Skip to content

Thinking on Borrowed Compute

For most of the AI cycle, Europe's dependence on US compute, model and cloud providers was an accepted commercial fact, the shape of a market Europe long ago made the decision not to lead. In June 2026 it became something else. A US export-control directive suspended access to a frontier model for everyone outside the United States, including the European staff of the company that built it.

15 June 2026  ·  15 min read aieuropesovereignty

In a little over three years, a technology almost nobody used has become one the economy cannot easily do without. ChatGPT launched at the end of November 2022; by the middle of 2026 it was reaching on the order of a billion people a week, the fastest any consumer product has reached that scale.1 The consumer numbers are the least of it. These systems have become load-bearing where work is done.

In the wider economy, around seven in ten organisations now use generative AI in at least one function, roughly nine in ten of the Fortune 500 run on ChatGPT, and about one person in six worldwide used a generative tool in the second half of last year.2 Among software developers the shift is close to total: some eighty-four per cent use AI coding tools, and GitHub’s assistant alone now writes nearly half the code of the developers who use it.3 In science and medicine more than a hundred and seventy AI-originated drug programmes are in clinical development, the US regulator has issued a framework for AI in drug discovery, and the model that mapped essentially every known protein won its makers a Nobel Prize.4 In law, ninety-two per cent of professionals report using at least one AI tool, with personal use among lawyers rising from a third to better than two-thirds in a single year.5 Chipmakers and carmakers talk of a “ChatGPT moment for robotics” as humanoids move onto production lines, though the people building them concede the field is oversold and that performance falls away outside the laboratory.6 And in policing, facial recognition, automated plate readers and real-time crime centres have moved from pilots into routine operation, accuracy and bias failures and all.7

Not all of this is proven, evenly spread or free of hype, and an honest account says so. But the direction is not in doubt. In under four years AI has moved from novelty to infrastructure, and infrastructure is the kind of thing whose absence you notice only when it stops. Almost all of it traces back to a few US providers and one chip supply chain, and Europe owns very little of either. So when one of those providers is switched off by a government’s order, it is not a tech-industry squabble. It is a stress test of infrastructure that much of the productive economy now leans on. One such test arrived on a Friday evening in June.

On 12 June 2026 the US government issued an export-control directive suspending access to Anthropic’s two most capable models, Claude Fable 5 and Mythos 5, for any foreign national, inside the United States or out, including Anthropic’s own non-citizen employees. With no way to enforce that per user at speed, the company disabled both models for everyone.8 The directive hit the frontier tier; the previous model, Opus 4.8, stayed available.89 What vanished overnight was the leading capability. Even so, Bloomberg called the order unprecedented, the first time a widely deployed model had been pulled by government order. A developer in Berlin who had ended Thursday with the most capable coding assistant ever built woke on Friday to a refusal. It was also the second time in four months the administration had moved against this particular company.

Underneath the incident sits a far larger position, and the European Commission’s own account of it is harder to argue with than any outside estimate. By Brussels’ reckoning the European Union depends on non-EU providers for more than four-fifths of its digital products, services, infrastructure and intellectual property, makes around a tenth of the world’s semiconductors, and has more than seventy per cent of its cloud market in the hands of three US companies.10 On computing power specifically, independent work puts Europe’s share of the global total at roughly five per cent, with each of the three largest US labs commanding more than all of Europe combined.10 This dependence is mostly invisible, because it is built into tools that European institutions, firms and citizens already use, and it deepens on its own, without anyone choosing it at the point of use. The June suspension was one visible instance of a very large, quiet exposure.

A dependence a foreign government can switch off is a control point, not a supplier relationship. Europe’s own institutions have started to say as much. Days before the suspension, the Commission presented a Tech Sovereignty Package that casts digital dependence as a strategic vulnerability to be corrected by law; the president of the Commission put the stakes as the technologies that keep hospitals running, grids stable and services secure; and the commissioner now responsible carries the title Tech Sovereignty, Security and Democracy.1112 The package would, among other things, let Brussels bar non-compliant cloud providers from sensitive contracts and override commercial chip agreements in a crisis.13 The security framing is no longer rhetorical.

