Microsoft announced it will make its Maia 300 AI chip the cornerstone of its AI infrastructure from 2027, aiming to scale deployments measured in gigawatts and surpass current GPU performance in running AI inference tasks.
- Maia 300 boosts AI inference throughput by 40% per megawatt over leading GPUs
- Full rack output is up to 7x higher than Maia 200, Microsoft’s earlier AI chip
- Microsoft plans over one million Maia 300 chips deployed in 2027
What happened
Microsoft revealed plans to transition to its Maia 300 custom AI chip as the primary hardware for running AI workloads in its cloud services starting in 2027. The Maia 300 is the third generation of Microsoft’s internally developed AI processors and is built specifically for inference, the execution phase of AI models after training is complete. It targets applications involving extended conversations, complex reasoning, and AI agents performing multi-step tasks.
The company aims for large-scale deployments measured in gigawatts of power consumption, implying production and installation of potentially over one million Maia 300 chips within the next year. Each rack consists of 96 Maia 300 chips drawing about 167 kilowatts and is designed to deliver substantial performance improvements compared with previous generations and competing GPU-based systems.
Why it matters
Microsoft’s Maia 300 chip advance signifies a strategic step toward reducing reliance on third-party GPU suppliers like Nvidia, positioning Microsoft to optimize AI processing efficiency and accelerate innovation within its cloud platform. According to Microsoft, the Maia 300 offers roughly 40% more token output per megawatt of power compared to leading GPU systems, enhancing both cost-effectiveness and environmental sustainability in AI workloads.
The architecture of Maia 300 splits functions into modular chiplets for compute, memory, and data transport, employing TSMC’s advanced N2P fabrication process and HBM4 memory stacks. This design enables Microsoft to maximize performance per watt and per dollar while scaling massively. It aligns with recent industry trends where major cloud providers, including Google and Amazon, are developing proprietary AI chips to meet growing AI demand.
What to watch next
Microsoft plans to unveil further technical details and deployment strategies for Maia 300 at its upcoming Ignite conference in San Francisco, scheduled for late 2026. Tracking the actual scale of chip rollouts and their impact on the performance and cost of Microsoft’s AI services over the next year will be critical indicators of success.
Moreover, ongoing supply chain developments, particularly Microsoft’s collaboration with TSMC to secure manufacturing capacity for over 300,000 Maia 300 chips, will shape the timeline and scale of adoption. Observers should also monitor responses from competitors developing AI chips and how these investments affect the broader AI hardware market and cloud offerings.