● Covered by 2 sources · 3 reportsMedium impact2 neutral1 positive
Microsoft to Deploy Maia AI Accelerators for Multi-Gigawatt Inference by 2027
Key points
- Microsoft will deploy Maia for multi-gigawatt inference starting 2027.
- Maia 200 is in production for Copilot and MAI team.
- Maia 300 is inference-first with 288 GB HBM4 and 18 TB/s bandwidth.
- Custom AI transport on Ethernet connects up to 12,288 accelerators.
Maia Accelerators for Microsoft Services
Microsoft plans to scale its Azure Maia family of AI accelerators to multi-gigawatt deployments beginning in 2027. This initiative positions Maia as the primary inference system for Microsoft's internal products and services. The Maia 200 is currently in production, supporting Microsoft Copilot and the MAI team, with GitHub Copilot and Microsoft Foundry scheduled for deployment by year-end across the US, Europe, and Australia.
Maia 300 Specifications and Performance
The Maia 300, set for production deployments in 2027, is designed for inference, specifically targeting long-context, memory-intensive, and decode-heavy tasks. Each Maia 300 accelerator includes eight HBM4 stacks, providing 288 GB of memory and 18 TB/s bandwidth within a 950-watt TDP. A single rack can house 96 accelerators, totaling 27.6 TB of HBM and 1.7 PB/s aggregate bandwidth. Compute chiplets use TSMC N2P, with separate dies for memory-controller and I/O functions, which Microsoft states increases compute FLOPs by 36% and memory bandwidth by 2.5 times compared to Maia 200.
Advanced Networking Architecture
Maia's networking architecture features a custom AI transport built on Ethernet, implemented by its AI network controller. This creates a unified fabric for communication across accelerators, allowing any peer in a superpod to initiate communication without host or discrete NIC involvement. Two accelerators are directly linked within a tray, a switched fabric connects 96 accelerators in a rack, and optical switches link up to 128 racks into a superpod of 12,288 accelerators. The design supports three latency domains for efficient tensor-parallel traffic management.
Efficiency Claims and Industry Response
Microsoft claims Maia 200 delivers up to 1.7 times more tokens per watt for built-in AI in Teams and Copilot compared to leading GPU infrastructure. However, the announcement's leading efficiency claims have drawn questions from practitioners due to their general nature, contrasting with the detailed architectural specifics provided for data movement and networking.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
The daily brief
One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.
One email a day. Unsubscribe in one click, any time.
Today's brief
Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.
▶ Play today's brief
New every morning, and the back catalogue is archived by date.
How outlets covered it
Microsoft announced plans to scale its Azure Maia family of AI accelerators to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services. Maia 200 is already in production, and the upcoming Maia 300 is designed for long-context, memory-intensive inference workloads, featuring a custom AI transport built on Ethernet.
Microsoft announced Maia 300, its next-generation AI accelerator, and Cobalt 200, an agent-native CPU. These new hardware components are designed to improve efficiency and performance for emerging AI workloads like long-context reasoning and agentic tasks.
Microsoft has released details on the architecture of its third-generation Azure Maia 300 AI accelerator, which is optimized for large language model (LLM) inference. The Maia 300 is designed as a complete system, co-designed with software, silicon, and infrastructure, to handle the demands of cloud-scale agentic inference workloads.