The control is also not only a matter of rules. The first of those two moves came in February, when the Pentagon designated the company a supply-chain risk, the first such label applied to a US firm and one historically reserved for foreign adversaries such as Huawei, and ordered it out of federal agencies. The trigger was Anthropic’s refusal to drop two safeguards the government wanted gone: against mass domestic surveillance, and against fully autonomous weapons. The company sued, calling it an unlawful campaign of retaliation; Senator Elizabeth Warren and a range of legal commentators read it the same way; the White House denied it, citing national security and the company’s future conduct; speech, it said, was not the basis.14 The merits will be settled, if at all, in court. Either way, European users depend on providers exposed to a government that has shown itself willing to act in ways widely seen as discretionary and punitive, and on this record a provider’s ethical restraint can raise its exposure rather than lower it.

Which disposes of the obvious hedge. If the danger is an erratic government, the instinct is to choose the most principled vendor and shelter behind its values. That does not work, because vendor values bend to the same government in both directions. Anthropic held its lines and was cut off. Within hours of that February designation, OpenAI signed a Pentagon deal that Anthropic had refused, and a visible consumer revolt followed: a campaign to cancel ChatGPT, a sharp rise in uninstalls and negative reviews, and Claude briefly the most downloaded free app in the United States.15 OpenAI maintains its deal preserved the key safeguards; critics call it capitulation; its chief executive conceded the arrangement looked opportunistic and later amended the terms. Whichever reading is right, the conclusion for a European is the same. No choice among US vendors yields a dependency insulated from US politics. Hold the line and you can be cut off; fold and you keep your access, but you remain governed by the power that the folding served. The only exit that changes the equation is one that does not run through a US vendor at all.

A sharper reading is harder to prove but difficult to avoid: a government that punishes restraint and rewards compliance is, over time, selecting which providers thrive. A counter-pressure exists, since that same compliance cost OpenAI a wave of users, so the outcome is still open. But it means the composition of the intelligence Europe runs on is being decided by a struggle between coercion and conscience inside US politics, in which Europe has no vote. If that is strategic, why have Europe’s usual instruments not already addressed it?

The answer is that those instruments were built to regulate other people’s technology. Building Europe’s own was never their job. Data protection and the AI Act govern how foreign systems may behave inside the single market, which protects citizens but does nothing to close the capability gap, and the gap is self-reinforcing: capital, compute and talent compound on whichever side already leads. An independent scenario published on 11 June puts it bluntly, describing Europe’s €200bn AI fund as largely repackaged money set against a single year of far larger US spending, and naming the coming summer as the last moment at which the trajectory might still be changed, with Mistral’s capital constraints as the illustration.10 The comfortable conclusion would be that the dependence is permanent.

It is not. The clearest evidence that the dominant cluster’s indispensability is contingent comes, awkwardly, from China. DeepSeek’s V4-Pro leads the open-weight field on coding, scoring around eighty-one per cent on a standard benchmark that independent trackers now corroborate, level with one of the proprietary frontier models, while shipping under an open licence, running on an efficient architecture at roughly a thirtieth of the output cost of the leading US systems, and increasingly served on non-Nvidia silicon.16 It matches the previous frontier tier, the one Anthropic’s Fable 5 has already left behind. And it is not Europe’s hedge. It is Chinese, and adopting it trades one dependence for another with its own costs of trust and security. It is, though, an existence proof of a direction: an open, self-hostable, auditable path that need not route through the dominant cluster at all. It is not trained independently of Nvidia; the weights and the serving are open, the training is not. And it does not show that computing power is optional, since building it took serious capability and infrastructure, on the argument of the scenario report already cited. That is why Europe needs both the open path and real investment in compute. The direction is real. The only remaining question is whether Europe takes it.

It has started to. The Tech Sovereignty Package pairs a revised Chips Act and a Cloud and AI Development Act with an Open Source Strategy that places open ecosystems at the centre of the sovereignty case, and trails a call for AI “gigafactories” expected in July.1117 It came nine days before the suspension, so it was not a response to it; the anxiety was already hardening into law, and the incident gave it a public illustration. The official posture has shifted from regulating dependence to attempting to build capability. But it is probably too slow. Forrester expects no European enterprise to move fully off the US hyperscalers this year, and the scenario report already cited argues that Europe is still treating sovereignty as a slogan while the United States runs it as an industrial race.1810

The dependence is now visible and named, the alternative path has been shown to exist, and whether Europe builds it inside a closing window is genuinely open. Fatalism and triumphalism are equal and opposite errors here: the one says nothing can be done, the other assumes Brussels will see it through. The indicators are concrete. Whether the gigafactories arrive at real scale and speed. Whether the €200bn becomes hardware or stays an announcement. Whether Mistral and its peers are funded to compete. Whether a single large European enterprise meaningfully moves off the US hyperscalers. On current form the honest answer to most of those is not yet, and perhaps not in time.

