The Pulse
Microsoft Makes Maia Its Primary Inference System
Microsoft says it will scale Azure Maia to multi-gigawatt deployments beginning in 2027 and make Maia its primary inference system for Microsoft products and services. Maia 200 is already handling production workloads, while Maia 300 is pla
AI.info Team · Written with AI ·
Multi-gigawatt Maia deployments start in 2027
Microsoft plans to make its Azure Maia accelerators the primary inference system for Microsoft products and services, scaling the hardware family to multi-gigawatt deployments beginning in 2027. The announcement, posted October 12 by Scott Guthrie, Microsoft’s executive vice president of Cloud + AI, sets out a broader role for the company’s in-house silicon than serving selected workloads.
“We will scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services,” Guthrie writes in the company’s announcement.
The target describes planned deployments, not capacity already installed or a disclosed amount of electricity consumed. Microsoft does not name the data centers or provide a rollout schedule beyond the 2027 start.
Maia 200 is already serving Copilot
Maia 200 is in production, running Microsoft Copilot workloads and supporting Microsoft’s MAI team as it develops models. Microsoft says the accelerator runs OpenAI models, its own MAI models and leading open-source model families. The company plans to add GitHub Copilot and Microsoft Foundry by the end of 2026.
Microsoft says Maia 200 is being deployed across the United States, Europe and Australia. For built-in AI capabilities in Teams, Copilot and Cowork experiences, it reports up to 1.7 times more tokens per watt than leading GPU infrastructure. That comparison applies to those listed experiences; the announcement does not present it as a general benchmark across all workloads.
Maia 300 targets longer contexts and agent tasks
Microsoft describes Maia 300, its third Maia generation, as an inference-first system for long-context reasoning, memory-intensive work and multi-step agent tasks. Production deployments are scheduled to begin in 2027. Its design combines accelerator silicon with memory, networking, software and rack infrastructure, rather than treating the chip as a standalone component.
The company says a full rack configuration of 96 accelerators delivers up to seven times the rack-scale token throughput of Maia 200. It also expects Maia 300 to provide 40% more tokens per megawatt than leading GPU-based systems. Those are Microsoft’s reported comparisons; its announcement does not describe an independent benchmark or specify the tested models and workload settings.
Memory and networking shape the system
Microsoft’s architecture post says each Maia 300 accelerator uses eight HBM4 stacks, with 18 terabytes per second of peak memory bandwidth and a 950-watt thermal design power. A 96-accelerator rack has a stated 167-kilowatt power envelope, 27.6 terabytes of HBM capacity and 1.7 petabytes per second of aggregate HBM bandwidth. These figures describe different levels of the system: the first set applies to an accelerator, while the rack totals combine all 96.
The architecture also uses multiple chiplets and a software-defined dataflow design intended to reduce data movement between compute, memory and networking resources. Microsoft says its rack design supports workloads ranging from long-context model serving to latency-sensitive conversations. The company’s technical post compares Maia 300 with Maia 200, reporting up to 2.5 times the performance per watt and up to seven times the rack-scale token throughput in a full 96-accelerator configuration.
Primary does not mean exclusive
Microsoft says Maia is to become the primary inference system for its products and services, but it does not say the company will stop using other chips. The announcement describes Microsoft’s platform as open, heterogeneous and able to use different models and silicon. Nor does it disclose Maia 300 pricing, customer availability, or whether Azure customers outside Microsoft’s own services will be able to reserve the planned capacity.
For now, the concrete rollout markers are Maia 200’s current production workloads and Maia 300’s planned 2027 production start. Microsoft has not specified how many gigawatts it expects to deploy in that first phase or which locations will host the systems.