Microsoft Maia 300: primary inference chip, multi-GW from 2027

Microsoft said on 12 October that its own Maia chips will become the primary inference system for its products from 2027, at multi-gigawatt scale, and the second-order effect is that more of Microsoft's serving bill starts to shift from merchant GPUs to TSMC N2P wafers, HBM4 stacks and 167 kW racks, while model training stays with Nvidia and AMD.

What happened?

• The announcement: On 12 October 2026 Scott Guthrie, Microsoft's executive vice president for Cloud and AI, wrote on the Official Microsoft Blog that Microsoft will scale its Azure Maia family to multi-gigawatt deployments starting in 2027, "making Maia the primary inference system for Microsoft products and services".
• The new chip: Maia 300 is the third generation. Microsoft expects it to deliver 40% more tokens per megawatt than leading GPU-based systems. A full rack of 96 accelerators delivers up to seven times the rack-scale token throughput of Maia 200. Production deployments begin in 2027.
• The hardware: Each Maia 300 carries eight HBM4 stacks with 18.0 TB/s of peak memory bandwidth inside a 950-watt TDP. The compute chiplets are made on TSMC's N2P process. A rack holds 96 accelerators in a 167-kilowatt power envelope, with 27.6 TB of HBM and 1.7 PB/s of aggregate HBM bandwidth.
• The current chip: Maia 200 already serves Microsoft Copilot and is deploying across the United States, Europe and Australia. Microsoft says it delivers up to 1.7 times more tokens per watt than leading GPU infrastructure for the built-in AI in Teams, Copilot and Cowork. GitHub Copilot and Microsoft Foundry are to follow by the end of the year.

Why does it matter?

The direct signal. Until now Maia ran alongside a fleet in which, as CNBC puts it, Nvidia GPUs fill Microsoft's existing data centers. Now Microsoft is naming Maia the default for serving answers, the part of AI that grows with every Copilot user and every agent task. CNBC reports that Alistair Speirs, a Microsoft general manager, declined to give a timeline for the multi-gigawatt build and said Microsoft keeps "a little bit of optionality". He also said Microsoft still partners with Nvidia and AMD for training models. So this is a split by job, not an exit from GPUs. If you need the difference between the two jobs, see our explainer on AI training versus inference.

Our arithmetic on the rack.

• 167 kW over 96 accelerators is about 1.74 kW per accelerator slot. The chip's own TDP is 950 W, so a bit under half of each slot's power goes to host processors, switches, memory, storage and cooling inside the rack.
• 27.6 TB of HBM over 96 chips is about 288 GB each, and 1.7 PB/s over 96 is about 17.7 TB/s each. Both match the per-chip figures in Microsoft's deep dive, so the rack numbers add up.
• One gigawatt of rack power is about 6,000 Maia 300 racks, or roughly 575,000 accelerators and 4.6 million HBM4 stacks. Microsoft has not said whether "multi-gigawatt" means rack power or whole-site power. If it means the whole site, the counts are lower.
• 40% more tokens per megawatt means the same token volume needs about 29% less power. That holds only if the claim survives real workloads; it is Microsoft's expectation, not a measured result.

The second-order signal: the bind moves down the stack. A hyperscaler that designs its own inference chip stops competing for finished GPU systems and starts competing for their parts. Maia 300 needs leading-edge N2P wafers at TSMC, a node family other chip designers also compete for. It needs eight HBM4 stacks per chip, and Microsoft has not named its HBM supplier; the field is SK hynix, Samsung and Micron, as covered in our Micron profile. And it needs halls that can take 167 kW racks with liquid cooling, which Microsoft's deep dive says removes more than 80% of the rack's heat. Microsoft's own pitch admits that memory is the limit: it calls data movement, not raw compute, "increasingly the bottleneck" in large inference systems.

What does it change?

Near term (to the end of 2026)

• Maia 200 takes more of Microsoft's own traffic. GitHub Copilot and Microsoft Foundry are due on Maia 200 by year-end. Each workload that moves is inference capacity Microsoft no longer has to buy from a GPU vendor.
• A first outside estimate. CNBC cites an Evercore ISI estimate that Maia will be about 10% of Microsoft's AI chip spending by the first quarter of 2027. That is an analyst estimate, not company guidance, but it shows the starting share is small.

