Microsoft Unveils Maia 300

Microsoft unveiled its third-generation in-house AI processor, the Maia 300, on Monday (October 12) at the Open Compute Project Global Summit 2026 in San Jose. The chip, designed specifically for inference workloads of large language models, marks a decisive step in the company's strategy to reduce its reliance on Nvidia GPUs, which currently dominate the AI infrastructure landscape.

Scott Guthrie, Executive Vice President of Microsoft's Cloud + AI Group, said the Maia 300 is expected to deliver "40% more tokens per megawatt compared to leading GPU-based systems" and that in a full 96-accelerator rack configuration, it "delivers up to 7x the rack-scale token throughput of Maia 200."

From Maia 100 to 300: an evolving line

The Maia 100 was introduced in November 2023 but remained limited to small-scale deployments. The Maia 200, presented in January this year, represented a tangible leap forward: built on TSMC's 3 nm process with native FP8/FP4 tensor cores, 216 GB of HBM3e memory at 7 TB/s, and 272 MB of on-chip SRAM, the Maia 200 already powers Microsoft's production workloads, including Microsoft 365 Copilot, OpenAI models, and Microsoft's internal MAI models. According to the company, the Maia 200 delivers 1.7 times more tokens per watt than next-generation GPU-based systems in Teams, Copilot, and Cowork applications.

The Maia 300, while lacking detailed technical specifications in public announcements, inherits the low-power architecture and inference optimization—the work that truly matters when a model meets a user and generates a bill. Training gets the conference glamour, but inference is what sustains margins.

The strategy behind the silicon

Microsoft is not alone in this movement. Google has developed Tensor Processing Units (TPUs) since 2016 and already recognizes direct revenue from external sales. Amazon fabricates the Trainium and is already negotiating its sale to other data centers. Meta and OpenAI also have custom chip projects in development. But what distinguishes Microsoft is the scale of ambition: the company negotiated with TSMC the manufacturing of more than 300,000 Maia 300 units for delivery in 2027, with a future ambition to exceed 1 million units.

To put the scale into perspective, Microsoft has already purchased 485,000 Nvidia Hopper GPUs—double the total purchased by its competitors combined. This is not an exit from Nvidia. It is a bet that in a few years, the largest share of tokens processed by the company will pass through Microsoft's own silicon.

Andrew Wall, General Manager of Azure Maia, said: "Microsoft continues to invest in custom silicon as part of our long-term AI infrastructure strategy. While we don't disclose production volumes, the reported figures do not reflect the scale of our program."

What to expect

The Maia 300's mass production will only begin in 2027. TSMC, which already operates with tight advanced packaging capacity through 2027, needs to convince Microsoft to adopt designs that truly compete with Nvidia's GPUs on performance per cost. Microsoft intends for the Maia 300 to become "the primary AI inference system for Microsoft's products and services."

Meanwhile, the race for custom silicon continues—not as a luxury, but as a strategic necessity. With costly GPUs, limited supply, and inference costs growing alongside demand for autonomous agents and increasingly large models, dependence on a single hardware vendor is a risk no technology company can shoulder alone. The Maia 300 may be the first chip that allows Microsoft to say, out loud: "we also manufacture."

Sources: The Verge, Yahoo Finance, Microsoft Blog

✓ Independent sources cross-checked and verified before publishing