Microsoft is getting serious about reducing its reliance on Nvidia chips. On Monday, Oct. 12, 2026, it unveiled Maia 300, its third-generation in-house AI chip, at the OCP Global Summit in San Jose. The chip is designed for inference, the work of running AI models after they are trained, which happens every time a user interacts with AI and is becoming a major cost for cloud providers. Microsoft said the Maia family will become the primary system for running AI across its products and services starting in 2027, with deployments scaling to multiple gigawatts, a signal that the software giant wants more control over its AI costs through its own hardware.
Microsoft Unveils Maia 300, Its Third-Gen AI Chip Built for Inference
Scott Guthrie, Microsoft's executive vice president of Cloud and AI, said Maia 300 will make cloud AI workloads more efficient and is better suited to reasoning models and AI agents that carry out multi-step tasks. The company expects Maia 300 to generate 40% more tokens, the units of text AI models produce, per megawatt of power than leading GPU-based systems, including those from Nvidia and AMD. However, Microsoft did not specify which GPU systems it used for comparison, and as of the launch no third-party company had tested the chip, so the figure is Microsoft's own claim.
Maia 300 Specs: 2nm-Class Chip, HBM4, 96 Per Rack
Microsoft split the chip into smaller pieces called chiplets for computing, memory and data transfer, with the computing parts built on TSMC's N2P process, part of its 2-nanometer family. Each chip carries eight stacks of HBM4 memory, with 288 GB of capacity and 18 TB/s of memory bandwidth within a 950-watt power budget. Compared with Maia 200, the new chip offers 2.5 times the memory bandwidth, 33% more capacity and roughly twice the HBM bandwidth per watt. In its full configuration, a rack holds 96 Maia 300 chips within about 167 kilowatts, roughly double the number of Maia 200 chips per rack.
Cutting Reliance on Nvidia as Big Tech Builds Its Own Chips
Maia 300 is part of Microsoft's plan to reduce dependence on Nvidia chips, which are expensive and hard to source. Maia 200, launched in January, already runs OpenAI's latest GPT-5.2 models, Microsoft Foundry and Microsoft 365 Copilot. Rivals all have their own silicon: Google began recognizing revenue from direct TPU sales in the quarter ended June, Amazon has its Trainium chips, and OpenAI has built a chip with Broadcom. Microsoft has been seen as lagging Google and Amazon here and is also trying to persuade major cloud customers such as Anthropic to adopt Maia, although no deal has been finalized. Microsoft will still build infrastructure using its own chips alongside Nvidia and AMD processors, rather than dropping Nvidia outright.
Gigawatt-Scale Plans Still Have to Prove Out in Production
On volumes, Reuters reported that the rollout could involve more than 1 million Maia 300 chips next year. In August, The Information reported that Microsoft was in talks with TSMC to manufacture more than 300,000 chips for delivery in 2027, with a longer-term goal of more than 1 million, though component shortages and capacity negotiations with TSMC could constrain those plans. Microsoft has said only that it will deploy at multi-gigawatt scale, without disclosing chip volumes, costs or what that scale means in practice. The real test for Maia 300 is therefore not paper performance but whether Microsoft can mass-produce it and convince major customers to use it.
What It Means for Investors: A Long-Term Win for Azure, No Immediate Hit to Nvidia
For Microsoft, if Maia 300 delivers as claimed, the cost of running AI services such as Copilot and Azure would fall, supporting cloud margins over time and reducing the risk of depending on a single chip supplier. For Nvidia, the near-term impact looks limited: Microsoft's chip only enters service in 2027, Microsoft will keep buying Nvidia GPUs, and Maia targets inference rather than large-scale model training, where Nvidia still holds a clear edge. Over the longer term, however, as Google, Amazon, OpenAI and Microsoft all lean more on their own chips, Nvidia's share of the fast-growing inference market could come under pressure. Clear beneficiaries include TSMC and HBM4 memory makers. Key things to watch are third-party test results, actual production volumes and which major cloud customers sign on to Maia.
