Export Controls Fracture the AI Order; Sovereignty Bets Accelerate

The Trump administration's export control order against Anthropic's Fable 5 and Mythos 5 models did in days what years of European AI policy debate could not: it provided a concrete, operational proof point that dependence on US-controlled frontier AI is a sovereign risk, not a theoretical one. JPMorgan and Goldman Sachs cut Anthropic access for Hong Kong staff within days. G7 leaders convened on 'trusted partner' frameworks. OVHcloud announced frontier model ambitions. France's Palantir divestment — confirmed independently but landing in the same news cycle — reinforced the same logic at the intelligence infrastructure layer. These are not coincidental signals; they are a coordinated strategic reorientation by allied governments who now have a named precedent to cite.

The regulatory mechanism itself is revealing a category error. Export control frameworks built for discrete physical goods with identifiable transfer points cannot cleanly apply to software models accessible via API. The White House's demand that Anthropic guarantee zero jailbreaks before rereleasing Fable 5 is, per independent security researchers, technically unachievable — suggesting policymakers are writing compliance mandates they do not understand. The result is a framework that functions primarily as a competitive handicap on US labs rather than meaningful capability containment, since foreign models with equivalent capabilities face no equivalent constraint. The Gottheimer bill's pre-deployment review proposal and the Anthropic-Pentagon friction both reflect the unresolved tension between a safety-first AI ecosystem and a defence establishment operating without a coherent civil-military AI integration framework.

Three capital-market signals this week point in the same direction. Goldman Sachs and Morgan Stanley are erecting information barriers to manage concurrent OpenAI and Anthropic IPO mandates — treating both listings as imminent. Nvidia is raising $25 billion in its first corporate bond sale in five years, repositioning from hardware vendor to infrastructure-class capital allocator. And Noam Shazeer's defection from Google to OpenAI, timed to the IPO roadshow, illustrates how talent concentration and capital concentration are now self-reinforcing: IPO proceeds fund compute, compute funds model quality, model quality attracts talent. Man Group's warning of a 'violent' potential correction in AI credit markets sits as a counterweight — debt financing of AI infrastructure may be mispricing duration and demand risk at exactly the moment the capital stack is expanding fastest.

The SpaceX acquisition of Cursor at $60 billion crystallises another consolidation dynamic: the agentic workflow interface layer is now valued above the models it routes to, because stickiness and distribution matter more than raw benchmark performance in a world where capable open-weights models are proliferating. China is running a parallel consolidation play through different instruments — Shanghai's Star Market IPO rule clarification for unprofitable AI labs, the ByteDance-Tencent pricing offensive, and Alibaba's Qwen Robot Suite push into embodied intelligence describe an end-to-end domestic stack being built deliberately to reduce dependence on US capital, US models, and US infrastructure simultaneously. The enterprise ROI reckoning adds a further compression: as CFOs demand cost-per-output accountability, the near-term revenue growth rates that IPO valuations are pricing in face a structural challenge from buyers who have already started cutting licenses.

Meta's conditional withdrawal of KOSA opposition — contingent on federal preemption of state AI laws — is the starkest illustration of a broader strategic maturation among AI incumbents: having failed to prevent state-level regulation in California, Colorado, Illinois, and Texas, major platforms are now using high-momentum federal bills as preemption instruments. The EU AI Omnibus follows the same structural logic in the opposite direction: a regulatory framework that was meant to establish a rights-protective floor has been amended at the final legislative stage to prioritise industry operability, weakening the very provisions that were its political justification. In both cases, the substantive policy outcome was shaped less by the stated purpose of the legislation than by jurisdictional and competitive interests operating within the legislative process.

The NO FAKES Act opposition from CDT and EFF — on First Amendment grounds — and the conservative coalition's resistance to age-verification mandates in KOSA on privacy grounds together create an unusual left-right pincer around the same legislative package, reflecting how AI governance has become a vector for pre-existing constitutional and regulatory disputes rather than a cleanly new policy domain. For policy professionals, the diagnostic implication is clear: every federal AI bill must be assessed not only on its substantive provisions but on its preemption scope, which is where the real regulatory stakes now reside. The EU Omnibus precedent and the Meta KOSA maneuver will both be studied as templates for how AI governance outcomes get determined through process rather than principle.