Microsoft has added NVIDIA Vera Rubin NVL72 GPU systems to Azure virtual machines and opened up NVIDIA open‑model services through Microsoft Foundry, while also extending AI workloads to edge devices and Windows PCs. For engineers, this means access to larger‑scale GPU clusters, built‑in model customization, and confidential‑computing capabilities that can be evaluated for both cloud and on‑premise deployments.
GPU‑powered Azure VMs and benchmark results
Azure now runs Vera Rubin NVL72 systems in its data centers, targeting demanding reasoning, agentic applications, and large‑scale training and inference. The preview builds on a joint code‑sign and performance optimisation effort, reflected in recent MLPerf results: Azure scaled to 10,240 NVIDIA GPUs with GB200 NVL72 systems for the DeepSeek‑V3 training benchmark, and the GB300 NVL72 systems led cloud inference submissions on DeepSeek‑R1 with more than a 10% throughput advantage.
Open‑model customization via Microsoft Foundry
Microsoft Foundry now provides a serverless Programmed Training service for post‑training workflows and a Managed Compute layer that accepts custom model weights. The Nemotron 3.5 family is available, covering use‑cases from long‑running agents (Super) to low‑latency reasoning (Bolt, upcoming) and specialised tasks such as embedding, reranking, parsing, and content safety. An EY case study showed a 55% accuracy lift and a 15× cost reduction when adapting Nemotron models with Programmed Training, illustrating the practical benefit of open‑model fine‑tuning.
Confidential computing and power‑efficiency extensions
NVIDIA and Microsoft are extending confidential‑computing support to bare‑metal and rack‑scale AI infrastructure. The offering includes attestation mechanisms that let customers verify the runtime environment before exposing sensitive models, data, or keys. Additionally, the partners are collaborating on power‑efficiency through NVIDIA MaxLPS, which can help data‑center operators extract more performance per watt from the new GPU clusters.
Edge, on‑premise, and Windows PC deployment options
Beyond the public cloud, the collaboration adds several on‑premise pathways:
- Azure Local now supports NVIDIA RTX PRO 6000 Blackwell GPUs from hardware partners, enabling AI workloads on premises.
- Support for NVIDIA Jetson Thor and Jetson Orin platforms brings the same model stack to industrial edge scenarios such as machine vision and robotics.
- Physical AI models, including the Cosmos suite, are available in Foundry Local for on‑site inference.
- Windows PCs equipped with NVIDIA RTX Spark GPUs will demonstrate local large‑language‑model experiences, allowing developers to run AI directly on devices.
- GPU‑accelerated Windows 365 Cloud PCs gain Direct Stream capabilities, supporting up to two 4K streams at 60 fps with low latency; a limited preview is open now, with a public preview slated for early 2027.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Engineers should start evaluating Vera Rubin NVL72 VMs for any workload that approaches the scale of the reported MLPerf training runs, especially when latency or throughput is a primary concern. Teams interested in model customization can prototype with Foundry’s Programmed Training and Managed Compute, using the Nemotron 3.5 variants that match their latency or throughput profile. Where data sensitivity is a factor, explore the confidential‑computing attestation flow to ensure compliance before moving proprietary models to shared hardware. Finally, assess edge and on‑premise options—Jetson platforms, RTX PRO 6000, and Windows PC GPUs—to determine if moving inference closer to data sources reduces bandwidth costs or improves latency for your applications.



