Microsoft Deploys Maia 300 Chips to Cut Nvidia Dependence in Azure Data Centers
Microsoft will run future proprietary artificial intelligence workloads primarily on its in-house custom accelerators starting in 2027, according to reporting by heise.de. The company plans to deploy large volumes of its third-generation Maia 300 chips within its Azure data centers to reduce its reliance on external hardware suppliers such as Nvidia.
The strategic shift follows the initial rollout of the Maia 100 in 2023 and the subsequent introduction of the Maia 200, which delivers 10 petaflops at FP4 precision. By moving infrastructure ownership to custom silicon, Microsoft aims to optimize performance specifically for proprietary services like Microsoft 365 Copilot, GitHub Copilot, and Microsoft Foundry, as goldesel.de reported.
Hardware Architecture and Performance Metrics of Maia 300
Fabricated using TSMC’s N2P production technology, the Maia 300 integrates two I/O chiplets, four memory-controller chiplets, and eight HBM4 stacks with 72 gigabytes of capacity each into a single package. According to specifications detailed by heise.de, the eight HBM4 stacks achieve a collective data transfer rate of approximately 18 terabytes per second. Each Maia 300 processor delivers 6.9 petaflops at FP8 precision and 13.8 petaflops at FP4 precision.
Four water-cooled Maia 300 accelerators fit into a single rack slot, with 24 slots forming a pod or rack containing 96 chips total. A standard Maia 300 rack generates seven times the tokens per second compared to a Maia 200 rack while drawing up to 167 kilowatts of power—a consumption profile comparable to Nvidia’s Vera Rubin NVL72 system housing 72 processors, as heise.de noted.
Deployment Timeline and Capital Investment Strategy
Microsoft intends to populate Azure data centers with Maia 300 pods representing several gigawatts of total electrical capacity starting in 2027. Operating at a power draw of roughly 1,000 watts per chip, the deployment scale will encompass several million custom accelerators, requiring billions of dollars in capital expenditure alongside companion purchases of hundreds of petabytes of HBM4 memory capacity, heise.de reported.
Despite the large-scale integration of proprietary hardware, goldesel.de noted that Microsoft plans to maintain external procurement relationships by continuing to deploy graphics processing units from Nvidia and AMD. Analysts suggest this dual-sourcing strategy allows the cloud provider to manage inference costs and supply chain constraints without fully abandoning third-party silicon.
Market Positioning Against Amazon and Google Cloud
Microsoft entered the custom AI chip market approximately three years ago, trailing competitors Amazon Web Services and Google Cloud, which have fielded multiple generations of proprietary processors such as AWS Inferentia, Trainium, and Google TPUs. While Microsoft previously relied heavily on rented Nvidia instances within Azure, the company is positioning its Maia platform to directly handle demanding inference workloads.
Performance comparisons provided by heise.de indicate that the preceding Maia 200 chip—produced on TSMC’s 3-nanometer N3P node with 272 megabytes of SRAM and 216 gigabytes of HBM3e memory—offered triple the FP4 processing power of third-generation AWS Trainium and outperformed Google TPU v7 at FP8 operations during internal benchmarking.