# Microsoft Launches Maia 300 AI Accelerator

> Microsoft releases Maia 300 chip for faster AI inference with 40% more tokens per megawatt than Nvidia.
- Title: Microsoft Launches Maia 300 AI Accelerator · SiliconANGLE
- Summary: Microsoft releases Maia 300 chip for faster AI inference with 40% more tokens per megawatt than Nvidia. Microsoft debuts Maia 300 inference accelerator that…
- Keywords: microsoft, maia 300, inference accelerator, nvidia, tokens, megawatt, hbm4, chiplet, technology, enterprise, Launches, Maia
- Source: SiliconANGLE — https://siliconangle.com/2026/10/12/microsoft-debuts-multichiplet-maia-300-inference-accelerator
- Author: Maria Deutscher
- Published: 2026-10-12T20:11:12+00:00
- Read time: 1 min
- Topics: microsoft, maia 300, inference accelerator, ai, nvidia, tokens, megawatt, hbm4, chiplet, technology, enterprise
## Microsoft debuts Maia 300
Microsoft Corp. today pulled back the curtains on Maia 300, its newest internally developed artificial intelligence accelerator.
## 40% more tokens per megawatt
Microsoft estimates it can process 40% more tokens per megawatt than leading gpu-based systems.
## Four chiplets with 288gb memory
Maia 300 features four computing chiplets using taiwan semiconductor manufacturing co.'s two-nanometer node and eight hbm4 modules with 288gb combined capacity.
## 20% less memory bandwidth
Maia 300 has about 20% less memory bandwidth than nvidia's rubin accelerator.
## Dataflow architecture
Maia 300 uses a dataflow architecture to minimize data movement between computing and memory circuits.
## Compressed data formats
Maia 300 compresses data into lightweight formats like mxfp4 to reduce memory footprint and data movement.
## Device control chiplet
Maia 300 includes a device control chiplet that manages power flow, collects telemetry, and handles data movement.
## 96 accelerators per rack
Microsoft plans to deploy 96 accelerators in liquid-cooled racks with seven times the token throughput of previous maia 200 systems.
## Key takeaway

Microsoft debuts Maia 300 inference accelerator that processes 40% more tokens per megawatt than leading GPU systems.