Anthropic Deploys Claude on NVIDIA GB300 in Microsoft Azure
Anthropic's Claude models (Opus 4.8 and Haiku 4) reached general availability in Microsoft Foundry on Azure on June 29, 2026, running on NVIDIA GB300 NVL72 Blackwell Ultra systems with Quantum-X800 InfiniBand networking, according to joint announcements from Microsoft, NVIDIA, and Anthropic. The launch builds on a strategic partnership the three companies announced in November 2025 and adds Azure-native enterprise controls: Microsoft Entra ID authentication, role-based access, zero-data-retention options, data-residency zones, and consolidated billing via Claude Consumption Units. NVIDIA says the integration also includes NVIDIA Verified Agent Skills and a Secure Agent Workspace Reference Design for infrastructure-level identity, credential, and runtime-policy controls. Microsoft cites early customers including Bolt, Momentic, and Everstar reporting production-scale throughput gains.
For enterprise ML and platform teams, this release removes a tradeoff that has slowed agentic AI adoption: teams no longer have to choose between frontier model quality (Claude) and Azure-native procurement, governance, and security controls. That combination, rather than the underlying GB300 hardware alone, is what Microsoft and NVIDIA are positioning as the actual unlock for moving agents from pilots into production.
What happened
Claude in Microsoft Foundry reached general availability on June 29, 2026, hosted on Azure and running on NVIDIA GB300 NVL72 Blackwell Ultra systems with Quantum-X800 InfiniBand networking, according to joint blog posts from Microsoft and NVIDIA. Two models are available at launch, Claude Opus 4.8 and Claude Haiku 4, accessible through the Messages API with prompt caching, extended thinking, and tool streaming, and through Foundry Agent Service for multi-step agent orchestration. The release builds on the Microsoft-NVIDIA-Anthropic strategic partnership announced in November 2025.
Technical context
The GB300 NVL72 is a rack-scale system pooling 72 Blackwell Ultra GPUs over NVLink and NVSwitch, paired with Quantum-X800 InfiniBand fabric; NVIDIA frames this as removing memory and interconnect bottlenecks for large-context, multi-agent workloads. NVIDIA also says the integration includes NVIDIA Verified Agent Skills and its Secure Agent Workspace Reference Design, giving enterprises a blueprint for running autonomous agents with identity, network access, credentials, and runtime policy controlled at the infrastructure level.
For practitioners
Microsoft's enterprise controls are the more consequential detail for platform teams evaluating adoption: Claude is billed through a single consolidated Claude Consumption Unit line on the Azure bill, authenticates through Microsoft Entra ID with standard Azure role-based access, supports Global and US data-residency zones, and offers a zero-data-retention mode where Anthropic does not retain prompts or completions after the API call. A model router can automatically send queries to the most cost-appropriate Claude model, which Microsoft says can cut costs up to 50% while improving user satisfaction. Anthropic remains the data processor and SLA provider for inference. Microsoft cites early production users, including Bolt (Fortune 500 throughput and reliability), Momentic (millions of tokens per minute for automated QA), and Everstar (a nuclear-industry safety-analysis workflow), though these are vendor-cited case studies rather than independently benchmarked results.
What to watch
Independent latency and throughput benchmarks for Claude on GB300 versus other Azure GPU tiers have not yet been published; watch for third-party comparisons, regional rollout of GB300-backed Foundry capacity beyond the initial data zones, and pricing detail once Claude Consumption Unit billing is broadly used in production. The Verified Agent Skills ecosystem is new, so its actual adoption and governance track record are also worth tracking before treating it as a settled standard.
Key Points
- 1Claude Opus 4.8 and Haiku 4 reached general availability in Microsoft Foundry on Azure on June 29, 2026, running on NVIDIA GB300 Blackwell Ultra hardware.
- 2Azure-native controls - Entra ID auth, zero data retention, and consolidated Claude Consumption Unit billing - aim to remove procurement and governance barriers.
- 3Microsoft's model router can cut costs up to 50 percent by auto-routing queries to the most cost-appropriate Claude model for each task.
Scoring Rationale
A three-way official launch (Microsoft, NVIDIA, Anthropic) with named production customers and concrete enterprise controls (billing, governance, data residency); notable for practitioners deciding where to run agentic Claude workloads, but a product-availability and partnership story rather than a research or capability breakthrough.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
