Microsoft launches Maia 300: promises primary inference system from 2027, but limited reach so far
Microsoft unveiled its third generation of in-house AI chips, Maia 300, on Monday, promising that the Maia family will become the company's primary inference system from 2027.
Microsoft unveiled its third generation of in-house AI chips, Maia 300, on Monday, promising that the Maia family will become the company's primary inference system from 2027. But behind the big numbers lies a more nuanced picture: analyst estimates point to Maia so far accounting for a small share of Microsoft's AI chip spending, and the company is at the same time betting on the next generations from Nvidia and AMD.
What happened
Microsoft unveiled Maia 300 on Monday, 12 October 2026, at the OCP Global Summit in San Jose. It is the third generation of the company's own AI chips, following Maia 100 and Maia 200, and the company said large-scale rollout begins in 2027, according to Yahoo Finance.
In a blog post, Azure chief Scott Guthrie wrote that Microsoft will scale the Azure Maia family to multigigawatt deployments from 2027, making Maia "the primary AI inference system for Microsoft's products and services," according to The Verge. Chief executive Satya Nadella reinforced the message in a post on X, describing the launch as "our biggest bet yet on our own silicon," according to GeekWire via MSN.
What Microsoft claims
The numbers Microsoft presented are ambitious, but it is important to distinguish the company's claims from verified facts. All the performance figures come from Microsoft itself and have not been independently verified.
Guthrie claimed that Maia 300 will deliver "40 percent more tokens per megawatt compared to leading GPU-based systems." He added that a full configuration with 96 accelerators provides "up to 7 times rack-scale token throughput compared to Maia 200," according to The Verge. It is worth noting that Microsoft has not defined which "leading GPU-based systems" form the basis of the comparison — it is unclear whether it refers to Nvidia's current or upcoming chips, or which workloads were measured.
How the chip is built
The technical specifications give a concrete picture of the progress from Maia 200:
- Chiplet architecture: Maia 300 is built on TSMC's N2P process node and separates compute, memory controller and I/O into separate chiplets, according to Microsoft's Azure hardware and infrastructure team, speaking to The Verge.
- Memory: The chip has eight HBM4 units, up from six HBM3E units in Maia 200. Capacity is 288 gigabytes, and bandwidth has been raised to 18 terabytes per second, from 7 terabytes per second.
- Power: Thermal design power is 950 watts.
- Rack level: A rack holds 96 accelerators, twice as many as the previous generation. Total HBM capacity per rack has been increased to 27.6 terabytes from 10.4 terabytes, and bandwidth to 1,728 terabytes per second from 341 terabytes per second, according to Seoul Economic Daily.
The sharp increase in HBM content also has commercial significance beyond Microsoft: Seoul Economic Daily points out that it strengthens demand for Korean HBM memory.
The strategic reality
Microsoft's framing is clear: the company wants to reduce its dependence on Nvidia. Yahoo Finance covered the launch precisely in that context. But the evidence of a transformative shift is thin.
First, analysts at Evercore ISI estimate that Maia will account for roughly 10 percent of Microsoft's spending on AI chips, according to GeekWire via MSN. The figure is the analysts' own estimate, and the source text is truncated, so it is unclear what timeframe the estimate covers. In any case, it suggests that Nvidia's GPUs still make up the bulk of Microsoft's AI infrastructure.
Second, Microsoft and Nvidia are still working hand in hand. Nadella said at the company's earnings presentation in July that Microsoft will be among the first cloud providers to roll out next-generation systems based on Nvidia's Vera Rubin and AMD's Helios. Maia is thus positioned as a supplement, not a replacement — at least so far.
The rollout history also gives reason for caution. Maia 200 already powers Microsoft 365 Copilot as well as OpenAI and Microsoft models, with deployments in the US, Europe and Australia. But The Information reported in August that Maia 200 was running in only two US data centers as of summer 2026, according to GeekWire via MSN. The distance between that kind of early rollout and the goal of multigigawatt inference from 2027 is considerable, and it remains to be seen whether the pace changes.
Why it matters
Mustafa Suleyman, head of Microsoft AI, gave the most principled rationale for the bet: "We have to work across the entire stack, all the way down to the silicon. At that scale we need a platform we own, and that we co-design with our models. That's why running our Microsoft AI models on top of the Maia chip is a critical focus for our team," according to The Verge.
The logic behind the bet, according to the company, is that control over both model and silicon should enable co-optimization that is harder to achieve with standard GPUs. Inference — running models in production — is a cost driver for AI services, so if Microsoft's 40 percent claim holds up, it could matter to the economics behind Copilot and other services.
But it is still an "if." All the performance figures come from Microsoft's own marketing. The sources available are secondary coverage of the same announcement, and no independent testing or third-party benchmarks exist.
Open questions
Several things remain before Maia 300 can be judged on its own terms:
- The comparison baseline: Which "leading GPU-based systems" form the reference for the 40 percent claim, and under which workloads? Microsoft has not specified this.
- Verification: No independent measurements of tokens per megawatt or rack throughput exist.
- The details of the 2027 plan: How many data centers, in which regions, and with what capacity? "Multigigawatt" is a goal, not an accomplished reality.
- The pace: The history of Maia 200 — two US data centers as of summer 2026, according to The Information — suggests that the rollout of in-house chips has been slower than Microsoft's GPU-based capacity buildout. It remains to be seen whether Maia 300 changes that pattern.
What Microsoft has promised, then, is clear: its own silicon as the primary inference system from 2027. Whether the promise is kept depends on numbers that no one but Microsoft has measured yet.