[News] Microsoft Eyes Maia 300 Launch in Sep., Reportedly Seeks 300K TSMC Capacity as Google, AWS Race Ahead
As Google reportedly targets an aggressive 12–15 million TPUs by 2028, Microsoft is also stepping up its in-house AI chip push, with TSMC poised to benefit. Reuters, citing The Information, reports that Microsoft could unveil its next-generation Maia 300 as early as September and is in talks with TSMC to secure manufacturing capacity.
Microsoft is already eyeing capacity for more than 300,000 Maia 300 chips for 2027 delivery, while planning a sharp production ramp and courting major cloud customers such as Anthropic, the reports note. Still, component constraints and ongoing capacity talks with TSMC could temper those ambitions.
TSMC has long been a key partner for Microsoft, with its latest Maia 200 built on the foundry’s advanced 3nm process. Unveiled in early 2026, the inference-focused chip packs more than 140 billion transistors and is designed for large-scale AI workloads with a focus on performance per dollar, according to Microsoft.
Meanwhile, Runtimewire reports that Maia 200 features 216GB of HBM3e memory and supports lower-precision FP4 and FP8 computations. Notably, as highlighted by Reuters, Microsoft also packed the chip with a substantial amount of SRAM, a high-speed memory technology that can accelerate AI workloads handling large volumes of user requests.
Maia 300 Taking Shape
Microsoft debuted its first Maia AI chip in November 2023 but has struggled to keep pace with Alphabet and Amazon in scaling their in-house AI chip programs, Reuters reports. It is now pushing to expand Maia beyond its own data centers, with CNBC reporting in May that Microsoft was in talks to supply the chips to Anthropic. Maia 200 was also designed to run OpenAI’s GPT-5.2 model, according to CNBC.
Cost efficiency could give Maia another edge. Citing The Information, Runtimewire reports that Microsoft told investors in July that Maia 200 was 30% to 40% cheaper to operate than NVIDIA’s latest-generation chips when running OpenAI and Microsoft models. Internal testing has also pointed to further performance gains with Maia 300, the report suggests.
Maia’s Scale-Up Challenge
Still, Microsoft’s planned Maia 300 volume could remain modest compared with the scale of Anthropic’s existing commitments to rival platforms. AWS says nearly 1 million Trainium2 chips are already training and serving Claude through Project Rainier, while Alphabet has told investors that Anthropic plans to use as many as 1 million Google TPUs, according to Runtimewire.
Beyond customer adoption, Microsoft also faces intensifying competition for TSMC’s advanced manufacturing capacity. AWS’s Trainium and Inferentia chips are both manufactured by TSMC, Next Apple reports. Trainium3 uses TSMC’s 3nm process, while next-generation Trainium4 is slated to move to 2nm and launch by the end of 2027, the report suggests.
Google is pushing the scale even further. Citing Morgan Stanley, Commercial Times reports that Google could produce more than 3 million TPUs this year and as many as 5 million next year—far exceeding the more than 300,000 Maia 300 chips Microsoft is currently seeking capacity for. Its TPU 8, including the training-focused TPU 8t, is being developed using TSMC’s N3P process and CoWoS-S advanced packaging, with MediaTek contributing to the chip’s core design, Commercial Times notes.
Read more
- [News] Intel Advanced Packaging Reportedly Gains Traction vs. TSMC as Google, Amazon Weigh EMIB Adoption
- [News] Samsung Reportedly in Talks for Google 2nm TPU I/O Die Order; 2028 Mass Production Eyed
(Photo credit: Microsoft)
