Microsoft Unveils Azure Maia 300 AI Chip to Challenge Nvidia’s Dominance

Microsoft has unveiled its new Azure Maia 300 AI accelerator, a custom-designed chip intended to power the next generation of artificial intelligence services across its cloud computing platform. The company plans to begin deploying the hardware in 2027 as it expands its investment in large-scale AI infrastructure.

Announced on 12 October 2026, Maia 300 is the third generation of Microsoft’s in-house Maia chip programme. It is designed to handle demanding AI workloads, including complex reasoning, conversational AI and autonomous AI agents capable of carrying out multiple tasks.

Microsoft hopes the new processor will improve the efficiency of running AI models while reducing its reliance on third-party hardware suppliers, particularly Nvidia, whose graphics processing units are widely used in AI data centres.

Up to 40% more AI output per megawatt

Microsoft says Maia 300 is expected to deliver 40% more AI tokens per megawatt than leading GPU-based systems. Tokens are the units of text and other data that AI models process and generate, making throughput an important measure of how efficiently AI services operate.

The company also claims that a complete rack containing 96 Maia 300 accelerators can deliver up to seven times the token throughput of a rack based on its predecessor, Maia 200.

Each Maia 300 rack is designed to operate within a power envelope of approximately 167 kilowatts. The system combines 96 accelerators with high-bandwidth memory and specialised networking to support large AI models and serve more simultaneous requests.

The figures highlight Microsoft’s focus on improving performance without allowing electricity consumption and infrastructure costs to rise at the same rate as AI demand.

Advanced chip design and memory

Maia 300 uses a multi-chiplet architecture that separates computing, memory-controller and input/output functions. Rather than relying on a single large computing die, the design combines specialised components to improve efficiency and scalability.

The chip’s main computing components use TSMC’s advanced N2P manufacturing process. Microsoft has also incorporated eight stacks of high-bandwidth HBM4 memory, providing a claimed peak memory bandwidth of 18 terabytes per second.

This substantial memory bandwidth is important for AI inference, the process through which a trained model generates responses, produces content or carries out tasks. Moving data between processors and memory can become a major bottleneck when serving large models, particularly those working with lengthy conversations or extensive documents.

Microsoft says its design also supports advanced low-precision computing formats, helping reduce memory requirements and improve throughput.

Designed for Copilot and AI agents

The new hardware is intended to support a broad range of Microsoft’s AI services, including Microsoft Copilot, GitHub Copilot, Microsoft Foundry, Teams and Dragon Copilot.

These services increasingly depend on AI systems that do more than answer individual questions. Coding assistants, business agents and other AI tools may need to analyse large amounts of information, maintain context and complete multiple stages of work.

By optimising the processor, memory, networking and software together, Microsoft aims to deliver these capabilities more efficiently at data-centre scale.

Maia 300 builds on the Maia 200 platform, which Microsoft says is already being used for production workloads across several regions, including the United States, Europe and Australia.

Microsoft expands its investment in AI infrastructure

Microsoft plans to scale its Azure Maia systems to multi-gigawatt deployments beginning in 2027. The expansion reflects the enormous computing resources needed to operate increasingly sophisticated AI services.

Developing its own accelerators gives Microsoft greater control over hardware design and allows the company to optimise systems around its own workloads. It could also help diversify its supply of AI computing capacity as demand for Nvidia hardware and other specialised processors continues to grow.

However, Maia 300’s claimed performance and efficiency advantages are based on Microsoft’s own comparisons. Independent, real-world testing will be needed to establish how the system performs against competing hardware across different AI workloads.

The launch represents another step in the competition between major technology companies to build more powerful and cost-effective AI infrastructure. As AI services become more widely used, the ability to deliver greater computing capacity within the limits of available electricity, cooling and data-centre space is likely to become increasingly important.

Microsoft expects Maia 300 production deployments to begin in 2027, with further details about its hardware and infrastructure strategy expected at Microsoft Ignite in November 2026.