The U.S. Reversed Its AI Rule—Europe Kept Building Its Own Stack
Washington abandoned the AI Diffusion Rule, but the broader drive for sovereign AI continued. The Bureau of Industry and Security’s May 2025 notice records both the rule’s rescission and instructions not to enforce it, while preserving a targeted semiconductor-control agenda.
Europe subsequently widened its technological-sovereignty policy across chips, infrastructure, cloud services, software and AI. The European Commission’s June 2026 package includes a Chips Act 2.0 initiative, a Cloud and AI Development Act and an open-source strategy intended to reduce dependencies across the technology stack. Reversing one American rule therefore did not remove the incentive to secure alternatives.
The argument is about control, not complete self-sufficiency
The strongest version of the sovereign-AI case does not require every country to manufacture its own processors, train a frontier model and operate an entirely domestic software ecosystem. It requires enough authority and substitutability to keep essential systems running if a supplier, licence or government policy changes.
Andrew Ng made this distinction in a January 2026 Moneycontrol interview, arguing that India should avoid allowing another country or a single company to control its access to AI while rejecting the premise that every component must be built domestically. His proposed shortcut was participation in open-source ecosystems.
Sovereignty is therefore better understood as operational control than technological isolation. A government or company may rely on imported processors and internationally developed software while retaining the ability to run a model locally, inspect or modify it, move workloads between operators and continue using a stored version without renewed permission.
Open weights can improve that position at the model layer, but they do not settle the entire question. Compute capacity, energy, networking, security maintenance, specialist staff and compatible licences remain potential points of dependence. A downloadable model available only through infrastructure controlled elsewhere provides less practical autonomy than its distribution terms might suggest.
Washington reversed a broad framework, not export controls as a whole
The U.S. policy record does not support a simple story of restrictions becoming steadily broader. The abandoned diffusion framework would have divided access among country tiers and imposed compliance requirements on advanced computing products and certain model weights. The Commerce Department’s own rationale for reversing it included concern about regulatory burdens and damage to relationships with countries placed outside the preferred tier.
That reversal supports a narrow part of the backfire thesis: American policymakers recognized that a sweeping access regime could impose costs on domestic suppliers and unsettle partners. It does not establish that export restrictions as a category had failed or disappeared.
The same policy move was paired with measures addressing Chinese advanced computing chips, the use of U.S. chips for Chinese AI models and supply-chain diversion. For foreign buyers, the resulting signal was mixed: one proposed framework was removed, but access to advanced technology remained subject to destination, end-user and national-security rules.
This distinction matters because the commercial effect depends partly on predictability. A partner does not need to believe that access will certainly be terminated to value an alternative; a material possibility of future restrictions can be enough to influence procurement, infrastructure investment and model selection.
Europe’s response now covers the entire stack
Europe’s policy direction shows why sovereign AI cannot be reduced to producing a national chatbot. Its agenda connects semiconductor supply, computing capacity, cloud infrastructure, software, open-source development and the energy systems needed to support digital infrastructure.
The layers are related but not interchangeable:
- accelerators and large-scale computing capacity determine which models can be trained or operated;
- cloud and data-centre governance affects jurisdiction, continuity and control of sensitive workloads;
- model licences and available weights determine whether systems can be inspected, adapted or moved;
- software tooling and interoperable standards affect the cost of changing suppliers;
- operational expertise determines whether nominally local infrastructure can be secured and maintained.
A country can strengthen one layer while remaining dependent at another. Domestic data centres may still rely on imported accelerators, while an openly licensed model may depend on proprietary deployment software. Sovereignty is consequently a spectrum of replaceability, legal authority and operational capability rather than a status automatically obtained by locating servers inside national borders.
The European package also weakens any claim that U.S. export controls alone created the movement. Industrial competitiveness, cybersecurity, data governance, energy resilience and dependence on non-European suppliers all appear within the wider policy logic. American restrictions may reinforce the case for diversification without being its sole cause.
Where the backfire thesis holds—and where it goes too far
The evidence supports the claim that uncertainty over continued access gives governments and regulated industries a reason to diversify. A model or cloud platform that can be operated locally, transferred between suppliers or maintained through a broad developer community may carry strategic value even when a closed American service performs better on a benchmark.
The evidence does not establish that export controls are the “fastest” route to ending U.S. leadership. Investment in alternative infrastructure does not by itself show parity with leading American laboratories, and the effects of controls cannot be cleanly separated from subsidies, procurement policy, privacy requirements, research funding and ordinary technological competition.
Security and commercial outcomes must also be assessed separately. A restriction may slow access for an intended military or strategic target while simultaneously encouraging other buyers to fund substitutes. Evidence of substitution does not prove that the security measure failed, just as evidence of a security benefit does not erase the cost of pushing partners toward competing ecosystems.
The useful test is therefore narrower: does a policy impede its intended target, and does it also motivate customers outside that target to reduce their exposure to American suppliers? The first question concerns national security; the second concerns the durability of U.S. technological influence.
Reliable access has become part of the AI product
For governments and critical industries, model quality is no longer the entire purchasing decision. Licence portability, local deployment, hardware availability, contractual continuity, control of encryption keys and the practical ability to replace an operator now affect the value of an AI system.
This creates a competitive challenge that performance alone cannot resolve. Interoperable infrastructure, transparent licensing and credible continuity commitments can make American technology compatible with a partner’s sovereignty goals. Without those assurances, funding alternative models, clouds and computing capacity is a rational hedge rather than a symbolic rejection of U.S. technology.
The updated record points to gradual diversification, not the sudden collapse of American AI leadership. Washington recognized the diplomatic and commercial risks of one broad rule and reversed it, while Europe continued building a policy around greater technological control. The lasting consequence is a market in which confidence in future access has become a product feature—and policy uncertainty can strengthen the investment case for competing stacks.
Also read:
Subscribe to our newsletter
Get the latest Web3, AI, and crypto news delivered straight to your inbox.