SUSE Positions Hardware Choice as a Core Part of Sovereign AI

SUSE is positioning hardware portability as a core element of sovereign AI, allowing enterprises to use Nvidia accelerators today without rebuilding their AI infrastructure if they switch to another silicon platform. The company launched SUSE AI Factory with Nvidia in April and announced support for AMD silicon at AMD Advancing AI 2026, validating the AMD Instinct MI350P accelerator. Rhys Oxenham, VP and general manager, AI, at SUSE, said full sovereignty at every layer isn't realistic, but architectural choice at each layer is.

SUSE is positioning architectural choice as a key element of sovereign AI. It argues that enterprises should be able to use Nvidia accelerators today without having to rebuild their AI infrastructure if they later switch to another silicon platform. This hardware portability reflects a broader shift in how SUSE defines sovereignty—not as eliminating external dependencies, but as giving enterprises the flexibility to change or replace those dependencies. “The honest answer is every organization needs to decide its own risk profile on this spectrum,” Rhys Oxenham, VP and general manager, AI, at SUSE, said. “Full sovereignty at every layer isn’t realistic today, but architectural choice at each layer is.” SUSE launched AI Factory with Nvidia https://www.eetimes.com/suse-nvidia-launch-ai-infra-for-enterprise-ai-deployment-and-sovereignty/ in April as an infrastructure stack for enterprises building and running AI workloads across data center, edge, and cloud environments. For customers using SUSE AI Factory with Nvidia, the platform runs on SUSE Linux Enterprise Server and SUSE Kubernetes Engine, with Nvidia AI Enterprise embedded inside SUSE’s governance and security layer. View All https://www.eetimes.com/category/sponsored-content/ According to SUSE, this allows enterprises to keep their AI workloads within customer-controlled infrastructure rather than relying on third-party APIs. “Customers choose a dependency on Nvidia at the accelerated compute layer itself, specifically CUDA, the GPU operator, and the NIM microservices that ship as part of Nvidia AI Enterprise,” Oxenham said. “That is a deliberate, pragmatic tradeoff for top-tier speed and performance, with a pre-validated, fast-to-deploy stack rather than something hand-built from scratch.” An enterprise using SUSE AI Factory with Nvidia is therefore still dependent on Nvidia at the accelerated-compute layer. SUSE’s argument is that this dependency does not necessarily extend through the rest of the infrastructure stack. The company said the underlying AI Factory architecture remains hardware agnostic, allowing organizations to move to other accelerator ecosystems without replatforming. The company is also working with organizations running accelerator platforms other than Nvidia. At AMD Advancing AI 2026, SUSE announced support for AMD silicon and its GPU operator within SUSE AI Factory. According to AMD’s technical blog, SUSE has validated the AMD Instinct MI350P accelerator alongside its software blueprint engineering work. Linux and Kubernetes provide the abstraction layers above the hardware. “Organizations don’t need to replatform, redesign cluster topologies, or rewrite operational playbooks just because they are introducing a new accelerator,” Oxenham said. The changes are primarily limited to silicon-native dependencies at the bottom of the stack, at the Linux layer. These include drivers and GPU operators, along with corresponding runtime libraries such as AMD’s ROCm and Nvidia’s CUDA runtime environments. SUSE said the broader AI software ecosystem can support both implementations, which can reduce the changes required when customers move between accelerators. The approach is intended to allow workloads to run across different accelerator platforms as supply, cost, and performance requirements change. Managing vs. eliminating dependency SUSE applies the same principle beyond silicon to AI models and APIs. Enterprises can use proprietary AI models and services, but SUSE argues that they should avoid dependencies that make it difficult to switch later. “If, as an organization, you are building against proprietary, you are becoming much more locked into that implementation and you subject yourself to two core problems,” Oxenham said. The first is cost. If API prices increase, organizations remain tied to those services, and costs rise with usage. The second is access. “As we saw in a number of different events over the last few months, certain models were restricted for certain jurisdictions,” Oxenham said. “If you had built your intelligence based around a particular implementation, then you would have been immediately shut off from that implementation.” Oxenham also pointed to the growing capability of open-weight models as another way to reduce dependence on proprietary AI services. “The capabilities of some of the latest open-weight models, when benchmarked against the latest frontier models, are absolutely incredible,” Oxenham said. He said the narrowing capability gap makes private enterprise AI viable while allowing enterprises to maintain cost control and keep