Why the U.S. Is Leaving Its Best AI Partner on the Table
RAND's first transatlantic AI framework argues the U.S. has a functional AI relationship with Europe and treats it as tactical maintenance rather than a strategic platform.
RAND's first transatlantic AI framework argues the U.S. has a functional AI relationship with Europe and treats it as tactical maintenance rather than a strategic platform.
Washington has spent two decades building a working AI relationship with Europe. Dialogue channels, research ties, venture-capital flows and reciprocal market access all exist. Almost none of it is being used as strategy.
That gap is the subject of The Transatlantic Artificial Intelligence Calculus, a 50-plus-page RAND research report published July 7, 2026. Its author, policy analyst Marzia Giambertoni, walks the reader through a four-scenario framework built to answer a narrow question: which cooperation moves should Washington make with Europe, and in which AI future?
The report's argument is not that the partnership is broken. It is that the partnership works well enough to substitute for a real one. That substitution is the trap.
Two cooperation lanes get most of the report's attention. The first is model evaluation, testing and standards interoperability: the practical work of making sure a French-built evaluation suite scores a U.S. model the same way a German one does, and that a safety benchmark designed in Brussels is accepted inside an American lab. The second is lithography and semiconductor supply-chain resilience. Lithography is the advanced chipmaking equipment, dominated by Dutch and Japanese vendors, that sits upstream of every frontier training cluster. Both lanes already have a transatlantic footprint. Neither is being treated as strategic infrastructure.
The framework's four scenarios map cooperation moves to outcomes across plausible AI futures. Two axes define them. The first is the locus of AI value creation: does most of the economic value accrue to whoever builds the best models, or to whoever controls the deployment layer that delivers them to enterprises and consumers? The second is the degree of advanced-model openness, meaning whether frontier models are released as open weights (publicly downloadable parameter sets) or kept closed behind APIs.
Each combination produces a different recommendation. In a world where closed models win and value sits in the model layer, the United States gains little from deeper European entanglement and is better served by protecting its lead. In a world where open weights dominate, the calculus flips. Europe becomes a force multiplier for U.S. firms that want their standards adopted globally, and bilateral moves on evaluation interoperability start paying back. The lithography lane matters in both cases, but for opposite reasons: in the closed-model world it is about securing a chokepoint; in the open-weights world it is about making sure the European equipment base is paid to participate rather than neutralized. The report's full PDF walks through all four scenarios in detail.
The report refuses to crown a winning future. It asks instead which moves hold up across all of them.
Most transatlantic AI coverage treats the relationship as a regulatory alignment problem. Brussels regulates strictly under its AI Act; Washington stays light-touch; the two sides should harmonize. A companion analysis from Brookings lays out that divergence in detail. RAND's actual finding runs the other way. The cooperation infrastructure already exists and is genuinely functional, and the failure is that it has never been used as anything more than a diplomatic convenience. That changes what an administration is being asked to do. A regulatory-alignment problem gets solved by negotiation. A strategic-partnership problem gets solved by treating working cooperation as a platform for shared capability, not as the end state.
Treating the relationship as a platform looks like this. On the evaluation lane, it looks like joint funding for shared test suites, mutual recognition of model audits, and a habit of writing safety standards with one drafting team rather than two. On the lithography lane, it looks like procurement commitments, joint investment in next-generation equipment, and a willingness to pay European vendors for resilience rather than treating their prices as a tax on U.S. progress. None of this requires a treaty. All of it requires the U.S. side to stop treating its European AI footprint as background noise.
The report has limits. Its framework assumes the EU will remain a coherent negotiating partner on AI, which depends on internal politics the report does not model. Its scenarios are stylized, and the lines between open and closed weights blur in practice as labs release partial weights or restricted licenses. Its recommendations are conditional, and conditional recommendations are easy for a new administration to file and forget.
The United States does not need to invent a transatlantic AI partnership. It needs to use the one it already has, and to recognize that the cooperation architecture sitting in maintenance mode is a strategic asset, not a side project. The first concrete signal of whether Washington treats the architecture as background or as platform will be the FY27 budget request for joint AI evaluation work with EU partners—if the administration files one.