Microsoft Unveils Maia 300, a Custom AI Chip Aimed at Nvidia's Grip on the Market

Microsoft unveiled its Maia 300 AI chip on October 12 at the OCP Global Summit in San Jose, positioning it as the primary processor for AI inference across its own products and Azure cloud services. It is the third chip in Microsoft's Maia line, which the company first introduced in November 2023.

Microsoft says Maia 300 delivers 40% more tokens per megawatt than leading GPU-based AI systems, though it did not name which systems it compared against, so the figure is a vendor claim rather than an independently verified benchmark. The chip takes a different design approach than Nvidia's accelerators: instead of one large computing die surrounded by memory chips, Maia 300 splits its central processor into chiplets, using TSMC's 2-nanometer process for core computing and cheaper 3-nanometer chiplets for moving data.

At rack scale, Microsoft fits 96 Maia 300 chips into a single rack drawing roughly 167 kilowatts, compared with Nvidia's current GB300 racks, which hold 72 GPUs and draw about 142 kilowatts. Microsoft executive vice president Scott Guthrie said in a blog post that the company plans to scale Maia deployments to multiple gigawatts of compute starting in 2027.

On the supply side, multiple reports say Microsoft has been negotiating with TSMC for manufacturing capacity covering more than 300,000 Maia 300 chips for delivery in 2027, with longer-term ambitions above 1 million units. Microsoft has not confirmed specific volumes, and reports note that component availability and the pace of TSMC talks could limit how fast those plans play out.

Maia 200, the previous generation, is already in production and currently powers Microsoft's Copilot AI assistant. Microsoft still relies heavily on Nvidia and AMD GPUs for training large models, so Maia 300 is better understood as an attempt to gradually reduce that dependence for inference workloads rather than an immediate replacement.

The move fits a broader pattern among major cloud providers, which have been building their own AI silicon - Google's TPUs and Amazon's Trainium chips among them - to manage the cost and supply constraints of running AI at the scale current demand requires.