Microsoft shows Maia 300 AI accelerator at OCP 2026

Custom chip features 288GB HBM4 memory, 18TB/s bandwidth and 950W TDP

By LineZotpaper AI staff

Published27 October 2026

Read Time2 min

Illustration: Microsoft shows Maia 300 AI accelerator at OCP 2026

At the OCP Summit 2026, Microsoft showcased its next-generation in-house AI accelerator, the Maia 300, along with a complete four-accelerator node. The chip boasts 288GB of HBM4 memory with 18TB/s bandwidth and is designed for large-scale AI workloads, with Microsoft fitting 96 Maia 300 accelerators per rack and around 12,000 in a scale-up domain.

The Maia 300 was revealed at the Open Compute Project Summit 2026, following a deep-dive on the previous-generation Maia 200 at Hot Chips 2026. Microsoft brought an entire 4-accelerator node to the show floor, demonstrating the platform's design.

Each accelerator has a thermal design power of 950W and is cooled by a liquid-cooled cold plate. The system also includes a massive fan partition for cooling CPUs, SSDs, and other components, meaning it is not fully liquid-cooled.

The node features extensive cabling for scale-up connectivity, with Microsoft opting for a cable-heavy design in contrast to NVIDIA's approach with fewer cables in its Vera Rubin NVL72 nodes. The accelerators sit on individual PCBs with multiple cables, and the system uses busbar power and blind-mate liquid cooling connections at the rear.

Microsoft says it can fit 96 Maia 300 accelerators into a single rack and scale to roughly 12,000 accelerators in a scale-up domain. The chip provides 3.2TB/s of bidirectional scale-up bandwidth. The node is wider than standard open compute racks but not as wide as some other custom designs.

The author of the report noted the potential interest in a Maia-based developer workstation if Microsoft decided to sell the accelerators, given strong demand for high-end AI tower systems.

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Analysis

Why This Matters

Microsoft's custom AI chip development reduces reliance on external suppliers like NVIDIA and demonstrates the growing trend of hyperscalers designing their own silicon.
The Maia 300's high memory bandwidth and capacity are critical for training large language models and other AI workloads.
The system design choices, such as cable-heavy interconnects and liquid cooling, influence data center infrastructure and cooling strategies.

Background

Microsoft has been developing its own AI accelerators as part of the Maia family. The Maia 200 was detailed earlier in 2026 at Hot Chips. The Maia 300 represents the next step, with significant increases in memory and bandwidth. The OCP Summit is a key venue for open hardware designs, and Microsoft has been a contributor to open compute standards.

What to Watch

Whether Microsoft will eventually offer Maia accelerators for external sale, which could shake up the AI hardware market.
How the Maia 300 performs in Azure deployments compared to competing accelerators from NVIDIA and AMD.
The adoption of liquid cooling for high-TDP AI accelerators in data centers.

Sources

Microsoft Maia 300 Shown at OCP 2026 · ServeTheHome

Checked

12 claims in this story checked against 1 source: 9 supported.

Show each claim

SupportedMicrosoft showed the Maia 300 AI accelerator at OCP 2026.ServeTheHome: “For OCP, it showed the next-generation chip, the Maia 300.”
SupportedThe Maia 300 has 288GB of HBM4 memory.ServeTheHome: “This has eight HBM4 stacks for 288GB of memory”
SupportedThe Maia 300 has 18TB/s of memory bandwidth.ServeTheHome: “and 18TB/s of memory bandwidth”
SupportedThe Maia 300 has a 950W TDP.ServeTheHome: “One thing you can clearly see is that the 950W TDP accelerators are covered by a liquid-cooled cold plate.”
Not in the sourcesThe Maia 300 is designed for large-scale AI workloads.ServeTheHome: “Microsoft’s project manager for the Maia 300, the company’s new in-house AI acceleration platform.”
SupportedMicrosoft brought a complete four-accelerator node to OCP 2026.ServeTheHome: “Instead of just bringing a chip to the show, they brought an entire 4-accelerator node”
SupportedMicrosoft says it can fit 96 Maia 300 accelerators in a rack and roughly 12,000 in a scale-up domain.ServeTheHome: “Microsoft says that it fits 96 Maia’s into a rack and roughly 12K in the scale-up domain.”
Partly supportedThe Maia 300 was revealed at OCP 2026, after a deep-dive on the Maia 200 at Hot Chips 2026.ServeTheHome: “A few weeks before OCP 2026, at Hot Chips 2026, Microsoft did a deep-dive on the Maia 200. For OCP, it showed the next-generation chip, the Maia 300.”
SupportedThe Maia 300 has 3.2TB/s of bidirectional scale-up bandwidth.ServeTheHome: “and 3.2TB/s of scale-up bidirectional bandwidth.”
SupportedThe accelerators are covered by a liquid-cooled cold plate.ServeTheHome: “the 950W TDP accelerators are covered by a liquid-cooled cold plate.”
SupportedThe node is not fully liquid-cooled; a fan partition cools CPUs, SSDs and other components.ServeTheHome: “Another item we noticed is that this is not an all-liquid-cooled design. The CPU, SSDs, OCP NIC 3.0 devices, and many components around the AI accelerators are cooled by a massive fan partition that spans the chassis.”
Partly supportedThe node has extensive cabling for scale-up connectivity.ServeTheHome: “Our best guess is that these are UltraPass connectors that help scale within the node.”