Microsoft's Maia 300 explained: why its own AI chip matters more after Azure's $33 billion quarter

This week Microsoft disclosed Azure's quarterly revenue for the first time — $33.2 billion, up 48%. The fastest way to protect a margin on that much AI traffic is to stop paying someone else for every chip. That's the job of Maia 300.

What Maia 300 is

Microsoft's third-generation in-house accelerator, announced October 12, is inference-first: designed for long-context, memory-heavy, decode-heavy workloads — the kind generated by reasoning models and agents that think for a long time and write long answers.

Key specs and claims (Microsoft's figures, not independently verified):

• 288GB of HBM4 memory per accelerator.
• Chiplet design on TSMC's N2P process.
• 96-accelerator, liquid-cooled racks.
• 40% more tokens per megawatt than leading GPU-based systems.
• Up to 7x the rack-scale token throughput of Maia 200.

Why "tokens per megawatt" is the number to watch

AI data centers are increasingly limited by power, not floor space or even chip supply. If a given building can only draw so many megawatts, the operator that squeezes more answers out of each one wins on cost. Framing the chip around tokens per megawatt — rather than raw FLOPS — tells you Microsoft is optimizing for serving Copilot and Azure AI cheaply at enormous scale, not for winning training benchmarks.

The roadmap

• Maia 200 is already in production across the US, Europe and Australia, powering Copilot, Teams and other Microsoft services. By the end of 2026 it's expected to run GitHub Copilot and Microsoft Foundry workloads.
• Maia 300 goes into production deployment in 2027.
• Microsoft plans to scale the Maia family to multi-gigawatt deployments in 2027, positioning it as a primary inference system for its own products.

You can't rent one

Unlike AWS (Trainium, Inferentia) or Google Cloud (TPUs), Microsoft hasn't offered Maia as rentable Azure instances and hasn't said whether it will. Analyst Matt Kimball of Moor Insights & Strategy frames this as a services play rather than a chip war: customers get faster, more reliable AI services; Microsoft gets lower costs. He describes Maia 100 as a proof of concept, Maia 200 as the validation, and Maia 300 as the point where Microsoft decided to go big.

What it means for Nvidia

Not much in the short term. Microsoft still relies heavily on Nvidia and AMD GPUs, especially for training. But every hyperscaler moving its highest-volume inference onto in-house silicon caps how much of the AI boom flows to merchant chipmakers over time. With investors now scrutinizing the return on AI capex — see Alphabet's $500 billion bad day — cheaper inference is no longer a nice-to-have. It's the business case.