Microsoft Unveils Maia 300, Its First Chiplet-Based AI Accelerator, Targeting 2027 Production

Microsoft announced the Maia 300 inference accelerator on October 12 at the OCP Global Summit in San Jose, the same event where AMD also dropped a significant rack-scale product. The more interesting story, technically, is Maia 300.

The chip is the first in the Maia family built as a multi-chiplet part. As reported in detail by StorageReview, which reviewed Microsoft's technical briefing materials, each accelerator pairs a compute die on TSMC's N2P process with two I/O dies and four memory dies, flanked by eight HBM4 stacks. That disaggregation lets Microsoft apply the expensive 2nm node only where it earns its cost. The I/O and memory chiplets run on a less costly 3nm process, according to reporting by Notebookcheck citing Reuters.

The memory numbers matter here because inference workloads are usually memory-bound, not compute-bound. The eight HBM4 stacks give Maia 300 288GB of capacity at 18 TB/s of bandwidth. That's 2.5 times the bandwidth of Maia 200's 7 TB/s and a capacity jump from 216GB, according to specs reviewed by StorageReview. On compute, the gains are narrower: 13.8 petaFLOPS at FP4, up from Maia 200's 10.1, a 37% step.

The rack configuration is where the efficiency argument lives. Microsoft packs 96 Maia 300 accelerators into a single 167-kilowatt rack. The company claims that rack delivers seven times the token throughput of a comparable Maia 200 rack, and 40% more tokens per megawatt than what it describes as leading GPU-based systems, per figures in the Microsoft blog confirmed by secondary reporting from CNBC and Notebookcheck. A Maia 300 rack draws more power than a Maia 200 rack, but the per-watt efficiency argument is what Microsoft is leading with.

The competitive framing is harder to evaluate cleanly. As SiliconAngle noted in its OCP coverage, Maia 300 matches Nvidia's Rubin accelerator on memory capacity at 288GB, but carries about 20% less memory bandwidth than Rubin. The architecture compensates with a dataflow design that tries to minimize round-trips to HBM: compute chiplets prioritize an internal SRAM cache and, when HBM access is unavoidable, route data to the nearest physical stack rather than across the die.

There is a structural reason Microsoft is building this at scale. At 96 chips per rack and production clusters targeting roughly 12,000 accelerators, according to the StorageReview briefing, Microsoft is describing a deployment that doesn't fit the model of renting compute from Nvidia at hyperscale margins. The company has reportedly negotiated volume for more than 300,000 Maia 300 units in 2027, with a longer-term goal above one million, according to Gadget Salvation's sourcing of Quartz reporting.

Production deployments begin in 2027, which means this announcement is a design disclosure, not a shipping product. The one number Microsoft hasn't published is a rack-level throughput comparison against Nvidia's GB300 system. Until that figure surfaces, the 40% tokens-per-megawatt claim is a vendor number against an unspecified baseline, and buyers should treat it accordingly.

The OCP venue is itself a signal. By disclosing Maia 300 at an open-standards conference alongside AMD's Helion rack contribution, Microsoft is positioning its hardware roadmap within an ecosystem that doesn't require proprietary rack designs or vendor-controlled software stacks. Whether that posture survives contact with 2027 production schedules is a separate question.

Sources cited:
- StorageReview (https://www.storagereview.com/news/microsoft-maia-300-hbm4-n2p-chiplet-96-accelerator-rack)
- Notebookcheck (https://www.notebookcheck.net/Microsoft-Maia-300-AI-chip-arrives-with-monstrous-288-GB-HBM4-memory-and-18-TB-s-bandwidth.1422220.0.html)
- SiliconAngle (https://siliconangle.com/2026/10/12/microsoft-debuts-multichiplet-maia-300-inference-accelerator/)
- Gadget Salvation (citing Quartz) (https://www.gadgetsalvation.com/blog/microsoft-ai-chips-maia-300-vs-maia-200)
- Igor's Lab (https://www.igorslab.de/en/microsoft-maia-300-288-gb-hbm4-18-tb-s-and-96-ai-accelerators-per-rack/)