Microsoft Unveils Third-Generation In-House AI Accelerator Maia 300
Microsoft says Maia 300 delivers 40% more tokens per megawatt than leading GPU-based systems and will scale in Azure data centers from 2027.
Microsoft unveiled its third-generation in-house AI accelerator, Maia 300, a custom-designed chip specialized for inference work on large AI models, according to ZDNet Korea.
Microsoft expects Maia 300 to process 40% more tokens per megawatt than leading GPU-based systems. A server rack with 96 Maia 300 accelerators provides 7 times higher rack-level token throughput than systems based on the previous-generation Maia 200.
Maia 300 was designed as an inference-focused architecture optimized for long-context processing, memory-intensive workloads and decoding-heavy AI tasks. It was co-designed with software, server racks, networking and data center infrastructure.
Deployment Plans
Microsoft plans to expand the Azure Maia family deployment to several gigawatts starting in 2027. Scott Guthrie said Microsoft aims to make the Azure Maia family a primary AI inference system for Microsoft products and services from 2027.
Currently, Microsoft uses Maia 200 to run AI services including OpenAI models in Microsoft 365 Copilot and Azure. Guthrie stated that from 2027 Microsoft will scale the Azure Maia family to several gigawatts as the main AI inference system for its products and services.
Quick answers
What is Maia 300?
Maia 300 is Microsoft's third-generation in-house AI accelerator, a custom-designed chip specialized for inference work on large AI models.
When will Microsoft deploy Maia 300?
Microsoft plans to expand the Azure Maia family deployment to several gigawatts starting in 2027.
How does Maia 300 compare to previous systems?
Microsoft expects Maia 300 to process 40% more tokens per megawatt than leading GPU-based systems, and a rack with 96 Maia 300 accelerators provides 7 times higher rack-level token throughput than systems based on Maia 200.