None of that is in the gift of a single company. What is in its gift is narrower: the architecture of its own dependence, which it can begin to change now, long before the gigafactories arrive. Europe has been thinking on borrowed intelligence for most of the cycle, on terms it does not set and cannot guarantee. In June it learned, in the plainest possible way, that a loan can be called in. The work now is to need the lender less.

References

  1. ChatGPT launched 30 November 2022. OpenAI reported ~900M weekly active users in early 2026, up from ~400M a year earlier; by mid-2026 it was reported to have crossed roughly a billion weekly/monthly active users, the fastest consumer product to that scale (Sensor Tower estimates via Reuters). It reached 100M users within about two months of launch. ↩

  2. McKinsey: more than 70% of organisations use generative AI in at least one business function. OpenAI: ~92% of the Fortune 500 use ChatGPT, with 7M+ enterprise seats. Microsoft: ~1 in 6 people worldwide used a generative-AI tool in H2 2025. Pew: ~34% of US adults have used ChatGPT. Gartner: worldwide AI spending ~$2.6tn in 2026. ↩

  3. Stack Overflow 2025 Developer Survey (n≈49,000): ~84% use or plan to use AI coding tools, up from 76% in 2024, even as trust in outputs fell. Microsoft / GitHub: Copilot generates ~46% of code written by active users and is deployed across ~90% of the Fortune 100. ↩

  4. AI drug-discovery market ~$2.6bn with 170+ AI-originated drug programmes in clinical development (IntuitionLabs, early 2026); McKinsey estimates generative AI could save pharma $60–110bn annually; the FDA issued “Guiding Principles of Good AI Practice in Drug Development” (January 2026). DeepMind’s AlphaFold predicted structures for essentially all ~200M known proteins; Hassabis and Jumper shared the 2024 Nobel Prize in Chemistry. ↩

  5. Wolters Kluwer 2026 Future Ready Lawyer Survey: ~92% of legal professionals use at least one AI tool. 8am 2026 Legal Industry Report (via the American Bar Association): personal generative-AI use among legal professionals rose from ~31% (2025) to ~69% (2026). ↩

  6. NVIDIA framed CES 2026 as a “ChatGPT moment for robotics”; foundation models for robots have matured. Bank of America projects ~90,000 humanoid-robot shipments in 2026; Figure’s humanoids ran in production at BMW’s Spartanburg plant; Unitree shipped 5,500+ units in 2025. Industry figures concede the field is oversold, with field performance well below lab results. WSJ, December 2025. ↩

  7. AI policing tools (facial recognition, automated licence-plate readers, real-time crime centres, drones, edge-AI body cameras) are in routine operational use across US and Canadian agencies rather than pilots; Deloitte reports widespread city adoption of facial recognition and biometrics. The Brennan Center cites independent audits finding 80–90% of ShotSpotter alerts unlinked to a confirmed gun offence, and facial-recognition wrongful arrests falling disproportionately on Black individuals. ↩

  8. Suspension of Claude Fable 5 and Mythos 5 following the June 2026 US export-control directive (foreign nationals including Anthropic’s European-based staff; both models disabled for all users; Opus 4.8 retained; Anthropic disputes the severity). Anthropic. https://www.anthropic.com/news/fable-mythos-access . Bloomberg called the order “unprecedented”: https://www.bloomberg.com/news/articles/2026-06-13/anthropic-says-us-limits-foreign-access-to-fable-5-mythos-5 ↩ ↩2

  9. Fable 5 and Mythos 5 launch, capability tier and access terms. Anthropic. https://www.anthropic.com/news/claude-fable-5-mythos-5 ↩

  10. EU structurally reliant on non-EU providers for more than 80% of digital products, services, infrastructure and IP; ~10% of global semiconductor production; more than 70% of the EU cloud market held by three US hyperscalers. European Commission communication accompanying the Tech Sovereignty Package, 3 June 2026. https://commission.europa.eu/news-and-media/news/strengthening-europes-tech-sovereignty-2026-06-03_en . Europe ~5% of global AI compute; each of OpenAI, Google and Anthropic commands more than all of Europe combined; €200bn InvestAI characterised as largely repackaged money; summer 2026 as the last actionable window. Bakker et al., Europe 2031, a June 2026 scenario forecast modelled on AI 2027; the present-day figures cited here are its factual baseline, the post-2026 trajectory fictional. https://europe2031.ai/compute-forecast/ ↩ ↩2 ↩3 ↩4

