Wall switch set to OFF inside a ring of dimmed European Union stars, representing lost access to frontier AI

On May 21, 2026, Paul Krugman concluded a long argument that Europe was not really in economic decline. He made one exception. Europe, he wrote, “can’t be sure that it will always have access to new technologies developed and produced in the other superpowers.” He added that “the risk of being cut off from strategically important technologies, once minimal, is now very real.”

Twenty-two days later, a US directive switched off a frontier AI model for every non-American. Well, technically, it also switched the model off for every American. The episode turned European tech sovereignty from an abstract policy goal into an operational risk.

His argument turns on a measurement paradox. Noah Smith and the Garicano brothers take the other side. Since 2000, measured US productivity grew much faster than Europe’s when economists held prices constant.

Yet European output per hour at current-price purchasing power parity (PPP) barely moved relative to the US. PPP adjusts currency comparisons for local prices. The European figure rose from about 86% of the US level to about 87%.

Krugman treats the flat current-price line as closer to reality. He sees the diverging constant-price line as an effect of fast-falling US technology prices. His critics treat the divergence as real. In their view, companies captured the gains as profits and equity instead of passing them to consumers.

Euro-area output per hour versus the US, 2000–2024: constant-price falls from 104% to 87%, while current-price PPP stays near 86–87%

On whether the typical European is getting poorer, I am mostly with Krugman. The median gap is roughly stable. His comparison between California and the rest of the US is hard to dismiss. Constant-price productivity is also a poor measure of how people live. Europe remains a fine place to live and is not falling apart.

Technology you don’t own

the big benefits of IT come from applying it, rather than creating it

Europe applies information technology well. True. But “apply, don’t create” assumes that access remains frictionless and apolitical. It assumes that models, graphics processing units (GPUs), and cloud services remain available at world prices. That assumption died on June 12.

Henry Farrell and Abraham Newman call the replacement weaponised interdependence. Economic networks often depend on a few hubs. Control of one hub can let its owner cut off users downstream. They call this the chokepoint effect.

Europe’s ASML lithography chokepoint

Extreme-ultraviolet (EUV) lithography uses light to print the smallest circuits on advanced chips. ASML is the only supplier and earned €32.7 billion in 2025 sales. That is Europe’s one real card.

Chip fabrication is different. Taiwan’s TSMC held around 70% of the foundry market. A foundry manufactures chips designed by other companies.

Nvidia supplied 80% to 90% of AI-accelerator revenue and well over half of global AI computing capacity. An accelerator is a chip designed for intensive AI calculations. Nvidia’s CUDA software platform reinforces that position.

Synopsys and Cadence formed a US near-duopoly above 90% in chip-design software. Electronic design automation (EDA) software helps engineers design chips. Siemens EDA was the only European player at about 13%, and it still relied on US technology.

AWS, Microsoft, and Google supplied about 70% of the European Union cloud market. Those services also sit under US legal jurisdiction. As of June 21, 2026, US laboratories dominated the closed-model frontier. Mistral was the main European exception, while DeepSeek showed that China could build a parallel stack under sanctions.

Europe controls 100% of EUV; outside suppliers control 70% of fabrication, 85% of GPUs, 90% of design, and 70% of cloud and frontier models

The map is lopsided. Europe owns the hardest single chokepoint in the chain and buys almost everything else. The US held 74.5% of measured global GPU-cluster performance. China held 14.1%, and the European Union held 4.8%.

Even the ASML card is partly held in Washington. ASML machines contain US technology. The Foreign Direct Product Rule can restrict foreign-made products that use specified US technology.

US AI export controls at Washington’s discretion

The usual objection is that governments threaten technology restrictions but rarely apply them to allies. By June 2026, they had started doing so.

In January 2025, the US issued the AI Diffusion Rule. It divided countries into tiers for access to AI computing power. The rule placed ten European Union members in the licence-free top tier. It placed seventeen members in a capped second tier.

In May 2025, the administration rescinded the rule. Its notice said the rule “would have undermined U.S. diplomatic relations with dozens of countries by downgrading them to second-tier status.” Tiered access for allies had become too costly diplomatically.

What is worse than a bad rule for a dependent ally? No rule at all. It is discretion.

The administration made deal-by-deal grants to Gulf states. It also received a 15% share of Nvidia’s China revenue in exchange for export licences. One lawmaker said this told China and US allies that “American national security principles are negotiable for the right fee.”

And then came Fable 5. The June 12 directive required Anthropic to block the models for “any foreign national, whether inside or outside the United States.” Anthropic could not check nationality in real time, so the practical result was a worldwide shutoff.

Anthropic disputed the directive’s basis. It called the flagged flaw “a narrow, non-universal jailbreak” that other models also allowed. It warned that the standard “would essentially halt all new model deployments.”

Analysts at Lawfare and Tech Policy Press identified a broader change. Officials had extended export powers designed for physical goods to a live API. An application programming interface (API) lets software send requests to a model. Individual enforcement decisions had started to function like a rule.

Tail risk

Krugman’s welfare case and the security case answer different questions: the average outcome versus the extreme loss. Dependence is cheap and harmless most of the time. It becomes a catastrophe in the bad state of the world. Nobody prices flood insurance from the average year.

The strongest objections do not sink the point. Open weights spread after release. DeepSeek-R1 matched a leading US model at about one twenty-seventh of the cost of producing responses. As of June 21, Epoch estimated that open models trailed the closed frontier by roughly four months. Granted.

But the binding dependence is not the downloadable file. It is the cloud, API, update cycle, and advanced chips underneath. Buyers cannot substitute these components easily.

ASML gives Europe some leverage in return. A US cutoff could invite a Dutch response that blocks the leading-edge chip chain, including US factories. Also true. That possibility lowers the chance of a cutoff. It does not reduce the damage if one occurs. For this argument, severity matters more than probability.

China’s parallel stack shows that separation is survivable. It also shows that separation is expensive and probably leaves the parallel system a step behind.

European technology sovereignty: the scorecard

The giveaway is that Europe’s own institutions already accept the argument under gentler names. Mario Draghi’s competitiveness report called the dependence an existential challenge. It warned of “slow agony” and proposed roughly €800 billion in annual investment.

The report made a still sharper concession. It concluded that “it is too late for the EU to try and develop systematic challengers to the major US cloud providers.”

The EU Chips Act carried a €43 billion headline, but only about €3.3 billion was new EU money. The US CHIPS Act provided $52.7 billion. The European Court of Auditors called the bloc’s target of 20% global chip share by 2030 “very unlikely.” It estimated a more realistic share of 11.7%.

Europe’s reflex is to regulate from strength. It then quietly retreats when the AI Act starts to impose material costs.

US CHIPS Act provides $52.7 billion; the EU headline is €43 billion, including only €3.3 billion in new EU money

The June 2026 technology-sovereignty package carried its own admission: “we want to be sure nobody has a kill switch.”

Former French prime minister Édouard Philippe was more direct after Fable 5. Infrastructure whose models and computing capacity you do not control “is an infrastructure that others can unplug.”

A week later, Group of Seven leaders met at Evian while VivaTech opened in Paris. By then, the dependence was no longer a white-paper abstraction.

Footnote

Krugman moved the worry from productivity to geopolitics and treated that as filing it away. It was the opposite. European policymakers should not fear the output-per-hour line that he correctly takes apart. They should fear becoming revocable second-tier customers in somebody else’s licensing system.