Claude in Microsoft Foundry on Azure: GA on NVIDIA Blackwell GPUs

Azure’s mid‑July wave of updates deserves attention not because one item is flashy, but because several changes together reshape the operational surface for cloud platform teams. The single most consequential: Claude is now GA in Microsoft Foundry on Azure, running on NVIDIA Blackwell Ultra GPUs (Foundry GA announced mid‑July 2026). Pair that with Foundry’s preview of Fireworks AI and GPT‑style preview models and Microsoft’s Agent Framework — platform teams aren’t just serving LLMs anymore; they’re hosting multi‑agent, GPU‑heavy orchestration with real production constraints.

The new operational surface: Foundry + agents

Foundry GA for Claude (and preview support for Fireworks AI and GPT‑style preview models) means low‑latency inferencing on NVIDIA Blackwell‑series GPUs is now a supported Azure path. What matters for operators: this is an explicit signal that Azure expects customers to run agent‑centric architectures close to managed GPUs, not just query single LLM endpoints. The Microsoft Agent Framework patterns being published in preview push multi‑agent choreography — agents that call other agents, systems, or even run ephemeral code. That’s powerful, and it creates two hard problems at scale: telemetry partitioning and trust boundaries.

Microsoft did not ignore those problems. Application Insights introduced a GenAIContent table with access controls in preview to help separate sensitive generative AI telemetry, and Microsoft Defender for Cloud security assessments for Azure Database for PostgreSQL Flexible Server reached GA (July 17, 2026). Those are necessary moves. They’re not sufficient. My take: Foundry on Blackwell‑series GPUs is the right call for throughput and latency demands, but platform teams must treat agent orchestration as a first‑class attack surface and assume telemetry will contain PII unless it’s explicitly isolated.

Stateful AKS gets sane defaults

Across the stack, AKS delivered one of the quieter but more operationally important changes: encryption in transit for Azure Files NFS shares is GA (July 17, 2026). For teams running stateful workloads with NFS PVs — databases, ML feature stores, and legacy apps — this finally moves encryption‑in‑flight from manual network engineering into the platform’s default posture. Expect minor churn: CSI drivers and NFS client mount options may need configuration tweaks, and there’s a measurable CPU/network latency trade‑off on heavy IO patterns. Still: this should be enabled by default for production clusters.

You can read more on the AKS changes in our piece on AKS encryption in transit for Azure Files NFS mounts: AKS: encryption in transit for Azure Files NFS mounts.

IPv6, metrics, and the visibility you actually need

Two networking/observability changes complete the picture. Azure VPN Gateway added IPv6 support (GA), enabling dual‑stack and IPv6‑only hybrid connectivity — useful for modern device fleets and interop with IPv6‑first on‑prem networks. Azure Monitor launched advanced platform metrics in public preview (July 15, 2026), surfacing richer service‑level telemetry that’s directly useful for SLOs, anomaly detection, and cost attribution.

Advanced platform metrics will change how you write SLOs for managed services: higher cardinality, platform‑level counters mean fewer blind spots, but they also mean more data and more decisions about retention and ingestion costs. Do not treat these as optional observability sugar; they are the data you’ll need to debug multi‑agent behaviors on Foundry or cross‑service networking issues when IPv6 paths diverge from IPv4.

What platform teams should actually do next

Treat Foundry as a GPU + orchestration platform, not just an LLM endpoint. Plan network, cost, and observability for multi‑agent flows.
Enable AKS NFS encryption‑in‑transit for production stateful workloads; benchmark IO patterns and adjust CSI options.
Use Application Insights GenAIContent isolation (preview) for any telemetry that could contain model inputs/outputs; if you don’t, you’ll regret it when compliance audits start.

Final thought

This is a coherent release set. Microsoft is gluing high‑performance model serving, agent orchestration patterns, tighter DB assessments, and the observability primitives teams need to run them. That’s overdue. The real challenge for platform engineers isn’t whether these features exist — it’s operationalizing them together without creating new leak paths or runaway GPU bills. If you’re building agent‑first services on Azure this year, start rethinking telemetry boundaries and SLOs before your first multi‑agent chaos incident.

Claude in Microsoft Foundry on Azure: GA on NVIDIA Blackwell GPUs

Sources

AKS: Azure CNI Overlay GA with Application Gateway Integration and Layer 7 Policies GA

Microsoft Foundry Adds Anthropic Claude (GA), GPT-4o Support, and Foundry Agent Service GA

AKS: encryption in transit for Azure Files NFS mounts