Innovation Capture Pipeline

The innovation Europe produces is captured before it reaches scale.

The Nobel That Moved

In February 2023, a team of researchers at Meta’s FAIR lab in Paris published LLaMA, the model that democratised frontier AI. Trained on publicly available data, released with open weights, it became the foundation for thousands of downstream models.

By every metric that matters, America had captured it. The research was conducted under Meta’s employment, the intellectual property American, the paper carrying a California address. The researchers were in Paris; the value was in Menlo Park.

Five of the fourteen LLaMA authors left Meta to co-found or join Mistral AI, which within two years became Europe’s only frontier AI company, valued at six billion dollars and funded largely by American venture capital.

The traditional version of this capture is visible. Someone leaves, something gets acquired, value transfers. In October 2024, the Nobel Prize in Chemistry went to Demis Hassabis and John Jumper for AlphaFold. Hassabis is British, studied at Cambridge and UCL, founded DeepMind in London. Google acquired it in 2014. The Nobel was announced as a Google achievement.[1]

The pattern predates AI.

The World Wide Web was invented at CERN and commercialised by American companies. The MP3 was developed at the Fraunhofer Institute and monetised by Apple.

Bluetooth was designed by Ericsson and became an American platform standard. Spotify was built in Stockholm and listed on the New York Stock Exchange.

Skype was built in Tallinn and acquired by Microsoft for $8.5 billion. The Wikipedia entry filed it under American technology companies.[2]

In each case, European publicly funded research produced a capability that the American commercial ecosystem captured, scaled, and profited from. The captures were visible. They were also, individually, explicable: better capital, larger markets, faster commercialisation. The structural cause, the absence of a European institutional buyer, was invisible because each case looked like a one-off.

In European policy discussions, Skype is cited as evidence that America innovates and Europe does not. The acquired innovation generates the narrative that prevents the institutional response that would stop the acquisition.

These are the visible captures. Everyone sees the departure, the acquisition, the headline. LLaMA revealed something different: the possibility that the capture had already happened while everyone was still in Paris.

Extraction-in-Place

The most sophisticated capture is the one nobody complains about.

When a researcher leaves Paris for San Francisco, it registers as brain drain.

Newspapers write about it. Politicians respond.

When a researcher stays in Paris but works for an American company, producing assets owned by an American corporation under standard employment IP assignment, it registers as employment. The value leaves without the person, and the departure is invisible because there was no departure.

Yann LeCun convinced Meta to open FAIR in Paris in 2015 because the talent was there.

Everything needed to build a world-class AI research lab existed in France: the mathematical tradition, the grandes écoles, the research culture. The missing ingredient was a single institutional decision.

Build the lab, fund it properly, give it compute, let it recruit at competitive salaries. LeCun made that decision.

He made it inside Meta, because no European institution would make it.[3]

FAIR Paris became what a European AI research institute should have been:

well-funded, working on frontier problems, staffed by the best researchers in the field. The funding came from American advertising revenue, the intellectual property went to an American corporation, and the commercial applications served American shareholders.

Europe got the prestige of hosting a world-class lab. America got the returns.

The institutional ghost is the right metaphor.

FAIR Paris occupies the space where a European institution should exist. It filled a gap that European policymaking left open.

When Europe talks about building sovereign AI capability, the conversation begins from a position where the leading AI research lab on European soil is already American. The starting line was moved before the race was announced.

DeepMind tells the same story in a different register.

Demis Hassabis studied at Cambridge and University College London. He founded the company in London in 2010.

The research culture, the university connections, the available talent were all British. Google acquired DeepMind in 2014 for approximately £400 million.

The researchers stayed in London. The intellectual property moved to Mountain View.

When AlphaFold solved protein structure prediction, one of the most significant scientific advances of the decade, it was announced as a Google achievement. The Nobel Prize in Chemistry went to Hassabis and John Jumper in October 2024.

British education, British talent, British research culture, American ownership, American returns.[4]

DeepMind is extraction-in-place at institutional scale.

Hundreds of British-trained researchers work in London, publish from London, and raise their children in London schools. The research is conducted on British soil.

The strategic decisions about deployment, licensing, and commercial application are made in California. A British computer science PhD costs the public purse roughly £250,000 over its duration.

DeepMind employed hundreds of them. The value of their work accrues to Alphabet shareholders.

