Microsoft targets September Maia 300 reveal, 300K TSMC talks

Microsoft's homegrown AI accelerator program has spent most of the past year looking like a hedge against Nvidia rather than a serious replacement. That framing is starting to shift. The Information reported that Microsoft is targeting a public unveiling of its next-generation Maia 300 as soon as September and is in talks with TSMC for more than 300,000 units to be produced for delivery in 2027, with a longer-term ambition of capacity for over one million chips if the supply chain cooperates.

That is a step change from the current Maia 200, which the reporting describes as having been produced only in the tens of thousands so far. The strategic case Microsoft is making internally and to prospective customers is that it can run both its own models and OpenAI workloads more cheaply on custom silicon than on the Nvidia GPUs that still anchor Azure, cutting the bill for the inference wave everyone in the industry is now spending against. It slots into a busy stretch of hyperscaler-silicon moves across our AI infrastructure coverage. We also reported this week that Chinese AI labs still train on Nvidia because of CUDA lock-in, which is the same dependence Microsoft's Maia push is trying to loosen.

The more delicate part of the story is the target customer list. According to Bloomberg's summary of the report, Anthropic is among the names Microsoft hopes to win over, at the same time that Anthropic is standing up its own internal semiconductor team to design custom chips for Claude. Microsoft is effectively pitching its accelerator to a lab that is preparing to build its own, which is a strange sales motion even if near-term compute scarcity is enough to force a shorter-term deal.

A few things are conspicuously absent. The 300,000-unit number is a figure Microsoft and TSMC are still negotiating, not a signed purchase order, and the report itself points at component supplies and unresolved packaging deals as the constraints on the million-unit stretch goal. There is no disclosed price, no committed external launch customer, and no benchmark for Maia 300 against whatever Nvidia is shipping by 2027. All of this is currently one outlet's reporting, so treat the specifics as a direction of travel rather than a settled plan.

If the ramp lands anywhere near what is being described, the more interesting outcome is not that Microsoft finally competes with Nvidia on a spec sheet. It is that a hyperscaler for the first time has enough of its own inference silicon to move the price it charges its own customers, and to hand big AI labs a genuine second source the next time GPU allocations get tight.