their AI workloads under their control. For SUSE, the choice of AI models and services is therefore part of the same sovereignty question as the choice of silicon—enterprises need the ability to change what they depend on. India moves AI from experimentation toward production In India, SUSE sees the sovereignty discussion alongside a growing need to move AI workloads from experimentation into production. SUSE identifies two customer categories for its AI Factory platform in India—service providers and end users. Service providers have traditionally provided infrastructure-as-a-service IaaS and are now offering GPU as a service. “These service providers really need a platform,” Marshal Correia, general manager for India and South Asia, said. “The IndiaAI Mission’s compute initiative has helped make enough GPUs available to data center providers in India.” The second category is end users. These customers have experimented with use cases and now need a platform as those use cases scale. Cloud-native technology plays a major role in this layer. “Underneath LLM and above the hardware, that is the layer that we are working on,” Correia said. “This has, of course, a lot of tools, slicing, dicing, etc., security.” A large amount of work last year focused on proofs of concept, with some moving into production and many not making that transition. With greater availability of GPUs and infrastructure, SUSE sees an opportunity for customers to establish a platform or software factory that can support a growing number of applications. SUSE also sees demand from enterprises building their own infrastructure for generative AI, MLOps, and domain-specific workloads such as computer vision, as well as from organizations building multi-tenant infrastructure such as gigafactories, neoclouds, and public-sector systems. Oxenham added that these customers need infrastructure for workload isolation, access control, observability, and security. Enterprises running their own AI workloads also need a platform on which applications can move from experimentation into production, he said. Sovereignty differs by region SUSE’s definition of sovereignty also varies by market, although Oxenham said the underlying need for control is common across regions. In the U.S., private, customer-controlled infrastructure is largely driven by commercial velocity, IP protection, and financial predictability. CIOs want to eliminate unpredictable consumption-based API costs, avoid vendor lock-in, and maintain deployment flexibility. In Europe, the driver is heavily precautionary and regulatory. It is anchored in compliance, data sovereignty, and risk-tiering under the EU AI Act, along with a desire to diversify from incumbent hardware and software ecosystems. In India, the drivers include national AI ambitions and regulatory requirements, while customers also face the practical challenge of moving AI applications from experimentation into production. Oxenham said he sees sovereignty less as a binary switch based purely on data residency or jurisdiction, and more as a spectrum rooted in operational resilience, pivotability, and active risk management. “Viewing sovereignty solely through the lens of data location misses the bigger picture,” he said. “True sovereignty is about evaluating tradeoffs at every layer of the stack so an organization is never locked in or left at the mercy of external constraints.” That includes third-party Linux, orchestration, AI frameworks, models, and silicon. Air-gapped deployments add another layer of control The question of dependency becomes more direct in air-gapped environments, where enterprises need to keep data, model artifacts, and system activity within an isolated network. SUSE said its AI Factory architecture supports such deployments, with SUSE Security and SUSE Observability used to enforce and monitor network and system boundaries. Software updates can be staged through signed artifact bundles in local registries. “For fully disconnected deployments, this pre-validated architecture makes updates mechanically straightforward,” Oxenham said. “Because the entire stack is tested and verified as a single unit, customers simply mirror or stage the signed artifact bundles into local registries within their isolated network.” Air-gapped deployments illustrate the same distinction. The external technology does not disappear. Instead, the enterprise controls what enters the environment and how the software is operated and updated once inside it. The central tradeoff remains—an enterprise can retain control over its data, infrastructure, and higher-level software layers while still depending on Nvidia, AMD, or another accelerator vendor for silicon-specific components. For SUSE, sovereignty therefore rests less on eliminating dependency and more on preserving the ability to change dependencies when needed. Also read: SUSE Extends Single-Kernel Linux Strategy from Edge to Data Center https://www.eetimes.com/suse-extends-single-kernel-linux-strategy-from-edge-to-data-center/ SUSE Launches Industrial Edge Platform Following Losant Acquisition https://www.eetimes.com/suse-launches-industrial-edge-platform-following-losant-acquisition/