The third-generation custom inference chip is expected to roll out at scale in 2027 as Microsoft works to lean less on Nvidia
Microsoft wants to run more of its AI on chips it designed itself. On October 12, 2026, the company unveiled the Azure Maia 300, its third-generation custom AI inference chip.
The plan is to deploy it at scale in Azure data centers beginning in 2027. That puts Microsoft on a path toward owning more of the hardware behind Copilot.
Microsoft CEO Satya Nadella isn’t underselling it.
“Our biggest bet yet on our own silicon.”
What the Maia 300 is built to do
The Maia 300 is built for inference workloads such as conversational AI and long-context reasoning, where a model has to digest large amounts of text before responding.
Under the hood, the chip uses a chiplet architecture built on TSMC’s N2P process node.
The Maia 300 is reported to deliver 40% more tokens per megawatt than current leading GPU solutions and provides up to 7x the throughput of its predecessor, the Maia 200.
Where Maia fits in Microsoft’s lineup
The Maia 200 is already operational in data centers across the US, Europe, and Australia, supporting Microsoft 365 Copilot and GitHub Copilot.
The Maia 300 is meant to go further. Microsoft aims for it to serve as the core inference system for Copilot experiences, GitHub Copilot, and various models built by OpenAI and Microsoft itself.
The longer-term goal is for Maia to become the primary AI inference system across Microsoft’s products and services. If that happens, the company expects margins to improve over time.
Microsoft’s capital expenditure plans exceed $175 billion annually, and Maia is expected to account for only around 10% of Microsoft’s AI chip spending by early 2027.
Why Microsoft is building its own chips
Custom silicon offers a way to trim that bill. Estimates suggest the shift toward Maia could produce multi-billion-dollar savings, though those benefits are expected to build slowly rather than arrive all at once.
What this means for Nvidia, AMD, and the chip market
Nvidia and AMD aren’t getting written out of the story. Microsoft continues to invest in third-party solutions from both, and those partnerships are expected to remain critical, especially for AI training.
The Maia 300’s reliance on TSMC’s N2P node shows the foundry stays in the picture whether AI chips come from Nvidia or from a hyperscaler’s in-house team.
The things to watch are concrete: how quickly Maia 300 deployments ramp in 2027, whether the chip’s share of Microsoft’s AI spending climbs beyond the expected 10%, and whether margin improvements start showing up in the company’s results.
