Microsoft to make its own Maia chips its primary AI inference platform from 2027
The company announced on Monday that its own AI chip is already running Microsoft Copilot in production, and that Maia 300 will make the Maia family the company's primary inference platform at multi-gigawatt scale. The efficiency figures are, for now, only Microsoft's own expectations.
Microsoft announced on October 12, 2026, that the Azure Maia 200 is now in production and that its successor, the Maia 300, will arrive in 2027. In a blog post, Scott Guthrie, executive vice president of Cloud + AI, writes that the company will scale the Maia family to "multi-gigawatt deployments from 2027" and that Maia will become the "primary inference system for Microsoft's products and services" (Microsoft Blog).
The announcement is the clearest signal yet that Microsoft intends to move AI inference—the running of models in production—onto its own silicon, and thereby reduce its dependence on Nvidia's GPUs in this area. The qualification matters: Maia applies only to inference. According to Alistair Speirs, a general manager at Microsoft, the company is continuing its partnerships with both Nvidia and AMD for model training.
What is being announced
Maia 200 is already in operation. According to Microsoft's blog post, the chip serves 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. The chip runs OpenAI models, Microsoft's own MAI models, and open-source model families, the company says.
The Maia 300 launches in 2027, but Microsoft declined to be more specific about timing. Asked about the timeline for the multi-gigawatt rollout, Speirs pointed to flexibility: "We reserve a bit of flexibility there, because this is an industry that is changing very rapidly, and we want to roll out the right silicon for the right models we use, at the right time," he told CNBC (CNBC).
Guthrie describes the Maia 300 as a complete system, not just a chip: "Maia 300 is a complete AI system designed for the new types of AI workloads that are emerging today: reasoning over long contexts, low-latency conversations, and persistent multi-step agentic tasks," he told The Verge (The Verge).
The specifications
According to Microsoft's own figures, the Maia 300 is built on eight HBM4 memory stacks delivering 18.0 TB/s of maximum memory bandwidth, within a power envelope of 950 watts (TDP). The compute chiplets are manufactured on TSMC's N2P process.
At the rack level, a full Maia 300 rack is to hold 96 accelerators, 27.6 terabytes of HBM memory, and roughly 1.7 petabytes per second of aggregate bandwidth, within a power budget of 167 kilowatts. The figures are reported by Il Sole 24 Ore, citing Microsoft's published specifications (Il Sole 24 Ore).
The efficiency claims—not verified measurements
The two most concrete efficiency figures are Microsoft's own expectations, not independent measurements:
- 40% more tokens per megawatt compared with "leading GPU-based systems." Microsoft itself writes that the company "expects" this as a result of efficiency gains across the system.
- Up to 7x rack-level throughput over the Maia 200, in a full configuration of 96 accelerators. The comparison thus concerns entire racks, not individual chips.
Il Sole 24 Ore points out that Microsoft's materials neither identify the reference GPUs nor describe the measurement conditions for the comparisons. There is, in other words, no independent benchmark yet to substantiate the figures.
What Maia means for the Nvidia dependence
The announcement primarily affects the inference side of Microsoft's AI operations—the portion of the workload that is growing fastest, since every user request to Copilot or other models requires compute. Because Microsoft continues to train models on Nvidia and AMD hardware, the initiative is about moving the execution of models, not about replacing those vendors generally.
Analysts at Evercore ISI estimate in a note to clients that Maia will account for around 10% of Microsoft's spending on AI chips by the first quarter of 2027—an estimate, not a Microsoft figure.
External access is for now lacking. Maia accelerators cannot be rented by external Azure customers today, and analyst Patrick Moorhead told CNBC that "no one has tested them outside Microsoft yet."
According to Cryptobriefing, Microsoft is negotiating with TSMC for the production of more than 300,000 Maia 300 chips, with delivery in 2027, potentially more than a million units further out, and has reportedly discussed broader Maia use with potential customers such as Anthropic. These are reports of ongoing negotiations—not confirmed deals—and they do not appear in Microsoft's own statements (Cryptobriefing).
What remains to be known
Three questions remain open after the announcement. First, independent benchmarks are needed to verify the 40% and 7x figures, including information on which GPUs they are compared against. Second, Microsoft has given no timeline for the multi-gigawatt rollout beyond "from 2027." Third, it is unclear when, or whether, Maia chips will become available to external customers, or what share of Microsoft's inference volume they will actually cover.