Maia 300 is Microsoft’s third-generation AI chip, set to begin handling production AI workloads in 2027. It is designed to support reasoning models, agentic workloads, and conversations that require faster responses.
Microsoft aims for Maia 300 to deliver 40% more performance than GPU-based systems per megawatt. In a system with 96 accelerators, it can achieve up to 7 times the token throughput of Maia 200. The goal is to make Microsoft more reliant on its own chips instead of primarily using GPUs from Nvidia and AMD.
Microsoft Is Building an AI System It Controls From Chips to Models
Maia 300 is designed to serve as the core system for processing AI workloads across Microsoft’s products and services, from Copilot and GitHub Copilot to Microsoft Foundry and the company’s own AI models.
This approach reflects Microsoft’s desire to control the entire AI stack, from silicon to models. The company believes that developing its own chips and co-designing them with its models will help create a platform suited to next-generation AI workloads and help it catch up with Google DeepMind, OpenAI, and Anthropic.
How Maia 300 Differs From Maia 200
Maia 200 launched in January and is already being used for production AI workloads, including Microsoft Copilot, GitHub Copilot, and Microsoft Foundry. Maia 300 was announced on October 12, 2026, and will begin rolling out to production in 2027.
| Factor | Maia 200 | Maia 300 |
|---|---|---|
| Announcement timing | January | October 12, 2026 |
| Usage | Already running production AI workloads | Begin rolling out to production in 2027 |
| AI models and workloads | Not specified by sources | Supports latest reasoning models and agentic workloads |
| Performance | Described as the best-performing AI infrastructure in the fleet | Tokens per megawatt increase by 40%, with up to 7 times rack-level throughput in a system with 96 accelerators |
When AI Needs to Respond Quickly, Think Deeply, and Work Continuously
Maia 300 is designed for AI workloads that do not end with a single question, including analysis based on long contexts, or long-context reasoning, as well as support for the latest reasoning models.
In conversational workloads, users want fast responses, while AI agents need to complete multiple consecutive steps. The chip therefore focuses on reducing latency and supporting more complex agentic workloads.
One Chip for Real-World AI Workloads in Azure
Maia 300 is positioned as the core system for AI inference across Microsoft’s products and services, building on the workloads Maia 200 already handles for Microsoft Copilot and GitHub Copilot.
In Azure, workloads will include Microsoft Foundry and Microsoft’s AI models, along with long-context reasoning, fast-response conversations, and multi-step agentic workloads. As it begins rolling out to production in 2027, the chip will play a role ranging from answering user questions to enabling AI to perform continuous sequences of tasks.
The Performance Metrics Microsoft Is Using to Support the Transition
| Factor | Maia 300 | Leading GPU systems |
|---|---|---|
| Tokens per megawatt | Expected to be more than 40% | Referenced as the baseline |
| Compared with Maia 200 | Up to 7 times | Not specified |
| Configuration | 96 accelerators | Not specified |
Microsoft’s figures indicate that Maia 300 focuses on reducing costs and improving the efficiency of AI workloads in the cloud, especially workloads that require continuous processing.
With 96 accelerators, the company says rack-level token throughput could reach up to 7 times that of Maia 200, making it suitable for services that must handle large numbers of requests and execute multiple consecutive steps.
How It Compares With Nvidia and AMD Based on the Available Information
Strategically, Microsoft is positioning Maia as the core system for AI inference across its own products and services, with plans to expand its use in Azure from 2027 onward to reduce reliance on chips from Nvidia and AMD.
In terms of performance, Microsoft says Maia 300 delivers 40 percent more tokens per megawatt than systems using leading GPUs. However, there is still no information about competing GPU models, pricing, or independent test results, so it is not possible to conclude that it outperforms Nvidia and AMD in every respect.
The New Chip Architecture Behind Its Efficiency
Maia 300 uses a chiplet architecture on TSMC’s N2P process, separating compute, memory controllers, and I/O.
This design increases the amount of silicon available for processing and allows each component to use the technology best suited to its function. It is therefore a key factor behind the chip’s focus on efficiency.
Maia 300’s Strengths and Limitations Based on Current Information
Pros
- +Microsoft expects to achieve 40 percent more tokens per megawatt than GPU systems
- +Supports reasoning models, agentic workloads, and conversational workloads that require low latency
- +Designed to work with Microsoft’s AI models and services
Cons
- −Will begin rolling out for production AI workloads in 2027
- −No pricing information or third-party test results are available yet
- −Details about use outside Microsoft’s ecosystem remain unclear
What to Watch Beyond the Performance Numbers
The major issue is the expansion of the Azure Maia family to multi-gigawatt-scale deployments starting in 2027. This will require examining production, installation, and real-world operation—not just chip performance.
Maia 300 relies on TSMC’s N2P manufacturing technology and uses a chiplet architecture that separates compute, memory controllers, and I/O. A crucial condition is that Microsoft must coordinate across silicon, hardware, and AI models to ensure the system genuinely meets the company’s workload requirements. The available sources do not yet provide direct financial cost information.
How Maia 300 Could Change Microsoft’s Role in the AI Competition
Maia 300 moves Microsoft from being primarily a cloud provider toward greater control over its AI systems. The company aims to make it the core system for inference workloads across its own products and services while reducing reliance on chips from Nvidia and AMD. If it meets its targets, the chip will deliver 40% more tokens per megawatt and, in a system with 96 accelerators, up to 7 times the rack-level throughput of Maia 200.
The key point is not just the launch of a new chip, but Microsoft’s effort to control the AI platform from infrastructure through to its own models. The real results will become clear when Maia 300 begins supporting production AI workloads in 2027.