Microsoft Promises Maia 300 as Its Primary AI Inference System From 2027 — With Major Caveats
The company announced on Monday, October 12, 2026, that its in-house AI chips will be rolled out at multi-gigawatt scale and become the main system for running AI across Microsoft's products. But the efficiency figures are Microsoft's own claims without independent verification, training remains with Nvidia and AMD, and analysts estimate Maia will still account for only around ten percent of the company's AI chip consumption in early 2027.
On Monday, October 12, 2026, Microsoft announced the third generation of its in-house AI chip, the Maia 300, committing at the same time to scale the Azure Maia family to multi-gigawatt deployments from 2027 and to make Maia the primary inference system for Microsoft's products and services. That is stated in a blog post by Scott Guthrie, executive vice president of Cloud + AI, on Microsoft's official blog (Microsoft blog).
CEO Satya Nadella framed the announcement as "our biggest bet yet on our own silicon" in a post on X, according to GeekWire. Formally, it is a news announcement with one concrete message: Microsoft is shifting the center of gravity for AI inference — not AI training — from Nvidia's GPUs to its own chips. But there are significant caveats, both to the numbers and to the pace.
What the chip is
The Maia 300 is built on a new multi-chiplet architecture that separates compute, memory control and I/O into separate chip elements. According to Microsoft's blog post, this is intended to improve efficiency, scalability and cost structure. The compute chiplets are manufactured on TSMC's N2P process node to maximize performance per watt.
The chip has eight HBM4 stacks delivering 18.0 TB/s of peak memory bandwidth, within a thermal envelope of 950 watts. In a full configuration with 96 accelerators per rack, Microsoft claims the Maia 300 delivers up to seven times the rack-scale token throughput of its predecessor, the Maia 200. All of these figures are Microsoft's own claims — none of them have been independently verified.
Reuters also compares rack-level power consumption: Maia 300 racks contain 96 chips and use around 167 kilowatts, versus 142 kilowatts for Nvidia's latest GB300 rack with 72 GPUs (Reuters via Yahoo Finance). The figures say something about both efficiency and density, but are not direct performance measurements.
The efficiency claim highlighted in coverage is that Microsoft "expects" the Maia 300 to deliver 40 percent more tokens per megawatt than "leading GPU-based systems." The wording matters: it is a stated goal, not a benchmark result. Reuters notes that Microsoft has not named which systems the comparison is based on.
How big this is
Reuters reports that Microsoft says it plans to deploy potentially more than one million Maia 300 chips next year to back its internal AI services. The context is dramatic if the number holds: the predecessor Maia 200, launched in January 2026, was reportedly produced in limited runs — tens of thousands of units, according to Cryptobriefing. A jump from tens of thousands to potentially more than a million would, in that case, be a wholesale transformation of production scale, not an incremental ramp-up.
Microsoft has at the same time said it expects to spend more than 175 billion dollars in capital expenditures in the current calendar year, according to Reuters. The custom silicon push must be seen against that backdrop.
What it does not change
The most important caveat is that Maia is exclusively about inference — that is, running already-trained models. When it comes to training models, "Microsoft partners with Nvidia and AMD," Alistair Speirs, a general manager at Microsoft, told CNBC. So this is not about Microsoft dropping Nvidia, but about moving one portion of the workload.
Speirs also declined to give a timeline for the multi-gigawatt rollout: "We reserve some flexibility there, because this is an industry that is changing very rapidly, and we want to deploy the right silicon for the right models at the right time," he told CNBC. That makes Reuters' figure of "potentially more than a million chips next year" an indication of scale, not a confirmed deployment volume.
Analysts at Evercore ISI estimate in a note to clients that Maia will account for around 10 percent of Microsoft's AI chip spending by the first quarter of 2027. That means that even with the Maia 300 in full production, the bulk of Microsoft's AI chip consumption — and thus its dependence on Nvidia and AMD — remains unchanged in the short term.
Nadella himself emphasized, according to GeekWire with reference to the July earnings presentation, that Microsoft will be among the first cloud providers to deploy the next generation of systems based on Nvidia's Vera Rubin and AMD's Helios. Nvidia remains a key partner.
Why Microsoft is doing this
The motivation, according to the company itself, is control and cost. Mustafa Suleyman, CEO of Microsoft AI, told The Verge: "We have to work across the entire stack, all the way down to silicon. At that level, we need a platform we own, and that we co-design with our models. That's why it's a critical focus for our team to run Microsoft's AI models on top of the Maia chip."
The Maia 200 is already running production workloads, including Microsoft Copilot, and supports the MAI team in developing next-generation models. GitHub Copilot and Microsoft Foundry are to follow by the end of the year, according to Microsoft. The Maia chips are, according to Newsbytes, intended to make services like Microsoft 365 Copilot more efficient and cheaper to operate.
Industry context and open questions
The push coincides with a broader movement among hyperscalers and AI companies toward their own silicon: Google has begun selling its Tensor Processing Units directly to some customers, Amazon's Trainium chips have won AI companies such as Anthropic, Meta has rolled out its own AI chips, and OpenAI launched its first custom chip, developed with Broadcom, in June. Microsoft is thus following a path others have already opened.
Significant questions remain before the concrete figures can be taken as fact. The 40 percent efficiency claim lacks named comparison systems and independent verification. "No one has tested them outside of Microsoft yet," Patrick Moorhead, CEO of the analysis firm Moor Insights and Strategy, told CNBC — a difference from Google and Amazon, which rent their chips out to external customers. The Maia chips are so far run only internally, meaning outside customers or independent tests cannot verify the performance.
The pace of the rollout is also unresolved, with reference to Speirs' stated flexibility. And the Evercore analysis of roughly 10 percent of the AI chip budget in early 2027 is a reminder that even a fully successful Maia 300 rollout is an incremental change in Microsoft's total silicon portfolio — not a break with Nvidia.
What is genuinely new and concrete is the commitment: from 2027, Maia is to be the primary inference system for Microsoft's products and services. That sets a clear direction for where the AI workloads in Microsoft's data centers will run in the years to come — but how quickly that happens, and how much it actually reduces dependence on Nvidia, remains to be seen.