Microsoft and NVIDIA Are Building a Massive AI Superfactory
Microsoft and NVIDIA are teaming up in a big way to power the next generation of AI. Timed with the Microsoft Ignite conference, they announced a huge expansion of their collaboration that covers everything from data centers and GPUs to AI agents in Microsoft 365 and industrial robots.
At the center of all this is Microsofts new AI Superfactory. It connects the Fairwater data center in Wisconsin with a new facility in Atlanta, Georgia. Together they form one giant AI engine designed to train and run some of the most advanced models on the planet.
Under the hood, Microsoft is integrating hundreds of thousands of NVIDIA Blackwell GPUs for large scale training, plus more than 100,000 Blackwell Ultra GPUs in NVIDIA GB300 NVL72 systems worldwide for inference. This is the hardware that will power heavy hitters like OpenAI, the Microsoft AI Superintelligence team, Microsoft 365 Copilot and Microsoft Foundry services.
Customers such as Black Forest Labs are already using NVIDIA GB200 NVL72 systems to train next generation FLUX multimodal models for visual intelligence. To keep all of this connected, Microsoft is deploying NVIDIA Spectrum X Ethernet switches in its Fairwater AI data center. These switches are built to handle massive AI clusters with the speed and efficiency required for large scale training and inference.
On top of that infrastructure, Microsoft is rolling out new Azure NCv6 Series virtual machines powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. These VMs, now in public preview, provide right sized acceleration for workloads like multimodal AI agents, industrial digital twins using NVIDIA Omniverse, scientific simulations and high end visual computing.
The same GPU tech is not just locked in the cloud. Using Azure Local, enterprises can run these AI workloads closer to where their data lives in local data centers, factory floors or secure edge locations. That means lower latency and real time AI for places that cannot send everything to the public cloud.
- Train and run multimodal AI models at massive scale
- Build and deploy digital twins for factories and industrial systems
- Run high end visualization and simulation on premises or at the edge
Fungible AI Fleets and Cheaper Models
A big focus of the partnership is what NVIDIA and Microsoft call a fungible AI fleet. In simple terms, that is a flexible pool of GPUs and software that can run many different workloads efficiently instead of having rigid, single purpose hardware.
This is possible because the two companies are constantly tuning the full software stack on Azure for NVIDIA Blackwell and Hopper GPUs. They are optimizing everything from drivers and libraries to compilers and inference runtimes to squeeze out more performance over time.
These optimizations power the latest models from the Microsoft AI Superintelligence team, including:
- MAI 1 preview for text
- MAI Voice 1 for real time voice
- MAI Image 1 for high fidelity image generation
These models show up in products like Bing Image Creator and Microsoft Copilot, so the improvements are not just theoretical. They directly impact how fast and how cheaply users can run AI features.
Continuous tuning has already had a huge effect. NVIDIA notes that this work contributed to more than a 90 percent drop in the price of popular GPT models for end users on Azure in just two years. That makes previously unrealistic AI projects suddenly affordable for many organizations.
The optimization push also extends to Microsoft Foundry, where the NVIDIA TensorRT LLM library is used to boost throughput, reduce latency and cut costs for a range of popular open models. To measure and improve performance, the companies rely on the NVIDIA DGX Cloud Benchmarking suite. By hitting 95 percent of the reference performance for H100 training, Azure has been recognized as an Exemplar Cloud by NVIDIA.
All of this optimization work is not limited to generative AI. It also helps with data processing, vector search, databases, digital twins, scientific computing and 3D design. The idea is that once you have a well tuned AI fleet, you can throw almost any heavy workload at it.
From Enterprise Data to AI Agents and Physical AI
The partnership is also about bringing AI directly into the tools and data that enterprises already use. A big example is Microsoft SQL Server 2025. NVIDIA is integrating its Nemotron open models and NIM microservices into SQL Server so that companies can run retrieval augmented generation directly on their own data, in the cloud or on premises.
This approach keeps data close to where it is stored, uses GPUs efficiently and is designed for security and scale. It turns SQL Server into more than a database. It becomes an AI powered engine that can answer complex questions over enterprise data.
NVIDIA and Microsoft are also pushing deeper into agentic AI for everyday work. The NVIDIA NeMo Agent Toolkit now connects with Microsoft Agent 365. This lets developers build enterprise ready AI agents that plug into Microsoft 365 apps like Outlook, Teams, Word and SharePoint while respecting compliance and governance needs.
To power these agents, Microsoft Foundry offers NVIDIA Nemotron models for digital AI and NVIDIA Cosmos models for physical AI as secure NIM microservices. Developers can use them to build agents with multimodal understanding, multilingual reasoning, math and coding skills and even physical world awareness.
Security is another major theme. The companies are collaborating on new adversarial learning models built with NVIDIA Dynamo Triton and TensorRT tools. These models are designed to detect and respond to cyber threats in real time, claiming up to 160 times speedup compared with CPU based methods.
On the physical side, NVIDIA and Microsoft are enabling what they call physical AI. Through NVIDIA Omniverse libraries on Azure, developers can build and run full digital twins of factories, products and industrial systems. Partners like Synopsys, Sight Machine and SymphonyAI are using these tools for simulations, real time analytics and 3D visualization.
Robotics developers get access to NVIDIA Isaac Sim on Azure for synthetic data generation and simulation driven testing. Companies like Hexagon are building humanoid robots using the full NVIDIA robotics stack, while Wandelbots NOVA uses Isaac Sim and Isaac Lab to speed up the jump from simulation to real world deployment.
To tie it all together, NVIDIA and Microsoft are using a standardized digital engineering approach for OpenUSD, enabling smooth interoperability across 3D tools and workflows in the cloud.
This expanded collaboration also links to a separate partnership with Anthropic, focused on optimizing Anthropic models for NVIDIA hardware and tuning future NVIDIA architectures for Anthropic workloads. In other words, the ecosystem around Microsoft and NVIDIA is getting even bigger.
Original article and image: https://blogs.nvidia.com/blog/nvidia-microsoft-ai-superfactories/
