SUSE and NVIDIA sell "sovereignty" for AI agents: perimeter control becomes the product

By iTWire · October 26, 2026.

On October 23, open source infrastructure company SUSE announced new capabilities that let enterprises run autonomous AI agents while keeping control over their operations and data. At the core is the integration of the NVIDIA OpenShell Reference Architecture into SUSE AI Factory with NVIDIA. SUSE is a launch partner for the reference architecture.

According to SUSE, the architecture combines three NVIDIA components. NVIDIA Vera Rubin handles model inference, meaning it runs the model to generate responses. Vera CPUs run the agents with OpenShell. BlueField-4 DPUs (processors dedicated to networking and security) run NVIDIA Sentry, which monitors agents from outside the host server and enforces policies in the silicon itself.

SUSE adds Kubernetes sandboxing and API gateways to this. Its supervisor proxies inspect agent traffic, authorize access according to defined policies and block unauthorized network requests to prevent code or data leaks. Private model serving runs on NVIDIA NIM microservices. The company says this removes the per-token fees of public third-party APIs and replaces them with a more predictable investment in capacity.

The announcement includes several other pieces. SLES for NVIDIA, a variant of SUSE Linux Enterprise Server 16, promises day-one support for new silicon such as Vera Rubin and a single support provider for the operating system, kernel and drivers. SUSE is also launching a certification program for independent software vendors (ISVs) with two tiers: "Ready for" (partner self-validation) and "Certified for" (jointly validated configurations). It names six partners, including ClearML, Tensormesh, Traefik Labs and CloudCasa. The company also announced a partnership with Fujitsu on a hybrid quantum-AI platform.

As a use case, the Finance Secretariat of Guanajuato (Mexico) is testing the platform in a proof of concept for AI-based public services. It is a pilot, not a production deployment. All performance and savings figures in the announcement are company claims; the press release provides no third-party measurements.

**Our take.** The news matters less for the specific product than for what it reveals. A year ago the debate about agents centered on how capable they were. Now the selling point is control: where the agent runs, what it can touch and who audits its traffic. Sovereignty has become a product category.

That makes sense. An agent that executes code and calls tools is a new attack surface. Data exfiltration through agents is a real near-term risk, far more tangible than distant superintelligence scenarios. Putting the monitoring on separate hardware (the DPU) rather than in the same software as the agent is a sound design choice: a compromised agent cannot switch off a guard running on a different chip. But it is an architectural promise, and its robustness will need to be proven through independent audits.

There is also a less appealing side. The entire stack is NVIDIA from top to bottom: inference, CPUs, DPUs and microservices. "Sovereign" here means your data stays inside your perimeter, not that you are independent of any vendor. Dependence on a cloud API is swapped for dependence on a hardware and software ecosystem that is closed in its key components. Real sovereignty generally requires the ability to switch vendors, and that is harder with such tightly integrated reference architectures.

The Gartner forecast cited in the release (that by 2027, 35% of countries will be locked into region-specific AI platforms using proprietary contextual data) points to fragmentation. It ties into a trend already visible elsewhere: what agents can reach is decided by platforms and perimeters, not just by the model. The web and AI infrastructure are splitting into walled gardens, some run by companies, others by countries.

On the cost model, caution is warranted. Moving from paying per token to paying for capacity is more predictable, but it only pays off with sustained volume; a company that buys Vera Rubin racks takes on the risk of underusing them. Abundant cheap compute is the long-term horizon, but today control is paid for with capital.

The underlying positive is that the industry is building the governance plumbing before agents become ubiquitous. If that layer is standardized and opened up, it will allow useful automation to be deployed, including in healthcare and public administration, without every company having to choose between capability and security. If it stays tied to a single vendor, we will have solved the data leakage problem at the cost of concentrating power. That is what needs watching.