Microsoft says Maia 200 is in production and announces Maia 300 inference system, with multi-gigawatt Maia deployments starting 2027

Microsoft says its Maia 200 inference chip is now serving production workloads, including Microsoft Copilot and the MAI team's models, with GitHub Copilot and Foundry due by the end of 2026. Maia 200 is deploying in the US, Europe and Australia. Microsoft also unveiled Maia 300: TSMC N2P compute chiplets, 8 HBM4 memory stacks, 18.0 TB/s bandwidth and 950 W per chip. A 96-accelerator rack draws 167 kW and holds 27.6 TB of HBM. Maia 300 production deployments begin in 2027, and Microsoft plans multi-gigawatt Maia deployments from 2027 as its primary inference system.

If delivered, this would cut Microsoft's reliance on GPU suppliers for running its own AI products, and could lower its cost per query. That matters to Nvidia and AMD as competitors and to investors tracking Microsoft's capital spending. For enterprise customers and developers the effect is indirect for now. Nothing here changes what you can buy or the price you pay today, and Microsoft has not said Maia capacity will be offered to outside customers on the same terms. Maia 300 is a 2027 promise, so near-term decisions should not hinge on it.

The headline efficiency numbers (40% more tokens per megawatt than leading GPU systems, 1.7x tokens per watt, up to 7x rack throughput) are Microsoft's own, with no named comparison system, workload or independent test. "Multi-gigawatt" is a stated intention with no volumes, dates or budget, and the "era of Super Intelligence" framing is marketing. Maia 200 "in production" also does not say what share of Microsoft's inference it actually carries.

1. 01Independent or detailed benchmarks, or Microsoft disclosing the comparison system and workload behind the 40% tokens-per-megawatt and 1.7x tokens-per-watt claims
2. 02Whether GitHub Copilot and Foundry actually move onto Maia 200 by the end of 2026, and whether Azure offers Maia capacity to external customers
3. 03Microsoft's next earnings and capex commentary, plus any Nvidia or AMD purchase changes, to show whether Maia is displacing GPU spend or merely adding to it, and whether Maia 300 holds its 2027 schedule on TSMC N2P and HBM4 supply