At Microsoft Build 2026, Microsoft announced Foundry Managed Compute and Hugging Face models on Azure Foundry, a strategic initiative bridging open-source AI innovation with enterprise deployment. The integration places a curated catalog of Hugging Face open-weight models into the Foundry Model Catalog, enabling organizations to deploy production-grade AI workloads on managed GPU infrastructure with a single click. Azure Foundry consolidates access to frontier, open-source, and custom models through a single endpoint and unified software development kits. Alongside pay-per-token pricing and provisioned throughput, Foundry Managed Compute introduces a third deployment tier optimized for open-source models. Developers specify workload requirements such as parameter scale, context length, and performance priorities, while Azure automatically manages GPU topology, container updates, and runtime upgrades without manual redeployment. The integration addresses enterprise friction in open AI adoption. Microsoft operates a continuous curation pipeline that scans the Hugging Face ecosystem for trending models, verifying license compliance and conducting security audits. Models are screened to eliminate untrusted code, converted to SafeTensors format, and validated for API conformance. Pre-staged weights and CVE-scanned container images are stored in secure Azure storage, allowing deployments to operate entirely within private networks. Enterprise deployments leverage optimized inference runtimes matched to specific architectures. Large language models utilize vLLM or SGLang for high throughput and structured outputs, while embedding models run on Text Embeddings Inference. Quantized models operate efficiently via llama.cpp, and NVIDIA-accelerated workloads benefit from TensorRT-LLM. Specialized multimodal architectures deploy through hf-serve. All runtimes receive automatic security patches while preserving customer configuration policies. Deployed models integrate seamlessly with the Foundry Agent Service, providing built-in memory, knowledge grounding, and tool orchestration. Each model maintains standard observability features, including continuous evaluation and real-time monitoring. Deployment is standardized through predefined templates that pre-configure runtime settings and accelerator families, simplifying orchestration across multiple programming languages. All workloads route through a unified API compatible with the OpenAI SDK, ensuring consistent authentication and scoring. The Hugging Face Collection is currently available in public preview, supporting NVIDIA and AMD accelerators across global scopes. Microsoft plans to expand accelerator support and introduce Bring Your Own Weights functionality. By combining Hugging Faces ecosystem with Microsofts enterprise governance and networking infrastructure, the platform enables organizations to deploy and manage open-weight AI with the reliability required for mission-critical applications.