Microsoft has shown its Maia 300 AI accelerator at OCP 2026, with 288GB of HBM4 memory and a four-accelerator server node. The specifications point to a chip built for serious AI infrastructure, where memory bandwidth and how accelerators connect matter as much as the headline silicon.
Microsoft AI Watch analysis
What happened
ServeTheHome reports that Maia 300 has eight HBM4 stacks, 18TB/s of memory bandwidth and 3.2TB/s of bidirectional scale-up bandwidth. Microsoft says a rack can hold 96 Maia accelerators, with roughly 12,000 in the scale-up domain.
The company showed a four-accelerator node. ServeTheHome describes liquid cooling for the accelerators, alongside air cooling for the CPUs, SSDs and other components. It also notes the system’s extensive cabling and fabric connections. Those physical details matter: a chip is only part of the machinery needed to run large AI workloads.
Why it matters
The 288GB of HBM4 and 18TB/s bandwidth give a useful sense of the capacity Microsoft is targeting. More memory can let an accelerator handle larger models or workloads, while high bandwidth helps keep data moving. The four-chip node and rack-scale figures show Microsoft presenting Maia 300 as part of a larger system, not just a component to admire in isolation.
Our read
This is a substantial look at Microsoft’s AI hardware ambitions, with more to discuss than a new chip name and a glossy stand. The useful test will be how Maia 300 performs in Azure workloads and how its rack-scale system compares with competing platforms. For now, the disclosed specifications make the direction clear: Microsoft is building across the accelerator, server and scale-up fabric.
What to watch
- When Microsoft shares availability and deployment details for Maia 300.
- How the chip performs on real AI workloads, including inference and training.
- Whether Microsoft expands access to Maia beyond its own cloud infrastructure.
Discussion spark: Should AI infrastructure buyers prioritise raw accelerator specifications, or the full rack-scale system and software ecosystem around them?
Sources and evidence
- Source update (27 October 2026, 21:28 UTC)
not affiliated with or endorsed by Microsoft