Microsoft has announced Azure Maia 300, the third generation of its custom AI accelerator, as it prepares to expand its own AI infrastructure from 2027. The chip is designed to run AI models and services at cloud scale, including long-context reasoning, conversational experiences, and multi-step AI-agent tasks.
The announcement matters because the cost of running AI is increasingly shaped not only by model capability, but also by the power, memory, networking, and hardware needed to serve each response. Microsoft is betting that designing these pieces together can improve the economics of inference.
What Microsoft announced
In its October 12, 2026 announcement, Microsoft said it plans to begin production deployments of Maia 300 in 2027 and scale the Azure Maia family to multi-gigawatt deployments. The company described Maia as a primary inference system for Microsoft products and services.
Microsoft’s headline performance claims are:
- 40% more tokens per megawatt than leading GPU-based systems, according to Microsoft’s expectations.
- Up to seven times the rack-scale token throughput of Maia 200 in a full 96-accelerator configuration.
- A rack design containing 96 accelerators within a 167-kilowatt power envelope.
- Eight HBM4 memory stacks and a stated peak memory bandwidth of 18 TB/s per accelerator.
These are vendor-reported figures and comparisons, not independent benchmark results. The 40% claim compares Maia 300 with leading GPU-based systems, while the up-to-7x throughput figure compares a full Maia 300 rack configuration with Maia 200. They measure different things and should not be treated as interchangeable.
Why the chip is built for AI inference
Training a model and serving it to users are different workloads. Inference is the process of generating outputs after a model has been trained. At large scale, serving millions of requests can make memory bandwidth, data movement, networking, and power consumption major constraints.
Microsoft says Maia 300 uses a multi-chiplet architecture that separates compute, memory-controller, and input/output functions. Its design also emphasizes data locality, memory bandwidth, and lower-precision formats intended to reduce memory requirements for supported workloads.
Those choices are aimed at workloads such as long-context prompts and agents that perform several steps or tool calls. They do not, by themselves, prove that every Copilot interaction will become faster: real-world latency depends on the model, software, request size, serving configuration, and other infrastructure.
What this could mean for Copilot and Azure customers
Microsoft says its Maia infrastructure is intended to support Microsoft Copilot, Microsoft Foundry, GitHub-related coding experiences, Teams, and other AI workloads. Maia 200 is already serving production workloads, including Microsoft Copilot; Microsoft says GitHub Copilot and Microsoft Foundry workloads are expected to follow on Maia 200 by the end of 2026.
If Maia 300 delivers the efficiency gains Microsoft projects, it could help Microsoft serve more AI requests within a given data-center power budget and manage infrastructure costs as demand grows. It may also give Microsoft more control over how its hardware, models, and cloud services are optimized together.
That is a potential infrastructure benefit—not a promise of a specific Copilot price cut, a guaranteed improvement in response time, or a commitment that every Azure customer will use Maia 300. Microsoft has not established those outcomes in the announcement.
Does Maia 300 mean Microsoft will stop using Nvidia GPUs?
No such conclusion follows from this announcement. Microsoft is expanding its own accelerator program, but its cloud platform supports varied models and infrastructure. Custom silicon can reduce dependence on external accelerators for selected workloads without eliminating the need for other hardware.
The more useful question is how Microsoft allocates workloads between its own Maia systems and other accelerators, and whether the claimed efficiency translates into measurable customer-facing improvements. Independent comparisons and production results will be needed to assess that.
What to watch next
As Maia 300 moves toward planned 2027 deployment, the most useful evidence will be independent performance-per-watt testing, actual throughput under representative workloads, availability for Azure customers, and any documented effects on service latency or cost. Details about supported models, deployment regions, and customer access will also clarify how broadly the system can be used.
For now, Maia 300 is a significant infrastructure announcement, but its headline numbers should be read as Microsoft’s stated targets and system-level comparisons rather than proof of universal performance gains.