Microsoft targets AI self-sufficiency with MAI-1 model on 15,000 GPUs
Microsoft’s AI chief announced plans for AI self-sufficiency by developing advanced in-house foundation models and reducing reliance on OpenAI while preserving Azure API exclusivity through 2032. In August, Microsoft previewed the MAI-1 model trained on about 15,000 NVIDIA H100 GPUs and unveiled a Maia 200 inference chip to challenge Nvidia’s GPU dominance.
1. Strategic Shift to AI Self-Sufficiency
Microsoft’s new AI strategy emphasizes reducing dependency on third-party models by building its own foundation models, chips and software. The company retains OpenAI as a “frontier model partner” with IP rights and Azure exclusivity through 2032, but aims to negotiate more flexibly and host any winning models on its platform.
2. MAI-1 In-House Model Development
In August 2025, Microsoft unveiled MAI-1, a mixture-of-experts model pre- and post-trained on roughly 15,000 NVIDIA H100 GPUs for Copilot text use cases. This preview demonstrated Microsoft’s commitment to large-scale model training and its ability to deploy in-house AI solutions across Microsoft 365.
3. Maia 200 Chip and Multi-Vendor Hosting
Microsoft introduced the Maia 200 inference accelerator, designed to lower AI token generation costs and rival Nvidia’s CUDA ecosystem. Simultaneously, Azure now hosts models from Meta, Anthropic, Mistral and others, ensuring customers can access top-performing AI regardless of vendor while Microsoft promotes its own silicon.