The transatlantic divergence is real — EU adds prohibitions, US adds preemption, the compliance stack splits
The EU AI Act omnibus and the Trump EO are formally moving in opposite directions. EU adds specific prohibitions enforced under GDPR-style long-arm jurisdiction; US consolidates federal oversight to preempt state regulation. Convergence is not happening. Multinational AI companies now operate two distinct compliance stacks.
Through 2024-2025 the working assumption among multinational AI companies was that EU and US frameworks would gradually converge through coordination. The 2026 reality has overtaken that assumption. The transatlantic divergence is now structural — three layers of policy moving in opposite directions, with no visible coordination mechanism that would pull them back together.
Layer one is prohibitions. The EU's May 7 omnibus formally adds nudifier applications to the prohibited-practices list effective December 2, 2026. The US has no comparable federal prohibition and has signaled it will not. The two jurisdictions have publicly disagreed on which categories of AI use should be banned outright, and the gap is widening.
Layer two is enforcement architecture. The EU uses GDPR-style fines and market-access restrictions, with extraterritorial reach via the Brussels effect. The US uses federal preemption of state law as the primary leverage, with the December 2025 EO explicitly designed to prevent state-level AI safety laws (New York's, California's reactivated post-SB1047 efforts, others in draft). Different enforcement mechanisms, different leverage points, different compliance obligations.
Layer three is compliance cost calibration. The EU AI Act regulatory-sandbox deadline pushed to August 2 2027 acknowledges that high-risk-system compliance has substantial documentation and audit costs that need experimentation infrastructure to support. The US EO directionally reduces compliance burden by consolidating oversight federally rather than layering federal-on-top-of-state.
For multinational AI companies the operational consequence is that the 2024-vintage compliance plans — design once, deploy globally — have been overtaken. The 2026 reality is two distinct compliance stacks: EU stack with prohibitions, GDPR-grade audit trails, and sandbox experimentation; US stack with federal preemption, lighter audit requirements, and state-law dormancy. The two stacks have to be maintained independently and updated on independent regulatory schedules.
The throughline: convergence on AI regulation has been the polite assumption for three years. The empirical evidence in 2026 is that convergence is not happening — and that the divergence is deliberate rather than accidental. The compliance architecture has to adjust. The competitive advantage for the next 18 months goes to companies that build the two-stack discipline early rather than waiting for a convergence that isn't coming.
Consilium — AI Council and Parliament agree to simplify and streamline rules → · Data Innovation — New York's AI Safety Law Claims National Alignment but Delivers Fragmentation →