Microsoft announced Maia 300, its next-generation custom AI chip, as part of a multi-gigawatt scale rollout beginning in 2027.

Microsoft will expand its Azure Maia AI chip system to multi-gigawatt scale from 2027, with Maia 200 live now and Maia 300 due next year.

Microsoft has unveiled its next-generation artificial intelligence (AI) accelerator chip, Maia 300. The tech giant says it will expand the Azure Maia family of AI systems to multi-gigawatt deployments starting in 2027. The move signals a push to reduce Microsoft's dependency on companies such as Nvidia and AMD for the raw computing power needed for AI tasks.

Along with this expansion, the company says Maia 300, the successor to the current Maia 200 chip, will become the main system for running AI inference across its products and services. Inference is the step in which a trained AI model answers questions and completes tasks for users. Microsoft says it will begin deploying Maia 300 chips in 2027.

**Maia 300**

Microsoft describes Maia 300 as a complete system built for new kinds of AI work, including long-context reasoning, fast conversational replies and long, multi-step tasks carried out by AI agents. The company expects it to deliver 40 per cent more tokens per megawatt than leading GPU-based systems. In a full 96-accelerator configuration, the system offers up to 7 times the rack-scale throughput of Maia 200.

The Maia 300 chip is designed based on experience gained from running Maia 200. Together, the two generations will support a range of services, including:

- Multi-step agentic work that helps GitHub offer coding tools more efficiently.

- Cloud-scale model serving that helps Microsoft Foundry handle many types of models and setups.

- Inference and data generation for the MAI team.

- Fast replies that help Dragon Copilot ease the paperwork load on clinicians.

Microsoft says AI is becoming a basic layer in every app and business process. Companies are moving beyond simple chat toward groups of agents that can reason, act and work together. This calls for infrastructure that is secure, efficient and able to change as models change.

Maia 300 is the third generation of the program. It has an inference-first design for long-context tasks that require large amounts of memory, and it balances computing power, memory speed, networking and energy use at cloud scale.

**Key features of Maia 300**

- **Less data movement:** Moving data, rather than raw computing power, is now the main bottleneck in large inference systems. Maia 300 uses a software-defined dataflow design and AI-assisted tuning to reduce data movement. Microsoft says this gives it a clear performance-per-watt advantage over traditional GPU designs.

- **More memory speed per watt:** Eight HBM4 memory stacks deliver a peak bandwidth of 18 terabytes (TB) per second within a 950-watt power limit. Built-in SRAM on the chip also reduces the load on main memory.

- **Chiplet design:** Compute, memory control, and input and output now sit on separate chiplets, which improves efficiency, scale and cost. The compute chiplets use TSMC's N2P process.

- **Larger context:** Support for low-precision formats, including an improved MXFP4, reduces memory and bandwidth requirements. This allows larger context windows and higher throughput within the same power limit.

- **Networking at data-centre scale:** Networking is part of the same dataflow design. Fast links between accelerators suit tensor parallelism, and a custom AI transport carries data between accelerators and racks. Using Ethernet-based networking, systems can scale from a single rack to an entire data centre.

- **More from every rack:** One Maia 300 rack holds 96 accelerators within a 167-kilowatt power limit. It offers 27.6 TB of HBM memory and 1.7 petabytes (PB) per second of total memory bandwidth. This lets Microsoft serve more users and larger jobs from each rack, improving the cost of growing its AI services.

**Future plans**

Microsoft links Maia to what it calls Frontier Transformation. The company says no single model, chip or idea will define the future. It aims to offer an open platform that supports many models and chips, with each part tuned from the silicon up to the software agent. Its stated goal is to build responsible, affordable and scalable tools that help people get more from AI.

Microsoft will share more on its strategy, technology trends and customer success at Microsoft Ignite 2026 next month.