The extraction is invisible because nobody left.[5]

Meta did not need DARPA or In-Q-Tel funding directly.

The pipeline that funds American champions is examined in Paper 8. The American ecosystem produced Meta-scale companies through advertising revenue, venture capital, and Section 230 liability protection.

None of those preconditions were built in Europe. No European company reached equivalent scale. The institutional gap that FAIR Paris filled is a direct result.[6]

The Visible Pipeline

The dependency architecture is examined in Paper 6. The question is why the innovation is European and the returns are American. The answer has two parts. The first is a pipeline that moves talent, companies, and intellectual property from European origin to American ownership. The second, which LLaMA revealed, is a mechanism that achieves the same result without moving anything at all.

The visible pipeline begins with education. European universities produce talent at scale, funded by European taxpayers: ETH Zurich, Cambridge, Oxford, the Max Planck Institutes, INRIA, the grandes écoles. Seventy per cent of leading AI researchers working in the United States were educated outside the country. The single largest source is Europe.[7]

That talent enters the American ecosystem through four channels, each feeding the next.

Academic recruitment: Stanford offers ten times the compute budget and a path to industry; the European postdoc offers a two-year contract and a teaching load.

Corporate extraction: DeepMind in London, FAIR in Paris, Google Research in Zurich, recruiting infrastructure positioned at the source of supply.

Capital migration: a European deep-tech founder at Series B finds three potential lead investors in Europe and thirty in the United States; the terms include Delaware incorporation and a US board seat; the capital comes with a change of address.

Acquisition: eighty-two per cent of transatlantic technology acquisitions flow in one direction.

Thoma Bravo acquired Sophos for $3.9 billion. Broadcom acquired VMware for $61 billion.

The acquisition channel is the most visible. It is also the least important, because by the time a European company is acquired, the other three channels have usually already captured much of the value.[8]

Stripe is the defining case. Patrick and John Collison are Irish. The engineering began in Ireland. The company is incorporated in Delaware, headquartered in San Francisco, and valued at $65 billion. By every Irish metric, Ireland produced a world-class fintech company. By every financial metric, Delaware owns it.[9]

The final stage is attribution.

ARM was designed in a Cambridge office park. It powers ninety-nine per cent of the world’s smartphones.

SoftBank acquired it for £32 billion and listed it on the NASDAQ at a $54.5 billion valuation. The British origin became a historical footnote rather than an ongoing claim on value.

European policymakers, reading their own talent’s achievements under American corporate letterheads, conclude that Europe cannot innovate. The captured innovation generates the narrative that prevents the response that would stop the capture.[10]

The Honest Admission

The pipeline works because the American ecosystem is, in several measurable respects, better at the specific things that matter at the point of decision.

For a generation of European engineers and founders, Silicon Valley was the aspiration.

The culture seemed to emanate from there: the startups, the products, the feeling that ideas could become companies in months rather than years. Some of that was marketing.

Some of it was real. The capital and legal infrastructure were already there. The tax framework rewarded founders and employees in ways that European frameworks punished.

The market was already there: 330 million consumers under one regulatory jurisdiction, one language, one set of consumer protection rules.

Every individual who entered the pipeline made a rational decision.

That is what makes the pipeline so effective and so difficult to interrupt. The engineer who moved to San Francisco was not deceived.

The founder who incorporated in Delaware was not coerced. Each chose the option that maximised their career, their equity, their impact.

The pipeline runs on rational self-interest, not on ignorance.

The question is whether these advantages are natural or constructed.

The answer, as with every layer of the dependency architecture, is that they were built. The capital depth was created by decades of pension fund deregulation and tax incentives for venture investment.

The legal infrastructure was designed to attract incorporation. The market access was a consequence of continental scale and federal regulatory uniformity.

The bait is real. And it was designed.

Two Layers

The capture has two layers, and the distinction matters because Sweden has already proved which layer can be fixed alone.

Layer one is per-company conditions: the four mechanisms that determine whether a European founder can build a viable company on European soil.

Stock option taxation.

A Berlin engineer paying income tax at exercise on a forty-five per cent marginal rate took home forty per cent less than her San Francisco counterpart on the same nominal offer.

Legal infrastructure. Delaware’s Court of Chancery has two hundred years of corporate case law; European incorporation across multiple jurisdictions imposes complexity costs that fall hardest on the smallest companies.

