On November 26, at the Microsoft Ignite conference, NVIDIA announced a broad expansion of its collaboration with Microsoft, jointly driving full-stack AI innovation from underlying chips to upper-layer services, and from cloud computing to enterprise applications. The core of this collaboration is to build and deploy large-scale AI computing infrastructure powered by the next-generation NVIDIA Blackwell platform, aimed at providing developers and enterprises with unprecedented AI performance and efficiency.
A major highlight of the partnership is Microsoft’s deployment of the new NVIDIA Spectrum-X Ethernet network platform in its landmark Fairwater AI data center (reputed as the largest and most complex AI factory ever built). This facility connects data centers in Wisconsin and Atlanta, Georgia, and will integrate hundreds of thousands of NVIDIA Blackwell GPUs for large-scale AI model training. Concurrently, Microsoft will deploy over 100,000 Blackwell Ultra GPUs globally for inference tasks, providing powerful computing capacity for critical workloads including OpenAI and Microsoft 365 Copilot.
For developers and enterprise users, the new Azure NCv6 series virtual machines equipped with NVIDIA RTX PRO Blackwell GPUs have entered public preview. These virtual machines extend the capabilities of the Blackwell platform, capable of providing precisely scaled acceleration for various workloads such as multimodal agent AI, industrial digitalization (via NVIDIA Omniverse), scientific simulation, and visual computing. Through Azure Local Edge, these capabilities extend from the cloud to the edge, supporting low-latency, real-time AI applications and sovereign AI solution deployments.
The collaboration has established “flexibly provisioned clusters.” Through continuous full-stack software optimization, composite performance improvements have been achieved across various stages of the AI lifecycle and on multiple NVIDIA architectures. This co-design significantly enhances efficiency and reduces costs. For example, thanks to optimization efforts, the end-user price for popular GPT models on Azure has dropped by over 90% within two years. Microsoft was recognized as an exemplary cloud provider for achieving 95% of the reference architecture performance on Hopper GPU training.
To leverage the value of enterprise proprietary data, NVIDIA is integrating its Nemotron open models and NIM microservices with Microsoft SQL Server 2025, delivering secure, scalable Retrieval-Augmented Generation capabilities directly within enterprise data stores. Simultaneously, the NVIDIA NeMo Agent Toolkit is now integrated with Microsoft Agent 365, enabling developers to build and deploy compliant, enterprise-grade AI agents directly within the Microsoft 365 application ecosystem, including Outlook, Teams, Word, and SharePoint. The Microsoft Foundry service also offers secure NIM models, including Nemotron and Cosmos, supporting the development of complex agents with multimodal, multilingual reasoning, and physical AI capabilities.
In the cybersecurity domain, the two companies are researching new adversarial learning models based on the NVIDIA Dynamo-Triton framework and the TensorRT toolkit, reportedly, according to the latest news, achieving performance improvements of up to 160 times compared to CPU-based solutions, helping enterprises defend against real-time threats.
In the areas of physical AI and industrial digitalization, the collaboration utilizes the NVIDIA Omniverse library on Microsoft Azure to unlock end-to-end re-industrialization processes. Robotics developers can use tools like NVIDIA Isaac Sim for simulation and testing. Companies such as Hexagon and Wandelbots are leveraging this stack to accelerate the development and deployment of their humanoid robots. Furthermore, both parties are committed to promoting the OpenUSD standard to enable seamless interoperability in 3D workflows, simplifying cloud-based simulation and digital content creation.
