Unsplash
Microsoft has announced a large-scale expansion roadmap for its custom AI inference accelerator, the ‘Azure Maia’ series. According to an announcement on October 12, 2026, by Scott Guthrie, Executive Vice President of Cloud + AI, Microsoft plans to follow up its currently deployed ‘Azure Maia 200’ with the next-generation ‘Azure Maia 300’ starting in 2027, deploying it across data center infrastructure at a multi-gigawatt scale. This decision is a strategic move to center Microsoft’s primary AI service operations around its own custom silicon, effectively reducing reliance on external Graphics Processing Units (GPUs) and significantly improving the cost and power efficiency of rapidly growing AI inference workloads.
Mass Production of Azure Maia 200 and the 2027 ‘Maia 300’ Multi-Gigawatt Rollout
Microsoft confirmed that the Azure Maia 200 is already in active operation, handling production workloads for Microsoft Copilot and the development of next-generation models by the Microsoft AI (MAI) team. By the end of this year, the scope will expand to include developer tools such as GitHub Copilot and platform services like Microsoft Foundry. The Maia 200 is currently powering a wide range of models, including OpenAI’s latest AI models, Microsoft’s internal MAI models, and major open-source parameter models.
Looking ahead, Microsoft aims to introduce the next-generation ‘Azure Maia 300’ starting in 2027, scaling its AI infrastructure to a multi-gigawatt capacity. This initiative will solidify the Maia system as the core inference infrastructure for Microsoft’s key AI products and cloud services. Moving beyond simple chip supply, this system-level multi-gigawatt deployment is considered one of the largest self-reliance initiatives for AI silicon by a global tech giant to date.
40% Improvement in Tokens per Watt: Technical Differentiation to Reduce GPU Reliance
Based on the released technical specifications and performance targets, the next-gen Azure Maia 300 is designed as an intelligent system optimized for the latest AI workloads, including long-context reasoning, latency-sensitive conversational services, and multi-stage agentic operations. Notably, in terms of power-to-efficiency, the target is to improve token processing productivity by over 40% per megawatt (MW) compared to current market-leading GPU-based systems.
Furthermore, in a full 96-accelerator configuration per rack, the Maia 300 is expected to provide up to 7 times the token throughput at the rack scale compared to the existing Maia 200. Through deep hardware-software co-design, Microsoft has established a foundation to process real-time inference requests from millions of enterprise users with higher efficiency and lower costs.
Impact on Enterprise AI Service Costs and the Global Cloud Ecosystem
This announcement is expected to have significant economic implications for global enterprise customers grappling with high-performance GPU shortages and skyrocketing data center power costs. As tech giants continue to scale their AI models, the proportion of costs derived from inference—the actual service usage phase—is increasing exponentially compared to initial training costs.
As Microsoft expands its Azure Maia infrastructure to a multi-gigawatt scale, it is expected to significantly lower its own cloud service operating costs. This could lead to a stabilization of pricing for enterprise AI services and intensify the competition for custom accelerator self-reliance and cloud pricing among major players like Amazon Web Services (AWS) with Trainium/Inferentia and Google with TPUs. Simultaneously, the shift away from the external GPU-exclusive supply model and the growth of the custom ASIC market are accelerating the restructuring of the global semiconductor supply chain.
Implications for the East Asian ICT Industry and Data Center Market
The movement by global tech giants to develop custom AI accelerators and construct multi-gigawatt data centers has direct relevance to the ICT ecosystems in Korea, Japan, and the broader East Asian region. First, the mass production and deployment of high-performance inference chips like Maia 200 and 300 will continue to drive demand for High Bandwidth Memory (HBM) and cutting-edge foundry processes, closely linking to the supply chain strategies of major Korean semiconductor firms.
Second, as the transition to AI-integrated cloud services accelerates in the Japanese and Korean markets, the price competitiveness of Azure-based enterprise generative AI solutions is likely to improve. Particularly in regions where data center power consumption is a growing concern, the introduction of high-efficiency chip infrastructure—offering a 40% improvement in tokens per megawatt—presents a new alternative for companies looking to optimize Total Cost of Ownership (TCO) and improve energy efficiency.
Confirmed Facts and Unresolved Challenges
Confirmed Facts: Microsoft is currently running the Maia 200 for Microsoft Copilot and MAI workloads, with plans to expand with the Maia 300 at a multi-gigawatt scale starting in 2027.
Performance and Efficiency Goals: The Maia 300 targets a 40% improvement in token processing per megawatt compared to leading GPU systems and up to 7x throughput compared to Maia 200 in a 96-accelerator rack configuration.
Remaining Uncertainties: The exact semiconductor foundry process for the Maia 300, specific per-unit pricing, and the extent to which these standalone accelerator chips will be available for rental or use by general public cloud customers outside of Azure were not specified in the official documentation.
Official Sources and References
This article is based on official technical blog posts released by Microsoft on October 12 and September 28, 2026.