Market access. 450 million European consumers behind twenty-seven regulatory interpretations versus 330 million Americans behind one.

Narrative capture. The statement that America innovates and Europe regulates functions as a premise rather than a claim, and the premise produces the outcome that confirms it.[11]

Each is fixable by targeted policy.

Sweden proved this definitively. Stockholm produces 6.3 billion-dollar companies per million residents: subsidised home computer purchases in the 1990s, tax-advantaged investment accounts channelling household capital into startups, an engineering culture running from Ericsson through Volvo to Spotify.

Forty-one unicorns, per capita the most prolific producer of billion-dollar technology companies on the continent.[12]

Layer two is the macro conditions: capital depth and demand. The forces that determine whether a viable company stays once it scales.

Capital depth.

The American venture capital market deployed roughly $215 billion in 2024 against barely $52 billion across all European stages combined. A European Series C needing €100 million follows the capital to San Francisco because the cheque exists nowhere else.

Klarna raised $4.2 billion from Sequoia, Silver Lake, and Dragoneer because European capital pools were not deep enough to anchor it.[13] The mechanism that channels European pension funds, sovereign wealth, and household savings into European technology investment is examined in Paper 26.

Demand.

The customer at the end of the pipeline.

Sweden could not fix this. With ten million people and no sovereign procurement mandate, Sweden could grow companies but could not anchor them.

The companies grew because Sweden fixed enough of layer one. They left because Sweden, with no European procurement mandate behind it, could not fix layer two.

The demand-side mechanics are examined in Paper 8.

The exits tell the story.

Spotify listed on the New York Stock Exchange. Klarna’s capital came from American venture funds.

The engineer at ETH Zurich does not leave because San Francisco has better weather. She leaves because Google has a project that will ship to a billion users and the European alternative is a grant-funded programme that will produce a paper.

Fix layer one and the talent stays through the start-up phase. Fix layer two and the talent stays through scale.

What’s Buildable: Mistral, Helsing, AMI Labs

Mistral AI is the primary exception, and its story is inseparable from LLaMA’s.

Five of the fourteen LLaMA authors left Meta to build what Meta had shown was possible: a frontier AI company, headquartered in Paris, drawing on French talent. Within two years, Mistral reached a $6.2 billion valuation.

The company demonstrates that when European researchers are given the institutional conditions to build, they can compete with anyone.[14]

The conditions are revealing.

Mistral’s investors are American venture capital firms: Andreessen Horowitz, General Catalyst, and Lightspeed Venture Partners.

The company survives because there is enough European corporate demand for AI, from LVMH to Airbus to BNP Paribas, to sustain a credible alternative to American models. The demand is corporate.

If that corporate demand softens or if a larger American competitor prices below cost, Mistral’s independence depends on whether a European government steps in as anchor customer.

The exception proves the demand layer: Mistral exists because demand exists. Its fragility proves the layer is thin.[15]

Helsing is the defence-AI counterpart.

Founded in 2021 in Germany, headquartered in Munich, building autonomous systems and decision-support software for European militaries. The anchor customers are European, beginning with the Bundeswehr and extending to NATO partners.

The capital is European-led: General Catalyst, Lightspeed, and Saab.

In January 2025, Helsing raised €600 million at a $5.4 billion valuation, one of the largest defence-tech rounds in European history.

Helsing exists because European governments placed orders. The orders proved the demand was real.

The demand attracted the capital. The capital scaled the company. Each step depended on the previous step being European.

Helsing is what the Mistral pattern looks like when the demand layer is governmental rather than corporate, and it is the cleanest current case for the prescription that follows.[16]

ASML, headquartered in Veldhoven, is the exception that clarifies the boundary.

It is the only company on earth capable of manufacturing extreme ultraviolet lithography machines. TSMC, Samsung, and Intel all depend on it.

The pipeline cannot capture ASML because the technology is so specialised that acquisition would trigger national security intervention in every jurisdiction that depends on semiconductor manufacturing.

ASML survived because it cannot be absorbed. That is a condition almost no European technology company can replicate.[17]

BioNTech, founded in Mainz, tells the crisis version. It developed the mRNA vaccine technology that became one of the most consequential medical innovations in recent history. It survived as a European company because a pandemic created overnight sovereign demand that no peacetime European institution would have offered.[18]

Solvinity, a Dutch managed cloud provider, tells the most uncomfortable version.

