Microsoft is excited to announce a technology preview of Azure Confidential Computing on NVIDIA Vera Rubin NVL72, advancing our joint engineering work with NVIDIA toward future confidential AI offerings. We are starting with Confidential Bare Metal Instances (C-BMI), a first step that brings together customer control of the software stack and hardware-rooted protection for AI data and models while they are in use.
AI becomes more valuable when it can work with information that is not on the public internet: a patient’s history, a bank’s proprietary risk model, a citizen’s case file, or the business context an AI agent carries across systems. Using that information should not mean surrendering control of it.
In healthcare, private records can inform clinical workflows. In financial services, AI can detect financial-crime patterns while protecting customer data and proprietary models. Government assistants can help citizens navigate taxes and benefits using their own case context. Model developers need to safeguard weights and fine-tuning data as intellectual property. These uses call for protection on both sides of the interaction: the customer’s sensitive content and the model provider’s valuable assets.
Protecting sensitive AI without giving up control
Across these workloads, the challenge is not only where information is stored, but who can access it while AI uses it. Confidential computing adds hardware-backed protection for data and models during execution, reducing the trust customers must place in privileged infrastructure operators. Protection during execution complements policies such as zero data retention, which govern what a service keeps after processing.
Attestation provides signed evidence of the execution environment’s identity and measured security state. Combined with policies that release keys only to approved configurations, it gives data and model owners greater control over where their assets become usable. With carefully designed application and access policies, model providers can also use these protections to isolate their own operators from customer workload data. That isolation requires an end-to-end design; confidential hardware alone does not exclude authorized applications or tenant administrators.
For sovereign AI, these capabilities extend control beyond the location of a datacenter. Governments building national models or delivering public services need to determine who controls the keys, which environments may process sensitive information, and how those environments are verified. Confidential computing supports those requirements alongside residency, jurisdiction, and operational governance.
NVIDIA accelerated computing is essential for AI workloads, and the protections offered by confidential computing need to extend across these processors and interconnects. That is the foundation of our work with NVIDIA on Vera Rubin NVL72 and NVIDIA Confidential Computing.
A bare-metal foundation for rack-scale AI
Large models distribute computation across accelerators, making the interconnect part of the security architecture. NVIDIA describes Vera Rubin NVL72 as a rack-scale confidential domain spanning 72 NVIDIA Rubin GPUs, 36 NVIDIA Vera CPUs, and their interconnects, including NVIDIA NVLink and NVLink-C2C. Hardware-rooted device attestation provides evidence about those components. This platform architecture is the foundation for our work; the product demonstration at Microsoft Ignite has a narrower, single-node scope.
Confidential Bare Metal Instances (C-BMI) put the tenant’s host operating system, runtime and workload inside the trusted computing base (TCB), while placing the infrastructure operator outside that trust boundary. The tenant controls the software stack and must secure it accordingly. This differs from treating the host OS as untrusted: in this bare-metal model, the host kernel is part of what must be measured and trusted.
The C-BMI architecture combines CPU and GPU memory protection with encrypted physical links, including C2C, PCIe IDE, and NVLink. A dynamic root of trust for measurement (DRTM) reduces the tenant host’s TCB by keeping UEFI outside it. The host OS and workload remain trusted, so platform evidence and tenant software measurements must be evaluated together. Figure 1 brings these trust boundaries and protected connections together.
The tenant host is trusted; UEFI and the BMC sit outside its TCB. The demo uses local verification, not the remote service shown.
From platform evidence to a 397-billion-parameter model
The Ignite demonstration is designed around a single Vera Rubin NVL72 node with four Rubin GPUs. It begins by collecting CPU and GPU attestation evidence: a platform token carries platform firmware measurements, a tenant token carries host kernel measurements, and each of the four GPUs supplies firmware and configuration measurements. The signed evidence is rooted in a trusted NVIDIA root certificate. Together, these measurements describe the platform and tenant TCB underlying the AI workload, rather than merely reporting that a GPU is present.
A minimal local verifier checks signatures, challenges, and the binding between the platform and tenant tokens. It compares GPU measurements with NVIDIA known-good values and confirms that all four GPUs are in Confidential Bare Metal mode.
The workload sequence then loads a 397-billion-parameter model using vLLM with four-way tensor parallelism. One worker runs on each GPU. The planned flow checks that the model endpoint is healthy, shows the four workers and GPU memory usage, and submits an inference request that generates a response. The sequence links a measured confidential environment with a near-400-billion-parameter workload on four GPUs.
One foundation, more ways to build
Confidential BMI is the starting point for a broader deployment vision. Future work will enable Vera Rubin NVL72 support for confidential virtual machines, confidential containers through Azure Kubernetes Service (AKS) and Kata Containers with Confidential Containers (CoCo), and Microsoft Foundry. These are areas of development beyond the Confidential BMI technology preview, extending the design toward different levels of infrastructure control and application integration.
Across these operating models, the goal is to preserve control of sensitive assets from encrypted storage through computation, combining a verifiable environment with the data owner’s key policies and governance. AI should be able to work with the information organizations most need to protect. Our work with NVIDIA on Vera Rubin NVL72 is a step toward that future.
Learn more
Learn more about Azure Confidential Computing and its approach to protecting data in use.