Anthropic Lands Claude on Azure's NVIDIA GB300 Blackwell Ultra

For the last year the Anthropic stack has effectively meant 'Claude on AWS', and the NVIDIA stack has effectively meant 'everyone except Anthropic'. The post on NVIDIA's blog on Monday is the first real crack in both of those defaults. Claude is now generally available inside Microsoft Foundry on Azure, running on NVIDIA GB300 NVL72 systems with Quantum-X800 InfiniBand networking, and the reporting describes it as Anthropic's first deployment on NVIDIA hardware.

The build-on point is the November strategic partnership between Microsoft, NVIDIA and Anthropic, but the news this week is the actual GA. Microsoft is pairing the launch with a performance pitch: per the Azure blog post, Claude Sonnet on GB300 generates tokens about 40% faster than on H100 nodes and roughly 15% faster than on B200, which is the kind of latency budget that starts to matter once you chain sub-agents into a workflow.

Around the model itself, NVIDIA is shipping two pieces that read more like enterprise plumbing than product. Verified Agent Skills add provenance and security checks at the capability layer, and the Secure Agent Workspace Reference Design is a blueprint for running autonomous agents in a governed environment where identity, network access, credentials and runtime policy are controlled at the infrastructure level rather than left to the application. If you have been stuck on the question of how to let an agent actually do things without giving it the keys to the kingdom, that is the bit worth reading first.

The honest caveat is that the 40% number is Microsoft's, measured in Microsoft's environment, on the model Microsoft is selling. It is plausible, since GB300 is a real generational step over H100, but treat it as a vendor benchmark until customer runs catch up. The reporting also does not pin down which exact Claude point releases are GA today, how pricing compares to Bedrock or the first-party API, or how the throughput numbers were normalized for batch size and context length.

The forward-looking read is that Microsoft now has a credible 'Claude or OpenAI, your choice' story inside Foundry, NVIDIA picks up a flagship agentic showcase for GB300, and Anthropic gains a second hyperscaler channel with Azure-grade governance attached. None of these parties needed this deal to survive, which is what makes it a useful tell about where enterprise agent deployments are actually heading.