Microsoft plans Maia 300 production and wider inference deployment for 2027

Current production and the next generation

Microsoft's official announcement says Maia 200 is serving production workloads, including Microsoft Copilot, while Maia 300 production deployments will begin in 2027. The company also plans multi-gigawatt deployments of the Maia family starting that year.

Maia 300 is designed for inference workloads involving long context, conversational latency and sustained agent tasks. Microsoft describes the program as co-design across hardware, software and its own service workloads.

Read efficiency targets at system scale

Microsoft expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems. It also describes up to seven times Maia 200's rack-scale token throughput in a 96-accelerator configuration. These are the company's claims and expectations, rather than Franklin measurements.

The announcement emphasizes memory, data movement, networking and low-precision formats. Its 96-accelerator rack description uses a 167-kilowatt power envelope. Those system details help explain the comparison's scope, but they do not directly predict a customer's application response time or cloud bill.

Connect infrastructure claims to service outcomes

Franklin analysis: tokens per megawatt is relevant to constrained datacenter capacity, while a user also needs acceptable latency and task quality. A throughput claim becomes more useful when the workload, model, context length and power accounting are available beside it.

For planning, keep Maia 200's current production role separate from Maia 300's 2027 deployment schedule. Then identify the actual service and workload a project would use, rather than treating a hardware announcement as proof of current customer availability.

Microsoft's roadmap supports following those future deployment and performance disclosures. It does not yet establish a measured improvement for an arbitrary application. A useful subsequent comparison would report quality and latency under the same workload while making the complete system configuration visible.