Microsoft is preparing to significantly expand its in-house artificial intelligence infrastructure as demand for AI services continues to grow. The company plans to scale its Azure Maia accelerator family to multi-gigawatt deployments starting in 2027, with Maia 300 expected to play a major role in powering its next generation of AI workloads.
The announcement marks another step in Microsoft’s effort to develop its own AI hardware rather than relying entirely on third-party processors. Its Maia chips are designed to handle AI inference, the process through which trained models generate responses, analyse information and complete tasks for users.
Microsoft says Maia 200 is already running production workloads, including Microsoft Copilot. The company is preparing Maia 300 for more demanding applications, such as long-context reasoning, conversational AI and multi-step tasks performed by AI agents.
The expansion is intended to help Microsoft serve more AI requests efficiently while managing the energy and infrastructure costs associated with operating large-scale AI services.
Microsoft Maia 200 Is Already Running Production AI Workloads
Microsoft’s Azure Maia programme is designed to develop specialised AI accelerators that work closely with the company’s software, cloud infrastructure and AI models.
According to Microsoft, Maia 200 is now supporting production workloads, including Microsoft Copilot, as well as the development of models by its MAI team.
The company also plans to expand Maia 200’s use across additional services. GitHub Copilot and Microsoft Foundry are expected to follow by the end of the year, according to the announcement.
Maia 200 supports OpenAI models, Microsoft’s own MAI models and selected open-source model families. This allows Microsoft to use its hardware across different AI applications instead of designing infrastructure for just one model.
Microsoft says Maia 200 is being deployed across the United States, Europe and Australia. For built-in AI capabilities in Microsoft Teams, Copilot and Cowork experiences, the company reports up to 1.7 times more tokens per watt compared with leading GPU infrastructure.
Tokens represent units of text or other information processed by an AI model. Measuring tokens per watt helps indicate how much AI processing a system can deliver for a given amount of electricity.
These efficiency gains are important as Microsoft expands AI features across its cloud services, workplace software and developer tools.
Maia 300 Expected to Arrive in 2027
Microsoft’s next-generation Maia 300 accelerator is designed for the growing demands of advanced AI inference.
Unlike traditional workloads that mainly focus on running calculations, modern AI applications can require large amounts of memory, rapid data movement and sustained processing over many steps. These requirements become particularly important when models handle long conversations, analyse large documents or operate as agents that complete complex tasks.
Microsoft says Maia 300 has been designed around these requirements, with an emphasis on balancing computing power, memory bandwidth, networking and energy efficiency.
The company expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems. Microsoft also says a full configuration containing 96 accelerators can deliver up to seven times the rack-scale token throughput of Maia 200.
These are Microsoft’s own performance claims, and actual results may vary depending on the model, workload and system configuration.
The company plans to use Maia 300 as part of its broader AI infrastructure expansion beginning in 2027.
Key Technologies Behind Microsoft Maia 300
Microsoft has highlighted several hardware and system-level improvements designed to make Maia 300 more efficient when running advanced AI models.
1. Improved data movement
Large AI systems do not depend only on raw computing power. Moving information between processors, memory and other components can also limit performance.
Maia 300 uses a software-defined dataflow architecture alongside AI-assisted kernel optimisation. Microsoft says this approach is designed to reduce unnecessary data movement and improve how computing, memory and networking resources are used.
The goal is to spend more time performing useful AI calculations and less time moving data between different parts of the system.
2. High-bandwidth memory
AI models need fast access to information, particularly when processing long inputs or maintaining context across multiple interactions.
Microsoft says Maia 300 incorporates eight HBM4 memory stacks, with peak memory bandwidth of 18 terabytes per second within a 950-watt thermal design power rating.
High-bandwidth memory can help processors handle memory-intensive AI workloads more efficiently. Microsoft is also using specialised software techniques to make better use of on-chip memory and reduce pressure on external memory.
3. A multi-chiplet design
Maia 300 uses a multi-chiplet architecture that separates computing, memory-controller and input/output functions.
According to Microsoft, this design is intended to improve efficiency, scalability and cost structure. The compute chiplets are built using TSMC’s N2P manufacturing process.
A chiplet-based approach allows different parts of a processor to be organised into specialised components, potentially improving how a complex chip is designed and manufactured.
4. Support for larger AI contexts
Some AI applications need to process lengthy documents, extensive conversation histories or large amounts of information during a single task.
Microsoft says Maia 300 supports advanced low-precision formats, including enhanced MXFP4, to reduce memory and bandwidth requirements.
Lower-precision computation can help AI systems process more information using available hardware resources, although the results depend on the model and workload.
