects small weight-based replicators by late 2026.
Intel names its Crescent Island AI chip: GPU D-Series, with up to 480 GB of memory
Intel's inference accelerator, codenamed Crescent Island, is now the GPU D-Series. The first model, the D270, uses Xe3P and is rated at 1.3 PFLOPS of FP8 within an air-cooled 350 W card.