On October 12, 2026, Microsoft unveiled Maia 300, its third-generation in-house AI accelerator. In a blog post, Executive Vice President Scott Guthrie said 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. CNBC's headline frames this as reduced dependence on Nvidia. The fuller picture is more nuanced: this is about serving models (inference), not wholesale replacement of GPUs.
What Microsoft actually announced
The key facts, as Microsoft and press coverage describe them:
- Maia 200 is in production now. Microsoft says it is serving production workloads including Microsoft Copilot. GitHub Copilot and Microsoft Foundry are due to follow by the end of the year. It runs OpenAI models, Microsoft's own MAI models and open-source model families.
- Maia 300 production deployments start in 2027. Microsoft hasn't given a deployment timeline for the broader multi-gigawatt build-out. General manager Alistair Speirs told CNBC the company keeps "a little bit of optionality" because the industry changes quickly.
- Maia 200 is already spreading. Microsoft says it is deploying across the United States, Europe and Australia. It claims up to 1.7x more tokens per watt than leading GPU infrastructure for built-in AI features in Teams, Copilot and Cowork.
Maia 300 by the numbers
Microsoft's published specifications for Maia 300:
| Item | Microsoft's stated figure |
|---|---|
| Memory | Eight HBM4 stacks, 18.0 TB/s peak bandwidth |
| Power | 950 W TDP per accelerator |
| Process | Compute chiplets on TSMC N2P |
| Rack | 96 accelerators, 167 kW, 27.6 TB HBM, 1.7 PB/s aggregate bandwidth |
| Rack throughput | Up to 7x Maia 200 in a full 96-accelerator configuration |
| Efficiency vs GPUs | 40% more tokens per megawatt than leading GPU-based systems |
Other reports add detail on the chip. Heise reports 288 GB of HBM4 per chip and, per Microsoft, 6.9 Pflops at FP8 and 13.8 Pflops at FP4. TokenPost reports Microsoft's claim of more than 2.5 times Maia 200's performance per watt.
Part of the seven-times rack figure is simple density. Heise notes a Maia 200 pod has only half the number of AI chips (48). So some of the gain comes from packing in more silicon, not only from a better chip.
Microsoft describes the design as inference-first. It targets long-context, memory-heavy, decode-heavy work such as conversational assistants and multi-step agent tasks. Its argument is that moving data, not raw compute, is increasingly the bottleneck. It also highlights enhanced MXFP4 low-precision formats, software-managed on-die SRAM and a software-defined dataflow design.
Why Windows and Microsoft 365 admins should care
This is infrastructure, not a toggle in the admin center. Microsoft did not say what share of Copilot requests will run on Maia. It also did not say whether customers can pick Maia-backed capacity, or whether pricing or service terms will change.
The practical stakes are indirect:
- Cost of serving AI. Microsoft has been shifting toward usage-based pricing. In its fiscal Q3 call, it noted GitHub Copilot usage was weighing on cloud gross margin. Cheaper inference helps its economics, though nothing here promises lower customer prices.
- Copilot performance. Microsoft technical fellow Jim Kleewein told CNBC that parts of Microsoft 365 Copilot should speed up with Maia 300. That is an expectation, not a measured result.
- Capacity. In July, Microsoft said it expected to stay capacity-constrained at least through 2026. Power-efficient chips are one way to get more output from a fixed power budget.
Nvidia isn't going away
Microsoft's own executives say as much. At a Goldman Sachs conference in September, CFO Amy Hood said first-party chips don't mean relying only on them. She framed the plan as great price performance, whether from first-party or third-party parts... or in her own words, a mix across multiple partners. She pointed to Microsoft's CPU history: it started with one supplier, diversified to two, then added its own Cobalt chip. She tied that to margin improvement over a decade.
CNBC adds that Speirs said Microsoft still partners with Nvidia and AMD for training models. CNBC also noted that Nvidia CEO Jensen Huang appeared at a Microsoft event last week where a Surface laptop with an Nvidia processor was announced. Evercore ISI analysts estimate Maia will be about 10% of Microsoft's AI-chip spending by the first quarter of 2027. That is an analyst estimate, not a Microsoft figure.
The caveats
- Vendor-supplied numbers. The 1.7x, 40% and 7x claims come from Microsoft. The 40% comparison is against unnamed "leading GPU-based systems". Superpowerdaily notes Microsoft hasn't named the processor used for that comparison.
- Different baselines. The 7x figure compares against Maia 200. The 40% figure compares against GPUs. The 30% figure from Satya Nadella's July remarks compares Maia 200 against the latest hardware in Microsoft's own fleet. These are not interchangeable.
- No outside testing. Unlike Google's TPUs or Amazon's Trainium, Maia can't be rented by the hour. Analyst Patrick Moorhead told CNBC that nobody has tested them outside Microsoft yet.
- 2027 is a plan. Maia 300 isn't in production today, and "multi-gigawatt" has no date attached.
Bottom line
Microsoft is betting that owning the silicon, software and racks behind Copilot and Foundry will lower its cost per token and ease power constraints. If the claims hold at scale, that is a real structural advantage. For now, it's a roadmap with vendor benchmarks, and Microsoft's own CFO is telling investors to expect a mixed fleet. Microsoft says it will share more at Ignite 2026 next month in San Francisco.
Update: Additional details (October 12, 2026)
Microsoft’s architecture post adds that Maia 300 can scale to a 12,288-accelerator superpod across 128 racks. It says the chip delivers 36% more compute FLOPs, 2.5 times HBM bandwidth and 14% more scale-up bandwidth than Maia 200, with an accelerator able to send up to 750 GB/s to a peer under certain cross-rack job configurations.
The Azure post also says liquid cooling removes more than 80% of the rack’s thermal load. Microsoft describes a two-tier Ethernet-based fabric that lets workloads use hot-spare capacity in another rack without leaving the scale-up domain.
Update: Additional details (October 12, 2026)
Reuters reports that Microsoft could deploy potentially more than one million Maia chips in 2027, though Microsoft itself has not announced a chip-count target and has only committed publicly to multi-gigawatt Maia deployments.
Reuters also says Maia 300 splits functions across process nodes: compute functions use 2 nm technology while data-movement components use 3 nm silicon. Microsoft’s own specifications identify TSMC N2P for the compute chiplets but do not provide that breakdown.