Microsoft plans to significantly increase production of its next-generation in-house AI chips, and has been in talks with Taiwan Semiconductor Manufacturing Co. to secure manufacturing capacity for more than 300,000 units for delivery in 2027, according to a report cited by Bloomberg.
The company is also preparing to unveil its new Maia 300 accelerator this fall. A production run of that scale would mark Microsoft's most serious step yet toward reducing its dependence on Nvidia hardware for its data centers — following the path Google and Amazon have taken with their own custom silicon programs.
The supplier side of the equation underlines why. TSMC reported July revenue of NT$467.58 billion — roughly $14.5 billion — up about 45 percent from a year earlier, and recently raised its 2026 dollar-revenue growth outlook to more than 40 percent, per Yahoo Finance. The foundry continues to expand advanced packaging capacity such as CoWoS, the bottleneck technology for high-performance AI accelerators.
The custom-silicon race compounds
Every major cloud provider now designs its own AI chips, but volume is the differentiator: a 300,000-unit order moves Microsoft's program from supplementary to structural. For customers, the bet is that first-party silicon eventually translates into lower inference prices; for Nvidia, it is one more large buyer becoming a partial competitor.
Neither Microsoft nor TSMC has commented publicly on the reported capacity discussions.