Microsoft Targets More Than 300,000 Maia AI Chips as It Expands Beyond Nvidia

Microsoft is negotiating with Taiwan Semiconductor Manufacturing Co. for capacity to produce more than 300,000 next-generation Maia AI chips for delivery in 2027, according to The Information and reports carried by Bloomberg AI, Reuters AI and Theedgemalaysia. The reported order would substantially exceed the tens of thousands of Maia 200 chips produced for Microsoft’s data centers, according to Finance.yahoo.

The scale signals a sharper push to supply Microsoft’s Azure infrastructure with internally designed accelerators. The Information’s reporting also links the plan to delays and a reassessment of the company’s chip ambitions, while Reuters AI describes next-generation production as delayed to 2026. Microsoft has not publicly confirmed the reported TSMC volume or delivery schedule in the supplied reports.

Microsoft’s custom chips are designed primarily to support AI workloads inside its cloud business. CNBC reported in January that Maia 200 would receive wider customer availability after the first-generation Maia 100 was used mainly for internal projects. TechTarget reported that Microsoft itself will remain the largest Maia 200 user, while the company’s deployment could free Nvidia GPUs for Azure customers.

Microsoft’s January launch materials described Maia 200 as an inference accelerator, a category covering the repeated execution of trained models for chatbots, search, recommendations and other services. TechCrunch reported more than 100 billion transistors, over 10 petaflops at 4-bit precision and about 5 petaflops at 8-bit precision. Those specifications position Maia as infrastructure for high-volume model serving rather than a general replacement for every GPU workload.

Microsoft’s strategy targets Nvidia’s cost, supply and software advantages at the same time. Finance.yahoo reported that the Maia 200 rollout includes software tools aimed at Nvidia’s developer ecosystem, while the same report described the chip as part of Microsoft’s effort to take a larger share of the AI infrastructure market. Nvidia remains the dominant supplier of AI accelerators, giving Microsoft a difficult software and customer-compatibility problem beyond chip design.

The company’s own account frames the effort as a systems approach spanning silicon, servers, networking and cloud services. Microsoft News reported 2 custom chips, new industry partnerships and hardware optimized for both internal and customer workloads. The approach matters because accelerator performance depends on memory, cooling, interconnects and software as much as raw compute. Microsoft News separately reported microfluidics work that cools chips up to 3 times better, underscoring the physical constraints around larger AI systems.

TSMC stands to become the clearest direct manufacturing beneficiary if Microsoft secures the reported capacity. Finance.yahoo identified TSMC as the supply-chain beneficiary and said the proposed Maia 300 order would dwarf Maia 200 production. Manufacturingdive reported that Microsoft previously selected Intel as a foundry customer for an upcoming chip using Intel’s 18A process, showing that the company is pursuing a multi-foundry strategy rather than relying on a single manufacturer.

The commercial question is whether Maia moves beyond Microsoft’s own data centers. CNBC reported talks to supply custom chips to Anthropic after Microsoft’s $5 billion investment, although Maia 200 had not been made broadly available to customers at that point. Wider external access would give cloud and model companies another accelerator option, while internal deployment alone would mainly help Microsoft manage Azure capacity and costs.

Microsoft’s next public milestone is the reported Maia 300 reveal, targeted as soon as September, according to Finance.yahoo’s account of The Information’s reporting. The company’s immediate task is turning a design and capacity plan into reliable, high-volume production while integrating software that developers can use with fewer changes.

The 2027 delivery target also puts pressure on packaging, memory, networking, power and cooling suppliers. A larger Maia fleet can reduce Microsoft’s dependence on Nvidia for selected inference workloads and reserve Nvidia hardware for applications that still require its broader software stack. The reports establish a significant production plan, but they don’t establish final chip performance, customer pricing, broad availability or whether the full order will be completed.