Microsoft names Maia 300 its primary inference chip and plans multi-gigawatt Azure deployments from 2027
The AMW Read
Expected next step for a known hyperscaler chip program, but naming custom silicon the primary inference platform at multi-gigawatt scale changes how much inference demand reaches merchant accelerators.
Microsoft names Maia 300 its primary inference chip and plans multi-gigawatt Azure deployments from 2027
Microsoft has announced Maia 300, the next generation of its custom AI inference platform. It plans to expand Azure Maia to multi-gigawatt deployments starting in 2027, and Maia 300 will become the primary inference system for Microsoft products and services. Microsoft says the system targets long-context reasoning, latency-sensitive conversational use and sustained multi-step agentic tasks. Projected gains are 40% more tokens per megawatt and up to seven times the rack-scale token throughput of Maia 200. Microsoft says Maia 200 already delivers up to 1.7 times more tokens per watt for built-in AI features.
The key point is the word primary. Until now, hyperscaler custom silicon has been something extra that ran next to merchant accelerators. Microsoft is now committing its own first-party inference traffic to an in-house chip, at multi-gigawatt scale. That traffic includes Copilot, Teams, GitHub, Microsoft Foundry and Dragon Copilot. Measuring the chip in tokens per megawatt treats power, not chip count, as the limit on serving AI. This fits a wider pattern: in large data centers, energy supply increasingly sets how much inference a provider can sell. Note that all performance figures are Microsoft's own projections. The source gives no independent benchmarks and no comparison with third-party hardware.
For investors, an inference fleet that Microsoft controls shifts part of its future AI spending away from merchant GPU suppliers. Watch that shift as 2027 deployments begin. For builders on Azure and Foundry, the more immediate question is whether Microsoft's own services move to Maia first and leave other accelerators for outside customers. If so, model deployment options, price per token and latency on Azure could start to differ depending on which hardware a workload lands on.
