Anthropic’s Claude on Microsoft Foundry becomes generally available on NVIDIA GB300 in Azure

Anthropic’s Claude models, delivered through Microsoft Foundry and hosted on Microsoft Azure, have moved into general availability on NVIDIA’s GB300 Blackwell Ultra GPU infrastructure, according to NVIDIA. The change is positioned as a way for Azure-native enterprises to build and run more capable “agentic” AI systems, including autonomous and domain-specific sub-agents that can tackle multi-step business tasks.

In NVIDIA’s announcement, the company links improved inference performance and efficiency to lower total cost of ownership for enterprises deploying large-scale AI. “Inference” refers to the compute used to run a trained model on new inputs, rather than the training process. In NVIDIA’s framing, faster or more efficient inference can matter for both responsiveness and ongoing operating expenses when AI assistants are used continuously.

NVIDIA said Claude in Microsoft Foundry running on NVIDIA GB300 is deployed on NVIDIA NVL72 systems, paired with NVIDIA Quantum-X800 InfiniBand networking. It also says this configuration is intended to support “more powerful agentic systems,” including specialized agents that can operate across business domains.

A key part of the push is how developers can connect Claude to enterprise capabilities. NVIDIA said it is working with Anthropic to extend developer tooling by integrating NVIDIA tools into the Anthropic stack. The company described this as enabling “domain-specific abilities” for Claude agents through what NVIDIA calls NVIDIA verified agent skills, enabled by access to NVIDIA accelerated computing.

NVIDIA further pointed to deployment and governance as the practical hurdle for enterprise agents. Enterprises can run Claude agents on Azure using NVIDIA’s Secure Agent Workspace Reference Design, which NVIDIA describes as a blueprint for running autonomous agents in a governed environment. The design, as outlined by NVIDIA, aims to keep identity, network access, credentials, and runtime policy controlled at the infrastructure level.

The announcement also situates the GB300 availability within a broader partnership. NVIDIA said the expanded access is built on a strategic collaboration announced in November involving Microsoft, NVIDIA, and Anthropic, aimed at offering enterprise access to Claude and making Anthropic models available on NVIDIA-accelerated computing.

From a market perspective, the move underscores how the largest AI platform providers are pushing beyond chat-style assistants toward systems that can execute tasks with autonomy, while still being controlled inside corporate IT environments. For NVIDIA, adding GB300-backed capacity behind a major cloud distribution channel such as Azure is a way to align its newest data center platform with where enterprises are actually deploying models.

For Microsoft and its customers, the Foundry framing matters because it targets developers and enterprises already operating in Azure ecosystems. By making Claude available generally on NVIDIA GB300-backed infrastructure, the companies are effectively turning cloud compute choice and AI model access into a tighter bundle for building and scaling agent-based workloads.

Still, NVIDIA’s disclosure is largely focused on availability, architecture, and integration rather than performance benchmarks, pricing, or service-level details. The announcement does not provide specific latency, throughput, or cost figures for Claude on GB300, nor does it spell out the scope of “NVIDIA verified agent skills” beyond the concept of domain-specific abilities and the need for accelerated compute.

Looking ahead, what to watch is how widely enterprises adopt Claude through Foundry on GB300, and whether NVIDIA and Anthropic provide additional technical detail on the Secure Agent Workspace Reference Design, including how enterprises can operationalize identity and runtime controls for real-world agent deployments. Developers will also look for documentation that connects NVIDIA tools and verified agent skills to concrete application workflows in the Anthropic model stack.