Tom Becker

What AI Sovereignty Really Means for CIOs

The panel’s core message was that AI sovereignty for CIOs is not a slogan about buying local technology; it is a practical question of control, access, inspection, and dependency across the AI stack. After a brief closing fragment from the prior talk, where Simon argued that energy will be the fundamental constraint on AI and noted TeraFab’s ambition to become a fourth global AI-chip manufacturer alongside TSMC, Samsung, and Intel, the session shifted into a panel on what AI sovereignty means operationally. The moderator introduced a technology-focused panel with Markus Hacker from NVIDIA, Tom Becker from Mistral, and Dr. Tim Rietz from Respeak, hosted by Philipp Karmires of Bureau Veritas. Philipp framed the discussion as a chance to “look under the hood” at the technologies CIOs may depend on when running IT and AI infrastructure. The opening definitions established the panel’s central vocabulary. Markus defined AI sovereignty in one word as “control”: control over data, when and how it can be accessed, and whether it is used to train or enhance models. Tom agreed but added “access” as a second pillar, warning that companies should avoid building their core around a model, provider, or layer that can be cut off by either the provider or a government. Tim added “inspection,” arguing that organizations need to know which data flows to which model and be able to exit each layer of the stack without rebuilding the entire process. Together, those definitions turned sovereignty into an enterprise architecture issue: know your data flows, inspect your models and stack, and preserve exit options. The discussion then moved to dependency at hardware and software layers. Markus pushed back on a one-sided European view of dependency by explaining that Europe does not currently produce much of the hardware used in data centers, from servers to accelerated systems, networking, and storage. He argued that meaningful European semiconductor capability would require investment on the order of $40 billion, not $40 million, and asked who would bear that risk. At the same time, he emphasized “multilateral dependabilities,” naming Siemens EDA, an Austrian mask-writer producer, ASML, Zeiss, and Trumpf as European companies without which sophisticated global semiconductor production would not be possible. The point was not that dependency disappears, but that it is distributed through a complex global value chain. Tom then addressed the software and model layer through Mistral’s perspective. He distinguished between off-the-shelf productivity uses of AI, where dependency is less problematic because models can be replaced, and strategic or “iconic” use cases where companies put proprietary data into AI solutions. For those cases, he argued that organizations should not rely on a black box or risk their data being used to train a model without control. He said Mistral built open-weight models that can be deployed on-prem so customers can decide where they run them and reduce dependency. Tim’s response, from a research and regulated-government-market perspective, began by connecting open-weight models with practical controllability and comparing today’s AI work with earlier research workflows involving code, public datasets, and collected data.