The municipality of Amsterdam and the Dutch Ministry of Justice chose Solvinity specifically to reduce dependence on American firms and mitigate CLOUD Act exposure. The sovereign procurement decision was made.

It worked. Then in November 2025, American IT company Kyndryl announced its acquisition of Solvinity.

Amsterdam was back where it started without having changed a single decision. Sovereignty that depends on ownership is sovereignty that can be sold.

The demand layer is necessary. Without the capital and exit layer, the demand creates a company that someone else harvests.[19]

In March 2026, Yann LeCun launched AMI Labs with a $1.03 billion seed round, Europe’s largest ever, at a $3.5 billion pre-money valuation. The company is headquartered in Paris. Its founding team is drawn almost entirely from Meta’s FAIR. LeCun told MIT Technology Review that to pursue genuinely new research in AI, you have to go outside Silicon Valley, to Paris.[20]

The statement is revealing.

LeCun spent a decade building FAIR inside Meta. The institutional knowledge, the management experience, the network of researchers, the understanding of what frontier AI research requires at scale: all of it was developed on American corporate funding.

AMI Labs is structural proof of the institutional ghost thesis. When Europe finally builds the frontier AI lab it should have built a decade ago, it is built by someone who had to build it first inside an American company.

The capture had to happen before the return.

A decade later, the skills and institutional knowledge that FAIR Paris created flow back into a European company.

The value was extracted, compounded inside an American corporation, and returned at a price: $1.03 billion in capital, most of it from American investors, for capabilities that European universities originally produced. The return is welcome.

The fact that it required a decade of American corporate extraction is the diagnosis.

The Recursive Damage

The innovation capture pipeline transfers wealth. It also deepens the dependency the rest of this series documents.

Each capture extends the architecture.

European startups that incorporate in Delaware place their intellectual property under US jurisdiction. European companies acquired by American firms bring their customer relationships, their government contracts, and their technical talent under CLOUD Act reach.

Each captured company becomes a node in the dependency network (Paper 6). Each Delaware incorporation adds another entity subject to American disclosure law.

Each acquisition extends American jurisdictional reach into European infrastructure.

A European AI researcher hired by Google trains the models that outcompete European alternatives. A European cybersecurity firm acquired by Thoma Bravo provides the tools that European governments depend on.

The pipeline is accelerating, and the acceleration compounds.

AI talent is more valuable, more mobile, and more concentrated than any previous generation of technical talent. Each captured researcher trains the next generation of American AI systems, which outperform European alternatives, which reinforces the narrative that America innovates and Europe regulates, which shapes the next researcher’s career decision.

The pipeline does not just extract value from Europe. It uses the extracted value to widen the gap that makes the next extraction easier.

The recursive damage is the pipeline’s most effective feature and the one least visible to policymakers who measure brain drain by counting departures.

Conclusion

The innovation capture pipeline has two layers.

Layer one is per-company conditions Europe can fix through targeted policy. Sweden proved it: forty-one unicorns, exits still American.

Layer two is the macro conditions of capital and demand. Sweden could not fix this alone.

Capital depth requires European pension funds, sovereign wealth, and household savings flowing into European technology investment at scale.

Demand requires anchor customer mandates that turn compliance laws into procurement orders. Both mechanisms are owned by dedicated papers in this series.

The LLaMA team is scattered now. Three remain at Meta. Five are at Mistral. Others went to Cohere, Anthropic, Kyutai. The talent that France produced, that France educated, that France retained on French soil in a French office, belongs to a dozen different jurisdictions. The researchers did not fail. The institutions around them failed to build what needed building and buy what needed buying.

I understand the pull because I felt it. I grew up wanting to go to California. Most European technologists of my generation did. The pull was real, and the reason was structural: Europe could not match what America offered ambitious technologists, so the choice was between European life and American opportunity. The choice did not have to exist.

What has shifted is the recognition that the choice was never about value. Both societies are real achievements. The American model produces extraordinary companies. The European model produces a place where ordinary people can live without fear. They are not the same achievement. The choice between them was forced by missing institutions.

The fix is mechanical.

Match European ambition with European institutions, and the choice between Europe and getting rich stops being forced. The talent that already chose Europe stays, because Europe becomes a place where ambition can also win.

The companies that European education produces stay European, because the capital, the customer, and the legal infrastructure all sit on the same side of the Atlantic. The demand-side coalition is starting to organise; the operational evidence is in Paper 26.

