Microsoft announced Azure Maia 300 on October 12, 2026, and said production deployments are planned to begin in 2027. The third-generation accelerator is designed for AI inference. Microsoft also plans to begin multi-gigawatt deployments of the Maia family in 2027, according to its announcement.
A memory-heavy inference design
Microsoft designed Maia 300 for long-context reasoning and other inference workloads that put heavy demands on memory and data movement. Each accelerator has eight HBM4 stacks—high-bandwidth memory—with 288 GB of capacity and a peak bandwidth of 18 TB/s. Microsoft gives the accelerator a 950-watt thermal design power (TDP), a measure of its design power target. Its architecture overview also describes a chiplet design that separates compute, memory-controller and I/O functions.
Reuters reported that the compute chiplets use TSMC’s 2-nanometer technology and other chiplets use its 3-nanometer technology. Microsoft describes a rack with 96 accelerators arranged in 24 trays of four, within a 167-kW power envelope. Each tray also includes a host processor, memory, storage and cooling.
The racks connect through a custom, Ethernet-based two-tier scale-up network. Microsoft says a superpod can scale to as many as 128 racks and 12,288 accelerators, with accelerators communicating directly without routing through a host or a separate scale-out network interface card. For cooling, Microsoft describes liquid cooling for more than 80% of the rack’s thermal load from accelerators, voltage regulators and switch chips; air cooling handles memory, storage and other components. The system can use facility water or a heat-exchanger unit in air-cooled data centers.
What Microsoft’s performance claims measure
Microsoft claims Maia 300 can deliver about 40% more tokens per megawatt than “leading GPU-based systems.” The company does not name those systems or specify the comparison conditions. It also claims that a full 96-accelerator Maia 300 rack can deliver up to seven times Maia 200’s rack-scale token throughput. For memory, Microsoft says Maia 300 offers 2.5 times Maia 200’s HBM bandwidth, 33% more capacity and about twice the HBM bandwidth per watt.
Reuters reported the following rack-level figures for Maia 300 and NVIDIA’s GB300 configuration:
| Rack configuration | Processors per rack | Rack power |
| Maia 300 | 96 accelerators | About 167 kW |
| NVIDIA GB300 | 72 GPUs | About 142 kW |
The rack figures describe different configurations; they do not measure performance under matched conditions. Nor does Microsoft’s 40% claim provide a direct Maia 300–Google Ironwood result: its stated comparison group is unnamed GPU systems.
Maia’s planned role in Microsoft services
Microsoft describes Maia as infrastructure for its own products and services. In an October 12, 2026 analysis, Matt Kimball said Microsoft had not announced Maia instances that Azure customers could rent as of the date of his analysis. (Moor Insights & Strategy)
