Gregory Kurtzer, CEO of CIQ and founder of Rocky Linux, sat down virtually with Frank Everaardt, founder and editor-in-chief of the Dutch publication TechnologyInsider, for a conversation about sovereign AI. Everaardt told Kurtzer that nearly every interview he does now lands on one of two subjects, sovereignty or AI, and usually both at once. His readers across the Netherlands and Belgium are asking the same questions CIQ hears from organizations everywhere: how to control their own AI, and how to keep token costs from eating the budget.
The conversation covered a lot of ground and surfaced insights worth sharing. The full article is available here.
Easy to start, hard to scale
Everaardt opened with the practical question: how complicated is it to run AI yourself? Kurtzer's answer came in two parts. Starting is easy: an organization can get a small system such as an NVIDIA DGX Spark or an AMD Halo box, stand up an inferencing service, and test and validate models quickly, assuming it can get the hardware. The hard part is scaling into production: running large models efficiently, distributing inference across many systems, and driving token throughput and model accuracy at scale.
This is where organizations falter, Kurtzer said, because scale demands more than hardware. Teams have to think about the entire stack: which operating system they run, how it is optimized, how GPU drivers are managed, all the way down to the foundation. Kurtzer traces this perspective back to CIQ's roots in high-performance computing (HPC), where the company has been driving tightly coupled scale for decades. Performance compounds as workloads parallelize: a small regression multiplied across a cluster becomes a large bill, and a small optimization becomes a large saving. Whether the target is a database, an edge device, or AI training and inference, the workload should shape every layer beneath it.
Workflows that travel
Everaardt asked how easily a workload can move from a local box to a cloud provider. This is the problem Fuzzball was built to solve. Fuzzball is container-native and organized around workflows: scripted definitions of everything a job needs, whether that job is pre-training, fine-tuning, post-training, or a production inferencing service. Because the workflow carries its containers and manages data staging and caching as part of the definition, it is completely portable. The same workflow runs on a single DGX box, scales to a large cluster, or moves to a public cloud or neocloud, and Fuzzball matches the workload requirements to whatever hardware is available. CIQ is also building a catalog of templated workflows, pretested and prevalidated, so teams can deploy a complete training or inferencing system in a few clicks and run many of the same services they would otherwise get from an external provider.
CIQ describes the platform in three pieces. Fuzzball Substrate runs the compute layer, executing training and inference across CPU and GPU resources. Fuzzball Orchestrate manages incoming workloads, data movement, and placement within a cluster. Fuzzball Federate sits on top and joins many clusters together: multiple data centers, clouds, availability regions, even edge deployments, all behind one API and one workflow syntax.
Federation resonated with Everaardt because of a story gaining momentum in Europe: telecom operators discussing data highways that interconnect data centers, since countries like the Netherlands cannot easily build giant new facilities. Federation fits that model naturally. With clusters joined, workloads can be placed according to power availability and price, architecture, the cost of moving data, and data policies around locality, gravity, and security. One early Fuzzball use case involved exactly this: data centers running on renewable power. When the wind settles and capacity drops, jobs migrate to another site, and high-priority workloads can move to cloud where they finish fastest.
Why organizations are moving now
Two forces are driving the shift to sovereign AI. The first is cost. At many organizations, token costs are starting to outrun headcount costs. High-performance computing learned long ago that the most financially sound way to run sustained computation is on premises, bursting to cloud for peak capacity. AI is arriving at the same conclusion.
The second is controls and compliance. To get a useful answer from an external provider, an organization must hand over enough data and context for an informed response, which means leaking data. For hospitals, banks, aerospace firms, and chip manufacturers whose trade secrets live in their data, that was never acceptable. Media and entertainment surprised Kurtzer too; the controls those companies manage around rendering and production are extensive.
Everaardt raised shadow IT, and Kurtzer agreed it's a real risk. When employees expense personal AI subscriptions, the spend adds up fast, but the leakage of company data through personal accounts is the larger risk. Policy alone will not stop it. The durable answer, Kurtzer said, is for organizations to offer sanctioned, controlled AI that serves employees better than the tools they smuggle in.
Where the models are headed
Kurtzer is a firm believer that open and privately runnable models, Mistral among them, will keep advancing until they are competitive with the major labs. Open source wins the long race, he said. The nearer-term gains will come from harnesses and agents tuned for open models, since much of the power of today's best tools lives in the harness around the model. Freedom of choice matters here as well: different models excel at different tasks, and running several models that cross-validate each other works like different personas in an organization arguing toward the best answer. A model focused on one domain does not need the size of a general-purpose one.
During the conversation, Kurtzer repeated a line he had heard recently: if an organization's infrastructure depends on somebody else's permission, that is a massive single point of failure. Permission can be withdrawn by a company, a government, or a mandate. Organizations collected their easy wins on external providers; now they are asking how to run AI in environments they control. That transition, from validation to owned production, is the sovereign AI story of 2026, and Everaardt noted that his readers across Europe are asking exactly these questions.
Both came away from the conversation with the same takeaway: the questions European organizations are asking about sovereignty and AI costs are the same ones CIQ is building Fuzzball to answer.