These capabilities are intended to support larger context windows and higher throughput without requiring a proportional increase in power consumption.
5. Faster networking between accelerators
Running large AI models often requires multiple accelerators to work together. Communication speed between those accelerators can influence overall performance.
Maia 300 incorporates high-bandwidth connections between accelerators and a custom AI transport system designed for communication across accelerators and racks.
Microsoft says its networking approach uses Ethernet-based infrastructure to support scaling from individual racks to larger data-centre deployments.
The aim is to make computing, memory and networking work together as a more unified system.
Microsoft Targets Higher AI Performance From Every Rack
Microsoft says its Maia 300 rack configuration brings together 96 accelerators within a 167-kilowatt power envelope.
The company lists the configuration as providing 27.6 terabytes of HBM capacity and 1.7 petabytes per second of aggregate HBM bandwidth across the high-speed scale-up domain.
This level of infrastructure density is intended to help Microsoft serve more simultaneous AI interactions and support larger workloads within its data centres.
For cloud providers, improving the amount of AI processing delivered by each rack can have a direct effect on infrastructure economics. Better utilisation can help them serve more customers without increasing physical capacity at the same rate as demand.
However, overall efficiency also depends on cooling, networking, memory usage, model size and how effectively the hardware is utilised.
How Maia 300 Could Support Microsoft’s AI Products
Microsoft plans to use its Maia infrastructure across several areas of its AI business.
Microsoft Copilot: The company’s AI assistant can benefit from infrastructure designed to handle large numbers of user requests while maintaining responsive performance.
GitHub Copilot: AI-assisted software development involves code generation, explanation and other tasks that can require repeated model interactions. Microsoft expects Maia infrastructure to support more efficient coding experiences.
Microsoft Foundry: The platform supports businesses building and deploying AI applications. Additional accelerator capacity could help Microsoft serve different models and enterprise workloads.
MAI models: Microsoft’s own AI team can use the infrastructure to develop and run its models.
Dragon Copilot: Microsoft’s healthcare-focused assistant could benefit from scalable AI infrastructure designed to support clinical workflows and reduce administrative workloads.
These applications have different technical requirements. Microsoft’s strategy is to develop a hardware platform that can support a range of AI models and services rather than depending on a single use case.
Why Microsoft Is Developing Its Own AI Chips
The development of Maia reflects a wider shift among major technology companies toward specialised AI hardware.
As AI usage increases, operating costs become an important consideration. Companies must pay for processors, data centres, networking equipment, electricity and cooling to keep their services running.
Developing in-house accelerators gives Microsoft more control over how its infrastructure is designed and optimised for its own applications.
It also allows the company’s hardware and software teams to work together from the early stages of development. Microsoft can use information from real-world AI workloads to influence future chip designs and improve performance for the tasks its customers actually run.
However, custom chips do not eliminate the need for other processors. Microsoft’s approach remains model- and hardware-diverse, allowing it to use different technologies depending on workload requirements.
What Microsoft’s Multi-Gigawatt Expansion Means for AI Infrastructure
The planned expansion highlights the enormous amount of infrastructure needed to support the next phase of AI adoption.
AI assistants are increasingly moving beyond answering individual questions. Businesses are exploring systems that can analyse information, use tools, generate code and complete multi-step workflows with limited human intervention.
These agent-based applications can generate repeated model requests and require computing resources over longer periods. As adoption grows, infrastructure providers will need to improve performance while keeping operating costs manageable.
Microsoft’s multi-gigawatt plan shows that the company expects demand for AI computing to remain substantial. At the same time, efficiency improvements will be important because adding more computing capacity also increases electricity and infrastructure requirements.
The ability to deliver more AI processing per unit of power could become an important competitive factor for cloud providers.
What Comes Next?
Microsoft expects Maia 300 to arrive in 2027 as part of its wider Azure Maia expansion. In the meantime, Maia 200 is already supporting production AI services and helping inform the development of the next generation of hardware.
The company’s progress will depend on how effectively it deploys the new systems, integrates them with its software platforms and delivers the promised performance improvements under real-world workloads.
Microsoft is also expected to share additional information about its AI infrastructure strategy and adoption trends at Microsoft Ignite 2026.
Conclusion
Microsoft’s plan to expand Azure Maia to multi-gigawatt deployments signals a long-term investment in specialised AI infrastructure. Maia 200 is already serving production workloads, while Maia 300 is designed to improve inference performance, memory efficiency and networking at scale. If the company delivers on its stated targets, the new hardware could help support the growing use of AI assistants, coding tools and enterprise AI applications while improving the efficiency of Microsoft’s cloud operations.