The most sophisticated capture is the one nobody complains about.

Match European ambition with European institutions. Build a Europe where ambition stays.

[1] Demis Hassabis studied at Queens’ College, Cambridge and completed his PhD at University College London. He co-founded DeepMind Technologies in London in 2010. Google acquired DeepMind in January 2014 for approximately £400 million.

[2] World Wide Web: Tim Berners-Lee, CERN, 1989. MP3: Fraunhofer Institute for Integrated Circuits, Erlangen, patented 1989, commercialised through Apple’s iTunes Store (2003). Bluetooth: Ericsson, Lund, Sweden, 1994.

[3] Yann LeCun was appointed Director of AI Research at Facebook (now Meta) in December 2013. The FAIR Paris laboratory was established in 2015 under his leadership.

[4] Google Inc. acquired DeepMind Technologies Ltd for a reported £400 million (approximately $625 million) in January 2014. The 2024 Nobel Prize in Chemistry was awarded jointly to David Baker, Demis Hassabis, and John Jumper for protein structure prediction (AlphaFold).

[5] Estimate based on UK Research Council PhD funding (stipend, fees, and institutional costs) over a typical 3.5-4 year programme. Actual costs vary by institution and discipline.

[6] Communications Decency Act of 1996, 47 U.S.C. §230(c)(1).

[7] Center for Security and Emerging Technology (CSET), Georgetown University, Keeping Top AI Talent in the United States, 2019. The study found that approximately 70 per cent of leading AI researchers at US institutions were born or educated abroad.

[8] Thoma Bravo completed its acquisition of Sophos Group plc for approximately $3.9 billion in March 2020. Broadcom Inc. completed its acquisition of VMware Inc. for approximately $61 billion on 22 November 2023.

[9] Stripe Inc., co-founded by Patrick and John Collison (both from Limerick, Ireland) in 2010. Incorporated in Delaware, headquartered in San Francisco. Valuation reached $65 billion in February 2024.

[10] ARM (now Arm Holdings plc) was founded as Advanced RISC Machines Ltd in Cambridge, UK, in 1990. ARM-based processors power approximately 99 per cent of the world’s smartphones. SoftBank Group acquired ARM for £32 billion in September 2016. ARM listed on the NASDAQ on 14 September 2023 at a valuation of approximately $54.5 billion.

[11] The Delaware Court of Chancery, established 1792, is the preeminent US court for corporate law disputes. Delaware is the state of incorporation for more than 60 per cent of Fortune 500 companies.

[12] Swedish unicorn count as reported at the time of writing. Sweden’s technology ecosystem has produced among the highest per-capita rates of billion-dollar technology companies globally.

[13] Klarna Bank AB (Stockholm) raised cumulative funding of approximately $4.2 billion across multiple rounds. Key investors include Sequoia Capital, Silver Lake, and Dragoneer Investment Group. US venture capital deployment data: National Venture Capital Association annual reports. European venture capital data: Atomico State of European Tech and Invest Europe.

[14] Mistral AI, Series B funding round of €600 million ($645 million) closed June 2024, at a valuation of approximately $6.2 billion (€5.8 billion).

[15] Mistral AI’s Series A (December 2023) was led by Andreessen Horowitz with General Catalyst and Lightspeed Venture Partners. Series B (June 2024) was led by General Catalyst.

[16] Helsing GmbH, founded 2021, headquartered in Munich. Series C of €450 million in July 2024 at approximately $5.4 billion valuation, led by General Catalyst with Lightspeed, Saab, Accel and Plural. Series D of €600 million in June 2025 at €12 billion valuation. Software deployed with the Bundeswehr, French armed forces and UK Ministry of Defence.

[17] ASML Holding N.V., headquartered in Veldhoven, Netherlands. ASML is the sole commercial manufacturer of extreme ultraviolet (EUV) lithography systems used in advanced semiconductor fabrication.

[18] BioNTech SE, founded in 2008, headquartered in Mainz, Germany. The company developed the mRNA vaccine technology licensed to Pfizer for the BNT162b2 COVID-19 vaccine.

[19] Kyndryl Holdings Inc. (spun off from IBM in 2021) announced the acquisition of Solvinity Group B.V. in November 2025.

[20] Yann LeCun, interview with MIT Technology Review, January 2026, discussing the rationale for establishing AMI Labs in Paris.