Microsoft is seeking to sharply increase production of its next-generation AI chip that it designed in-house next year to attract large customers such as Anthropic, The Information reported on Aug. 10, citing two people familiar with the matter.

Microsoft plans to unveil the new "Maia300" chip this autumn, possibly as early as next month, even though adoption of its in-house Maia200 AI chip has been weak, the report said. Microsoft is negotiating with TSMC to secure production capacity of more than 300,000 Maia300 chips in 2027. That would be on a different scale from the tens of thousands of Maia200 chips produced so far.

Microsoft ultimately wants to secure production capacity of more than 1 million chips, but could be constrained by component supply and negotiations with TSMC, The Information reported.

Putting the Maia business on a stable footing is a key pillar of Satya Nadella (사티아 나델라) Microsoft CEO's top priority of reducing reliance on Nvidia, The Information reported. Maia200 was delayed after initial tests last year fell short of internal targets, and it has since been used only in some areas.

Nadella said in June that two data centres were using Maia200 and that Microsoft planned to expand usage, including overseas. As of late last month, however, Maia chips were still being used only at two U.S. data centres, The Information reported.

Microsoft is lagging Google and Amazon in commercialising its in-house AI chips. Google and Amazon supply TPU and Trainium chips, respectively, to large customers, and Google has started selling TPUs to customers beyond Google Cloud. By contrast, Microsoft has so far used Maia chips for OpenAI models and its Copilot MAI models, and most software still relies on Nvidia chips, The Information reported.

Microsoft appears to believe it can provide Maia300 to more customers such as Anthropic by highlighting cost savings. The Information previously reported that Anthropic has been discussing use of Maia chips with Microsoft for several months.

Microsoft plans to shift a larger share of its own AI usage to Maia even if it fails to secure large customers. Microsoft said last month that Maia200 runs models at costs 30 to 40 percent lower than the latest Nvidia chips.

Even so, Microsoft's targeted Maia300 production volume falls well short of Google and Amazon. By Morgan Stanley estimates, Google is set to produce more than 3 million TPU chips this year and 5 million next year.