Microsoft plans to deploy potentially more than a million Maia 300 chips next year (2027) to back its internal AI services.
Microsoft says Maia 300 delivers about 40% more useful computing per megawatt than leading GPU-based systems, without naming the systems compared.
A Maia 300 rack with 96 chips uses about 167 kilowatts, compared with 142 kilowatts for Nvidia's GB300 rack with 72 GPUs.
Microsoft has said it expects to spend more than $175 billion in capital expenditures this calendar year.
Microsoft announced the Maia 300 on Monday, 12 October, calling it the third generation of its custom chips for AI workloads. The company said it plans to deploy potentially more than a million of them next year to back its internal AI services. The Maia 300 is designed mainly for inference, the stage in which a trained model answers a request, and targets long-context reasoning, conversational AI and multi-step agent tasks. Microsoft plans to start the production rollout in 2027 and scale to multi-gigawatt deployments from that year. Its predecessor, Maia 200, was announced in January 2026 and already runs production workloads, including Microsoft Copilot and Microsoft's own AI models. Scott Guthrie, executive vice president of Microsoft's cloud group, put the efficiency case directly.
“We expect Maia 300 to deliver 40 percent more tokens per megawatt compared to leading GPU-based systems.”
The central processor in the Maia 300 is built from chiplets, specialized chips bonded together, and its compute chiplets are made on TSMC's 2-nanometer N2P process. Other chiplets use cheaper 3-nanometer technology for tasks such as moving data around the system. The design pairs eight stacks of HBM4 memory with a claimed peak bandwidth of 18TB/s within a 950W thermal design power. Racks holding 96 Maia 300 chips draw about 167 kilowatts, compared with 142 kilowatts for Nvidia's GB300 AI server racks, which hold 72 GPUs. Microsoft says the chip delivers about 40% more useful computing per megawatt than leading GPU-based systems, but it did not name the systems it compared against. Microsoft also says a 96-accelerator rack provides up to seven times the rack-scale token throughput of a Maia 200 rack.
Data from the article text
Rani Borkar, president of Azure hardware systems and infrastructure, described the design approach in an interview with Reuters.
“It's basically delivering more intelligence for every watt, every dollar, and every byte of memory by optimizing across the entire system.”
Microsoft has said it expects to spend more than $175 billion in capital expenditures this calendar year. According to The Verge, the efficiency gains would let Microsoft rely mainly on its in-house chips rather than those from Nvidia and AMD. Bloomberg reported that Microsoft is currently not planning to rent the underlying computing power of Maia chips directly to its cloud customers. Microsoft aims to lower the cost of the operations that Copilot runs each time it summarizes a meeting, suggests a line of code or reads a document. Mustafa Suleyman, Microsoft AI CEO, tied the chip to the company's own models.
“We have to work across the full stack, all the way down to the silicon.”
Microsoft presented Maia 100 in 2023, and Maia 200 was announced in January 2026. Maia 300 was unveiled on 12 October 2026 as the third generation of the line. The production rollout is set to begin in 2027, with Microsoft expecting multi-gigawatt deployments from that year. The milestones are laid out in the chart below.
Pollar is free and carries no ads in the text. Its members keep it going.
Support PollarAI & Tech · San Francisco
Microsoft's third-generation custom AI chip uses TSMC 2-nanometer technology in its central processor and is set to roll out in multi-gigawatt deployments starting in 2027.
AI text based on reports by Xataka, Bloomberg Business, Reuters and other publishers. Published .