Microsoft is preparing to move more of the artificial intelligence workloads behind its products onto its own silicon. The company announced Maia 300 on October 12, 2026, its third-generation custom AI accelerator, alongside plans to expand its Azure Maia infrastructure to multi-gigawatt deployments starting in 2027.

The move marks a significant step in Microsoft’s efforts to control more of the infrastructure powering its AI services, including Microsoft Copilot, GitHub Copilot, Microsoft Teams and Azure AI offerings. Rather than relying primarily on third-party graphics processing units (GPUs) for inference, Microsoft wants its Maia chips to become the primary system for running AI models across its products and services.

Microsoft says Maia 300 is designed to deliver 40% more tokens per megawatt than leading GPU-based systems. At the rack level, a fully configured Maia 300 system is expected to deliver up to seven times the token throughput of its predecessor, Maia 200. These figures are Microsoft’s own performance claims, rather than independently verified comparisons.

Production deployments of Maia 300 are scheduled to begin in 2027.

Maia 300 is built for the next generation of AI workloads

Maia 300 is designed around a growing challenge in AI infrastructure: running increasingly capable models efficiently after they have been trained. This process, known as inference, powers the responses users receive from chatbots, coding assistants, AI agents and other AI-enabled applications.

As AI systems move beyond answering individual prompts and begin handling longer conversations, processing large documents and completing multistep tasks, infrastructure needs to manage more than raw computing power. Memory bandwidth, data movement, networking and energy consumption all influence how efficiently a model can serve users.

Microsoft designed Maia 300 with these demands in mind. Its target workloads include long-context reasoning, latency-sensitive conversational experiences and sustained, multistep agentic tasks. These are particularly relevant as AI assistants evolve into systems that can plan work, use tools and execute actions across multiple steps.

The accelerator uses a multi-chiplet architecture that separates compute, memory-controller and input/output functions. Microsoft says this allows different parts of the chip to use manufacturing technologies suited to their specific jobs, improving efficiency and scalability.

The compute chiplets use TSMC’s N2P manufacturing process, while the broader design balances processing resources with memory and connectivity requirements.

Memory is another major part of the design. Each Maia 300 accelerator incorporates eight stacks of high-bandwidth memory (HBM4), providing 288GB of memory capacity and up to 18TB/s of peak bandwidth, according to Microsoft’s technical documentation.

The company is also introducing support for advanced low-precision formats, including enhanced MXFP4. These formats reduce the amount of data required to represent model calculations, potentially allowing larger models and longer context windows to run more efficiently within a given power budget.

The underlying idea is straightforward: an AI accelerator becomes more useful when it spends less energy moving data around and more time performing the computations that generate results.

A single rack can house 96 Maia 300 accelerators

Microsoft is developing Maia 300 as a complete infrastructure system, not simply an individual chip. Its rack configuration combines 96 accelerators in a system rated for 167 kilowatts of power, with 27.6TB of aggregate HBM capacity and 1.7 petabytes per second of combined memory bandwidth.

The system uses a custom networking architecture based on Ethernet technology to connect accelerators and coordinate workloads across racks. Microsoft says the design can scale to a superpod containing as many as 12,288 accelerators.

Liquid cooling handles more than 80% of the rack’s thermal load, while air cooling supports memory, storage and other components. This arrangement is designed to accommodate the high power density of modern AI infrastructure while giving Microsoft flexibility in how it deploys the systems across its data centers.

These details help explain why Microsoft’s efficiency claims extend beyond the silicon itself. The company is optimizing the accelerator, memory system, networking, cooling and software as a unified platform.

That matters because the cost of operating large AI systems depends on how much useful work a data center can deliver within its available electricity, cooling capacity and physical space.

Maia 200 is already running production AI workloads

Microsoft’s transition to custom silicon is already underway. Its second-generation Maia 200 accelerator is serving production workloads, including Microsoft Copilot, and supporting the development of the company’s own MAI models.

Microsoft says Maia 200 is being deployed across the United States, Europe and Australia. The company also reports that it delivers up to 1.7 times more tokens per watt for built-in AI capabilities in Microsoft Teams, Copilot and Cowork experiences compared with leading GPU infrastructure.

The company plans to extend its Maia infrastructure to more workloads, including GitHub Copilot, Microsoft Foundry and Dragon Copilot. These applications have different requirements, ranging from coding assistance and model serving to conversational experiences that need to respond quickly.

Maia 300 builds on Microsoft’s experience operating Maia 200 at scale. Its larger memory bandwidth, higher accelerator density and updated architecture are intended to support increasingly demanding inference workloads.

The company says a fully configured Maia 300 rack can deliver up to seven times the rack-scale token throughput of Maia 200. That is a substantial generational claim, although real-world results will depend on the models, workload configurations and operating conditions involved.

Microsoft wants more control over its AI infrastructure

The announcement is part of a broader shift across the AI industry. Microsoft, Google and Amazon are investing in custom silicon to better match their hardware to the workloads running across their cloud platforms.

For Microsoft, developing its own accelerators offers the opportunity to coordinate chip design with its models, software and data-center infrastructure. It can tune the entire system around the services it operates rather than depending exclusively on general-purpose accelerators designed by other companies.

Nvidia remains a major force in AI computing, and Microsoft’s announcement does not mean the company is abandoning third-party GPUs. Instead, its stated goal is to make Maia the primary inference system for its own products and services as deployments scale from 2027.

The distinction between training and inference is important here. Training involves teaching an AI model using large amounts of data, while inference is the process of running that model to produce outputs. Maia 300 is explicitly designed around inference, where improvements in performance per watt can translate into more AI interactions from the same energy budget.

Microsoft has not announced a retail price for Maia 300. The accelerator is intended for its cloud infrastructure rather than as a consumer product, so the more meaningful measures will be its deployment scale, operating efficiency and ability to support growing demand for AI services.

Microsoft plans to begin production deployments in 2027, when the Maia family is expected to become a much larger part of the infrastructure behind its AI ecosystem. The key question is how much of that demand its custom hardware can handle efficiently as AI assistants and autonomous agents become more widely used.