Microsoft is preparing to run more of its AI services on chips it designs itself.
The company unveiled Azure Maia 300 on Oct. 12, introducing its third-generation AI accelerator ahead of production deployments beginning in 2027. The system is designed to handle increasingly demanding AI inference workloads.
For MSPs and systems integrators serving Microsoft customers, the development could eventually influence the cloud AI services they deploy and manage.
More AI processing from each megawatt
Maia 300 is designed for AI inference, the process of generating responses or completing tasks after an AI model has been trained. Longer conversations and agents carrying out multiple steps need substantial processing capacity, adding to the growing demand for AI inference infrastructure.
Microsoft expects Maia 300 to deliver 40% more tokens — the units of text AI models process and generate — per megawatt than leading GPU-based systems. That would mean more AI processing within the same power budget.
A full rack of the new accelerators is also projected to deliver up to seven times the token throughput of the previous generation. The tech giant describes a redesigned chip architecture with improved data movement to help larger AI models run efficiently.
Both figures are Microsoft claims about the planned system. The announcement does not identify the GPU configurations or workloads used for comparison. Microsoft’s earlier Azure Maia 200 specifications provide a baseline for comparing the two generations.
| Feature | Maia 200 | Maia 300 |
| Production status | Already deployed | Production deployments planned for 2027 |
| Chip technology | TSMC 3nm | TSMC N2P compute chiplets |
| Memory per accelerator | 216 GB HBM3e | 288 GB HBM4 |
| Memory bandwidth | 7 TB/s | 18 TB/s |
| Accelerators per rack | 48 | 96 |
| Chip thermal design power (TDP) | 750W | 950W |
| Maximum scale-up domain | 6,144 accelerators | 12,288 accelerators |
| Microsoft service deployments | Copilot in production; GitHub Copilot and Microsoft Foundry planned by year-end 2026 | Planned use across Microsoft AI services |
While the newer chip has a higher individual power rating, Microsoft’s efficiency claims concern how much AI work the overall system can complete for the electricity consumed.
Existing Maia deployments set up the 2027 rollout
Maia 200 installations are expanding across the United States, Europe, and Australia. The hardware also runs models from OpenAI, Microsoft’s MAI family, and open-source developers, allowing the infrastructure to support different AI offerings.
For Maia 300, the official announcement outlines plans to expand the broader chip program to multi-gigawatt capacity. Microsoft expects Maia to become the primary AI inference system supporting its products and services.
The announcement does not specify the first Maia 300 deployment regions, a service-by-service rollout schedule, or related Azure AI pricing changes.
Partners should measure costs before promising savings
For MSPs, resellers, and systems integrators, a more efficient Azure datacenter does not automatically mean cheaper AI services for customers. Partners will still need to assess costs against the services and contract terms available to their clients.
If you manage or recommend Azure AI services, existing customer projects can provide a useful baseline for evaluating future improvements. Several steps can help partners prepare without making commercial assumptions.
- MSPs and cloud resellers should establish current costs. Record customer spending, service performance, and support requirements for existing AI deployments. Compare those figures against future pricing or service changes before revising customer proposals.
- Systems integrators should test representative AI workloads. Partners building AI agent solutions through Microsoft Foundry can measure response times, task completion rates, and token consumption. Results can help determine whether future performance improvements translate into meaningful benefits for customers.
- VARs and infrastructure advisers should clarify availability. Maia 300 is designed for Microsoft’s Azure datacenters, so customers evaluating cloud AI infrastructure options should continue comparing existing service capacity and contract terms rather than planning around a direct chip purchase.
Microsoft’s growing reliance on proprietary AI chips gives it greater control over the infrastructure behind Azure services. How much of that advantage reaches MSPs and their customers, however, remains largely in Microsoft’s hands.
More news: Park Place Technologies is extending its Microsoft cloud business across North America, creating more opportunities around Azure and managed services.