Microsoft doesn’t want to rent inference forever, which is precisely why the Microsoft Maia 300 exists. The company’s third-generation in-house AI accelerator was unveiled this week by Scott Guthrie, EVP of Microsoft Cloud + AI, and the spec sheet reads like a dare: compute chiplets built on TSMC’s N2P node, memory controller and I/O on 3nm, and eight stacks of HBM4 delivering 288 GB and 18 TB/s per accelerator.
The company is positioning it as an inference-first system – designed to become the primary inference engine across Microsoft products and services, from Copilot to Teams. In other words: the workloads that actually keep the lights on at Redmond, rather than the benchmark runs that win headlines.
What the Microsoft Maia 300 actually brings
The rack-level numbers are where the chip gets serious. Ninety-six accelerators per rack, each with a 950 W TDP inside a 167 kW power envelope. Against the Maia 200: +36% compute FLOPs, 2.5x the HBM bandwidth, and +14% scale-up bandwidth. Microsoft claims roughly 40% more tokens per megawatt than leading GPU-based systems – a vendor-reported figure that hasn’t been independently benchmarked, but one that frames the argument: this fight is now measured in tokens per dollar of electricity.
Production starts in 2027, with deployment targets described as multi-gigawatt and a stated ambition that could see more than one million Maia 300 accelerators in the field. That would put Microsoft squarely in the same custom-silicon race as Google with its TPUs and Amazon with Trainium – the hyperscaler playbook being: design your own silicon, or keep writing Nvidia checks with a perfectly straight face.
The announcement lands against a market still digesting the IMF’s warning that an AI market correction could spill into broader financial turbulence. Capacity keeps being built either way – power contracts, fabs and rack orders do not care about one fund report.
Honest fine print: the primary Microsoft blog post was Cloudflare-blocked during sourcing, so specs were cross-verified against TrendForce, Igor’sLAB, CNBC and The Verge reporting on the same announcement; efficiency claims remain Microsoft’s own. For a first-party chip now on its third generation, the Maia line has stopped being an experiment and started being strategy.