Longer term (2027 and after)

• Fewer halls per token. If the 40% claim holds, each new megawatt Microsoft adds serves more tokens. That eases Microsoft's own need for power, but it does not shrink total demand for powered data center space, because Microsoft is planning for multiple gigawatts.
• Larger fabrics. Microsoft's deep dive describes a superpod of 12,288 accelerators, which is 128 racks of 96. On our arithmetic that is about 21 MW of rack power for one superpod, a size that has to be planned into a hall from the start.
• Less of Microsoft's spending goes to merchant inference. This matters for how to read the lines in our guide to hyperscaler capex lines to watch. More of Microsoft's AI chip money would go to its own designs, TSMC and memory, and less to finished GPU servers for serving. Microsoft does not split that out today. For the wider picture, see what Microsoft does in our Hyperscalers layer.

What should we watch next?

• Microsoft Ignite, 17 to 20 November 2026, San Francisco. Microsoft says it will share more on its strategy there. A Maia deployment count, a region list or a power figure would firm up "multi-gigawatt".
• GitHub Copilot and Microsoft Foundry on Maia 200 by 31 December 2026. This is the first dated proof of the shift. The Verge wrote that Maia 200 already runs both; Microsoft's own blog says they follow by end of year, and we go with Microsoft's wording.
• Maia 300 production deployments in 2027. Microsoft has given a year, not a quarter. The first region and the first workload named will be the real start date.
• TSMC's third-quarter call on 15 October. Any comment on N2-family demand from cloud customers is the supply check on the node Maia 300 needs.
• What would weaken the catalyst: a slip in Maia 300 to late 2027 or beyond, the 40% figure not holding in independent tests (an analyst quoted by CNBC said nobody outside Microsoft has tested the chips yet), or Microsoft adding GPU inference capacity faster than Maia.

FAQ

Is Microsoft dropping Nvidia?
No. Microsoft says Maia becomes the primary system for inference, which means serving answers. CNBC reports that Microsoft still partners with Nvidia and AMD for training models, and Nvidia GPUs fill its existing data centers.

Can other companies rent Maia chips?
Not as reported. CNBC says that, unlike Google's TPUs or Amazon's Trainium, Microsoft's accelerators cannot be rented by the hour by other companies. Maia serves Microsoft's own products and services, including Microsoft Foundry.

When does Maia 300 actually arrive?
Microsoft says production deployments begin in 2027 and has not given a quarter. Maia 200 is already in production in the US, Europe and Australia.

Why does the 40% figure matter if power is the limit?
Because a data center hall has a fixed power supply. If each megawatt serves 40% more tokens, the same hall serves more users. On our arithmetic, a fixed token load would need about 29% less power, if Microsoft's expectation holds.

Takeaway: Microsoft's move to its own inference chips eases its reliance on merchant GPUs for serving, but it moves the bind to TSMC N2P wafers, HBM4 supply and power-dense halls, and the 2027 start has no quarter yet.

Figures come from the dated sources listed under References below. Our own arithmetic is shown and labelled as ours. Where reports disagree, we say so.

Last reviewed: 13 Oct 2026 · Next review: after Microsoft Ignite (17 to 20 Nov 2026) or when Microsoft names the first Maia 300 region
Spot an error? Tell us and we will correct it and note the change here.
Our editorial standards →

References

1. Official Microsoft Blog, Scott Guthrie, "Microsoft to expand Azure Maia system to multi-gigawatt deployments", 12 Oct 2026
2. Microsoft Azure Blog, Rani Borkar and Saurabh Dighe, "Deep dive into Azure Maia 300 architecture", 12 Oct 2026
3. CNBC, Jordan Novet, "Microsoft to rely more on homegrown AI chips, reducing dependence on Nvidia", 12 Oct 2026
4. The Verge, Tom Warren, "Microsoft's new AI chip makes it less reliant on Nvidia's GPUs", 12 Oct 2026
5. Microsoft Security Blog, "Secure what's next: Your guide to Microsoft Security at Microsoft Ignite 2026" (event dates 17 to 20 Nov 2026), 30 Sep 2026
6. Business Wire, "Tower Semiconductor and Semtech Expand SiGe Partnership to Scale Power-Efficient Analog Connectivity for Data Centers", 13 Oct 2026
7. FD (Het Financieele Dagblad), "Eerste klant voor groot AI-datacenter bij Rotterdam – nu gaat de klok tikken", 12 Oct 2026
8. WFYI Public Media, "Indiana approves AES plan to power Google data center in Monrovia", 12 Oct 2026
9. TSMC Investor Relations, "2026 Q3 Quarterly Results" (conference notice), checked 13 Oct 2026