What Is Azure Maia 300? Microsoft's Inference Chip Explained
The short answer
Azure Maia 300 is Microsoft’s third-generation custom AI chip, built for inference rather than training, announced on October 12, 2026 and due in production in 2027. Microsoft plans to scale the Maia family to multi-gigawatt deployments from 2027 and make it the primary inference system for Microsoft products and services. You will not rent it directly; you will meet it inside Copilot, GitHub Copilot and Microsoft Foundry.
Maia 300 at a glance
| Spec | Maia 300 (Microsoft, Oct 12, 2026) |
|---|---|
| Purpose | Inference-first: long context, low-latency chat, multi-step agents |
| Memory | 8 × HBM4 stacks, 18.0 TB/s peak bandwidth |
| Power | 950 W TDP per accelerator |
| Process | Compute chiplets on TSMC N2P; separate memory-controller and I/O chiplets |
| Formats | Low-precision, including enhanced MXFP4 (also for KV-cache) |
| Rack | 96 accelerators, 167 kW, 27.6 TB HBM, 1.7 PB/s aggregate bandwidth |
| Networking | Custom AI transport over Ethernet, tuned for tensor parallelism |
| Claimed gains | 40% more tokens per megawatt than leading GPU systems; up to 7× Maia 200 rack throughput |
| Availability | Production deployments 2027; no public VM or price |
All figures are Microsoft’s; none are independent benchmarks yet.
What is new compared with Maia 200
- Memory first. Inference on long contexts is limited by moving data, not raw compute, so Maia 300 spends its budget on HBM bandwidth per watt and uses on-die SRAM as a software-managed scratchpad.
- Chiplets. Splitting compute, memory control and I/O improves yield and cost structure.
- Bigger racks. 96 accelerators act as one scale-up domain for serving large models.
- Lower precision. MXFP4 shrinks weights and KV-cache, which Microsoft says allows longer context and higher throughput at the same power.
Maia 200 already runs Microsoft Copilot in production, deploys in the US, Europe and Australia, and runs OpenAI, MAI and open-source models.
What it means for builders
- Cheaper Microsoft-hosted inference, eventually. If the 40% tokens-per-megawatt claim holds, Foundry and Copilot serving costs fall in 2027 and later — but Microsoft has not linked Maia to any price change.
- No direct access to plan around. Unlike Google TPUs or AWS Trainium, there is no Maia instance to rent today; your choice remains Azure GPU VMs or managed Foundry endpoints.
- More model-vendor leverage for Microsoft. Running OpenAI and open models on its own silicon reduces Microsoft’s dependence on NVIDIA supply.
For how custom silicon compares across clouds, see Anthropic on Azure Maia vs Amazon Trainium vs Google TPU and Google TPU vs Broadcom vs Marvell custom AI silicon.
What to watch
- Microsoft Ignite 2026 (November, San Francisco and online), where Microsoft said it would share more on its infrastructure strategy.
- Independent benchmarks once Maia 300 serves public traffic in 2027.
- Whether Foundry exposes Maia-backed endpoints with their own price tier.
Last verified: November 4, 2026.