Microsoft is preparing to introduce Maia 300, the next generation of its internally designed artificial-intelligence accelerator, potentially as early as September 2026. The company is also discussing manufacturing capacity with Taiwan Semiconductor Manufacturing Company for more than 300,000 Maia 300 chips for delivery during 2027, people familiar with the programme said.
The reported volume would represent a substantial expansion from Microsoft’s earlier Maia deployments. The company ultimately wants access to capacity for more than one million Maia 300 units, although component availability and negotiations with TSMC could constrain the pace of production. Microsoft has not publicly confirmed the launch schedule or volume targets.
Maia is part of Microsoft’s effort to exercise greater control over the economics, energy efficiency and availability of the computing infrastructure behind Azure, Microsoft 365 Copilot and its broader AI portfolio. Custom accelerators allow hyperscale cloud operators to configure chips, networking, memory and software around their own workloads while maintaining Nvidia and AMD hardware in heterogeneous data-centre fleets.
Microsoft introduced Maia 200 in January as an inference-focused accelerator manufactured on TSMC’s 3-nanometre process. The chip incorporates 216 gigabytes of HBM3e memory delivering seven terabytes per second of bandwidth, 272 megabytes of on-chip SRAM and native support for lower-precision FP8 and FP4 calculations. Microsoft said the platform could deliver more than 10 petaFLOPS of FP4 performance and improve total cost of ownership by more than 30% against the latest hardware then deployed in its fleet.
Maia 300 is expected to extend Microsoft’s custom-silicon programme across both AI training and inference. The company has been building the associated software stack alongside the hardware, an essential requirement because accelerator adoption depends on compilers, model libraries, scheduling systems and orchestration tools as much as raw chip performance. The programme also gives Microsoft another route to secure computing capacity while advanced GPUs, high-bandwidth memory and packaging remain subject to intense global demand.
The roadmap intersects with Microsoft’s previously announced $17.5 billion investment in India between 2026 and 2029. That programme covers cloud and AI infrastructure, population-scale AI adoption and skills development, with Microsoft saying a new Indian data centre would begin operating in mid-2026 and give it its largest hyperscale presence in the country.
The Maia 300 report does not establish that the new accelerator will be installed in Microsoft’s Indian facilities. However, Microsoft’s custom chips are designed for Azure’s integrated global infrastructure, making their performance, availability and operating costs relevant to Indian enterprises consuming Azure AI, Microsoft Foundry and Copilot services. The programme will also affect the mix of proprietary and merchant silicon underpinning the company’s expanding Indian cloud capacity.