  11. EU Tech Sovereignty Package, 3 June 2026: digital dependence framed as a strategic vulnerability; Chips Act 2.0; Cloud and AI Development Act; EU Open Source Strategy; AI Gigafactories call expected July 2026. European Commission. https://commission.europa.eu/news-and-media/news/strengthening-europes-tech-sovereignty-2026-06-03_en . Analysis via Digital Watch Observatory. https://dig.watch/updates/the-eus-tech-sovereignty-package-and-the-future-of-european-digital-power ↩ ↩2

  12. von der Leyen framed the stakes as the technologies that keep hospitals running, energy grids stable and services secure. European Commission statement accompanying the Tech Sovereignty Package, 3 June 2026, as reported by Digital Watch Observatory and others. https://dig.watch/updates/the-eus-tech-sovereignty-package-and-the-future-of-european-digital-power ↩

  13. Proposed powers to bar non-compliant cloud providers from sensitive government contracts and emergency powers over commercial chip agreements, including overriding existing agreements during a supply crisis, and the Commissioner brief titled Tech Sovereignty, Security and Democracy. TechPolicy.Press. https://www.techpolicy.press/eu-unveils-sweeping-tech-sovereignty-push-balancing-autonomy-with-openness/ ↩

  14. February 2026: the Pentagon designated Anthropic a supply-chain risk, the first such designation of a US company and a label historically used for foreign adversaries such as Huawei; the administration ordered federal agencies to stop using its technology and the GSA removed it from federal platforms, after Anthropic refused Pentagon demands to drop safeguards against mass domestic surveillance and fully autonomous weapons. Anthropic sued, calling it an “unlawful campaign of retaliation”; Senator Elizabeth Warren and legal commentators echoed the retaliation reading; the White House denied it, citing national security and the company’s future conduct. Coverage frames the June export controls as the latest escalation of the same dispute. Mayer Brown https://www.mayerbrown.com/en/insights/publications/2026/03/pentagon-designates-anthropic-a-supply-chain-risk-what-government-contractors-need-to-know ; Warren letter (primary) https://www.warren.senate.gov/imo/media/doc/letters_redesignationofanthropicasnationalsecurityrisk.pdf ; Al Jazeera (White House denial) https://www.aljazeera.com/economy/2026/3/24/anthropic-challenges-us-pentagons-ban-in-san-francisco-court-showdown ; Anthropic statement (primary) https://www.anthropic.com/news/statement-department-of-war ↩

  15. OpenAI’s Department of Defense deal (announced around 28 February 2026, within hours of Anthropic’s blacklisting) and the resulting QuitGPT / “Cancel ChatGPT” backlash. Independent signals: Sensor Tower reported a ~295% spike in US ChatGPT uninstalls and a ~775% jump in negative reviews around 28 February, with Claude briefly the top free app on the US Apple App Store. OpenAI stated its deal barred domestic mass surveillance and required human control of lethal force; critics characterised it as accepting “any lawful purpose”; Altman conceded it “looked opportunistic and sloppy” and terms were later amended. Euronews https://www.euronews.com/next/2026/03/02/cancel-chatgpt-ai-boycott-surges-after-openai-pentagon-military-deal ↩

  16. DeepSeek V4-Pro-Max: SWE-bench Verified ~80.6%, the top open-weights score and level with Gemini 3.1 Pro (independently tracked via llm-stats, June 2026); open weights (MIT); efficient sparse-attention MoE; output ~28.7x cheaper than Claude Opus 4.8; non-Nvidia (Huawei Ascend) inference. DeepSeek API documentation https://api-docs.deepseek.com/news/news260424 ; independent tracking summarised at https://www.morphllm.com/deepseek-v4 ↩

  17. EU Open Source Strategy places open source at the centre of the sovereignty approach (reduce supplier lock-in, increase transparency, retain control). Digital Watch Observatory. https://dig.watch/updates/the-eus-tech-sovereignty-package-and-the-future-of-european-digital-power ↩

  18. No European enterprise will shift entirely off US hyperscalers in 2026. Forrester 2026 European Predictions. https://www.forrester.com/press-newsroom/forrester-european-2026-predictions ↩