Michael JohnsonOn October 12, Microsoft said it would scale the Azure Maia family to several gigawatts starting in 2027. The company describes Maia 200 as part of its active production inference infrastructure: the chip serves Microsoft workloads, including Copilot, as well as OpenAI models, its own MAI models, and families of open-weight models. GitHub Copilot and Microsoft Foundry are expected to follow by the end of 2026, according to the company. Production deployments of Maia 300 are planned for 2027. Microsoft aims to make Maia its primary inference system for its products and services as part of an open, heterogeneous infrastructure spanning different models and types of silicon. The announcement did not specify a timeline for Maia availability to Azure customers.
Microsoft says Maia 200 is being deployed in the United States, Europe, and Australia. For the built-in AI features in Teams, Copilot, and Cowork, the company cites a separate estimate: up to 1.7 times more tokens per watt than leading GPU infrastructure. This figure applies to the Microsoft services named and its operating conditions.
Maia 300 is designed for inference and long context
Microsoft positions Maia 300 for resource-intensive inference: processing long contexts, latency-sensitive conversational tasks, and multi-step agent scenarios. The company describes an architecture that balances compute, memory bandwidth, network communication, and power consumption. Microsoft identifies data movement as a bottleneck in large inference systems; the company’s aim is for its software-defined data-processing architecture to improve utilization of compute, network, and memory resources.
A rack with 96 Maia 300 accelerators is designed for 27.6 TB of HBM and aggregate memory bandwidth of 1.7 PB/s, at a stated power draw of 167 kW. Microsoft estimates that a full configuration with 96 accelerators could deliver up to seven times the aggregate rack-level token throughput of Maia 200. Separately, the company forecasts 40% more tokens per megawatt than leading GPU systems. These estimates apply to a future configuration; production deployments of Maia 300 are planned for 2027.
What the bet on custom accelerators means
A custom accelerator lets a cloud provider co-design chips, networking, software, and service workloads. Microsoft presents Maia as an example of this approach, designed around models and applications operating at scale. For a cloud business, this approach could expand the choice of inference infrastructure and help serve more AI requests within data-center power limits. The announcement did not disclose how this would affect Azure pricing or the chip’s availability for customer workloads.
For teams choosing infrastructure now, it is important to distinguish Maia 200’s current production workloads from the future Maia 300. Microsoft has not announced a customer availability timeline for Maia 300, a list of specific cloud services, or pricing. When planning deployment, teams should check accelerator availability for the required model and service, as well as the region, latency, and cost for the specific workload.
Microsoft says Maia 300 has eight HBM4 stacks, peak memory bandwidth of 18 TB/s at a TDP of 950 W, and compute chiplets built on TSMC’s N2P process. The system also includes separate compute, memory controller, and I/O blocks, and uses Ethernet networking to scale from a rack to a data center. The company detailed the architecture in its technical article on Maia 300. The published details show how Microsoft is developing its own inference accelerators alongside GPU infrastructure.