The core message of the panel was that Europe’s industrial AI opportunity depends on converting its engineering talent, deep manufacturing expertise, and valuable industrial data into repeatable systems that can create global champions. The discussion opened by contrasting past investor enthusiasm for pure software with a newer wave of capital moving toward the physical world, or “ATOM.” Ilan Leifermann argued that AI is now expected to bring software-like efficiency gains into industries, factories, defense, autonomous vehicles, and manufacturing. He pointed to examples such as defense challengers and said the ingredients are in place, even if the market has not yet seen many multi-billion or tens-of-billions-dollar exits from these industrial AI companies.

Patrick Dreisch then framed early movement as a compounding advantage. Drawing on Anthropic’s view of scaling laws, he said AI capability is improving on an exponential trajectory as infrastructure investment grows, while people and companies tend to think linearly. He cited studies measuring how many hours of human work agents can perform autonomously, saying that number is currently doubling every seven months. His concrete example was an internal Anthropic experiment building a small dog robot with Claude: a prototype that took six hours a year earlier took ten minutes with a newer model. The point was not simply that companies can wait for better models, but that organizations adopting now build the internal learning needed to exploit future capability jumps.

The panel then turned to safety, where Patrick emphasized codifying security mechanisms, responsible scaling policies, SOPs, and human workflows into code so AI systems can operate within standardized guardrails. He proposed separating tasks into revocable and irrevocable categories: automate where the impact can be managed, but keep humans involved for major decisions in areas such as energy grids. Nathan Gruber added that industrial environments involve more irrevocable decisions than many other sectors because hardware choices, production timelines, and data-location decisions become difficult to reverse. For factories, he argued, the biggest challenge is often not frontier intelligence but system design: gathering enterprise knowledge, including undocumented expert and tested knowledge, and integrating it into workflows such as industrial quality.

Finally, the panel addressed sovereignty and dependency risk. Asked whether European industrial companies should worry about relying on American models, Patrick said they should examine dependencies holistically across infrastructure, data, integrations, and frontier AI capability. He acknowledged Anthropic had a highly capable model pulled under U.S. export controls in early June, so companies need exit strategies and backup plans. At the same time, he noted AI supply chains are deeply interdependent and cited ASML as a Dutch company that represents a major leverage point and potential point of failure in chip supply. His practical conclusion was not to use one model provider for everything, but to make case-by-case decisions based on capability, risk, and business value.