Microsoft Maia: The Complete Guide

Microsoft still buys AI chips from NVIDIA and AMD at scale. Satya Nadella said in January 2026 that "because we can vertically integrate doesn't mean we just only vertically integrate" 3. Maia's job is cost: Microsoft says Maia 200 gives 30% better performance per dollar than the latest hardware in its fleet 1.

Microsoft's annual report does not name Maia; it describes "hardware, which includes custom-built silicon and strong partnerships with chip manufacturers" 6. Microsoft is incorporated in Washington State 9. Its chip lab, where new chips are probed and tested, is on the Redmond campus 10.

Microsoft's first AI hardware was not a custom chip. Project Brainwave, shown in August 2017, ran AI models on Intel FPGAs (reprogrammable chips) to speed up Bing 12, and was offered in preview on Azure in 2018 15. Maia 100 followed in November 2023, announced with Cobalt, Microsoft's Arm server processor; Microsoft said it would roll out to its data centres from early 2024 10. OpenAI's Sam Altman said the two companies had worked together "to refine and test it with our models" 10.

Maia 200 is described by Microsoft as its "first silicon and system platform optimized specifically" for inference 17. Microsoft calls the programme "multi-generational" 1. Press reports say a Maia 300 could be unveiled as early as September 2026, with Microsoft seeking TSMC capacity for more than 300,000 chips for 2027; Microsoft has not announced it 18.

Maia 100 was "one of the largest processors made on 5nm node using advanced packaging technology from TSMC" 19, about 820 square millimetres 14. Maia 200 packs more than 140 billion transistors 1. Microsoft does not say which TSMC factory makes Maia.

Other names appear only in the press: SK hynix as the sole memory supplier for Maia 200 20. Microsoft announced in February 2024 that it would make "a chip design" on Intel's 18A process, without saying which chip 21.

Microsoft's flagship AI sites run NVIDIA, not Maia. Its Fairwater data centre in Mount Pleasant, Wisconsin (315 acres, three buildings, 1.2 million square feet) is described with "hundreds of thousands of the latest NVIDIA GPUs" 23, and its Atlanta Fairwater site, operating since October 2025, with NVIDIA GB200 racks 24.

Maia 100 was split into 16 clusters of four tiles, each tile with its own tensor unit for matrix maths and a vector processor 14. It was the first chip to use the Microscaling (MX) formats, open 4-, 6- and 8-bit number formats that Microsoft developed with AMD, Arm, Intel, Meta, NVIDIA and Qualcomm through the Open Compute Project 19 13.

Microsoft's own comparisons for Maia 200: "three times the FP4 performance of the third generation Amazon Trainium" and FP8 performance "above Google's seventh generation TPU" 1. On paper that is consistent with the rivals' published figures (see our Trainium and TPU guides), but these are the companies' own figures, not independent tests.

What you can rent on Azure is NVIDIA and AMD hardware, and Microsoft publishes those prices: for example an 8-GPU H100 machine lists at $98.32 an hour in East US, and an 8-GPU AMD MI300X machine at $48.00 (see our NVIDIA and AMD guides).

Microsoft also has rights to OpenAI's hardware ideas. Under the October 2025 agreement it keeps rights to OpenAI's models and products through 2032, excluding OpenAI's consumer hardware 11. Nadella said in November 2025: "As they innovate even at the system level, we get access to all of it" 28.

Maia 100's design was shaped by Copilot and aimed at Copilot and Azure OpenAI Service 10. In July 2026 Microsoft said Maia 200 supports both OpenAI and Microsoft's own MAI models 5. Microsoft has not published what share of its AI work runs on Maia 5.

6. About two-thirds of quarterly capital spending goes on short-lived assets, meaning processors and GPUs 5.
7. Microsoft booked a $3.2 billion gain on its Anthropic investment in the June 2026 quarter 4.
8. Leases signed but not yet started total about $329 billion, as extracted from the 10-K 32.
9. Microsoft's flagship Fairwater AI data centres are described with NVIDIA chips only 23 24.
10. Maia 200 has no scale-out network at all: everything is scale-up over Ethernet 26.

The same June 2025 report said the delayed chip was expected to fall well short of NVIDIA's Blackwell, and that Microsoft had dropped plans for a training chip in early 2024 16. These are press reports, not Microsoft statements.

Microsoft's other Indian regions are Central India (Pune), South India (Chennai) and West India (Mumbai), plus two regions run with Jio 34. Early Hyderabad customers named by Microsoft include Adani Group, Bajaj Finserv and HDFC Bank 34.

Press reports say Maia 300 could be unveiled as early as September 2026, with more than 300,000 chips sought from TSMC for 2027 18; as of 22 September 2026 we found no Microsoft announcement.