Patrick de Carvalho
EN
Français Português English
All articles
Patrick de Carvalho May 25, 2026 · 31 min

The French taxpayer pays for the research, the Americans keep the patent

When the head of Mistral AI uses the word vassal in front of French MPs, it is not a figure of speech. The figures prove him right, and then some.

By Patrick de Carvalho, CEO Apps Velocity

Contents

Special report: Choose France, the research tax credit, nuclear power. Anatomy of a consented vassalage.

By Patrick de Carvalho

"In AI, France must act now if it is not to become a vassal." Arthur Mensch, chief executive of Mistral AI, speaking before French MPs, May 2026.

When the head of France's only artificial intelligence champion uses the word "vassal" in front of the national parliament, it is not a figure of speech. It is a field observation. And I am going to spend the next few pages showing you, figures in hand, that he is right, and that he is in fact understating the case.

I am not an engineer. I have been building tech companies for thirty years without ever writing a line of production code. My job is to read what technology does to company accounts, to territories, and to the balance of power between nations. And what I am reading in the news this June 2026 makes me angry. Not slogan anger. Anger built on figures.

In 1998, I launched one of the first online newsstands in France. Back then the debate was not digital sovereignty, it was whether the Internet was a fad. I saw, from the inside, how a technological wave redistributes the cards of economic power in the space of a few years. I saw dominant positions that were supposed to hold for decades tip over because one player had understood before the others where the real value sat. And the lesson I took from it, the one I have been repeating ever since, is that in every technological rupture there are those who own the strategic asset and those who merely rent it. The first group wins. The second pays.

Today, in the age of artificial intelligence, the strategic asset has a threefold name: brains, energy, data. And I am going to show you how France, which owns all three, is handing them over one by one to players who owned none of them to start with.

In short: France funds AI research through the research tax credit (Crédit Impôt Recherche, or CIR, the French tax break that reimburses part of a company's research spending), sells its low-carbon electricity cheaply to foreign data centres, and lets its public data drift towards players who neither funded nor produced those assets. The three levers of power in the AI era are being given away at the same time, and each one through a political choice, not through economic fate.

Act 1: Salesforce in Grenoble, or the invisible mechanism

Let us start with a concrete case, because that is where everything becomes clear.

Salesforce is the American publisher of customer relationship management software, CRM (Customer Relationship Management, the software that runs the relationship between a company and its customers). Head office in San Francisco. Market capitalisation of several hundred billion dollars. The group owns two research and development centres in France, R&D being the work of inventing and improving products: one in Grenoble, opened in 2013, a second in Paris. Researchers there work on the group's platform and on advanced search and integration features.

Excellent news, some will say. An American giant doing research on French soil, employing French engineers, strengthening our ecosystem. And that is true. I am not going to deny it, it would be dishonest, and honesty is the only thing that makes a case for the prosecution credible.

But let us look more closely at the mechanism, the one nobody mentions in the triumphant press releases.

When Salesforce, or Google, or Microsoft, or Amazon sets up an R&D centre in France and employs researchers there, that group is legally entitled to the research tax credit. The CIR is not reserved for French companies. Foreign structures can claim it as soon as they have taxable activity in France and carry out eligible research work there, provided the spending is incurred on French territory. It is written down, it is legal, it is documented.

In concrete terms, the French state reimburses 30% of these subsidiaries' R&D spending below the 100 million euro threshold. The largest line in that spending is the payroll of researchers and engineers, often 60 to 80% of the base. Unfiltered translation: part of the salary of the French engineers who work for Salesforce is paid by the French taxpayer.

And now the uncomfortable question, the one at the heart of this dossier. Those French engineers, trained free of charge by the schools of the Republic, paid in part by French taxes, produce what exactly? Patents, code, intellectual property. And who owns that intellectual property? Head office. San Francisco. It does not stay in France. It crosses the Atlantic at the speed of an intra-group transfer, and it goes on to feed the stock market valuation of a company that will, when the time comes, sell its software to French companies at full price.

France pays for the research. The Americans keep the patent. That is the sentence that sums up everything, and it is exactly what I put in the title.

Act 2: the research tax credit, X-ray of a tax break that went out of control

To grasp the scale of the thing, you have to look the scheme in the face, without the talking points from Bercy, the French finance ministry.

The research tax credit was created in 1983. The original idea was noble, even brilliant: reward companies that increase their research effort, in order to close the French gap in private innovation. For twenty years the mechanism was incremental, it rewarded the additional effort. Logical. Virtuous.

Then came the 2008 reform, under Nicolas Sarkozy. The CIR switched to a volume basis: you no longer reward the additional effort, you reimburse 30% of all R&D spending below 100 million euros, and 5% above that. The change looks technical. It is in fact seismic.

The figures speak. The CIR tax claim rose from 1.8 billion euros in 2007 to 4.45 billion in 2008, the very year of the reform. Then it never stopped climbing: 6.8 billion in 2018, 7.2 billion in 2023, 7.65 billion in 2024. The Cour des comptes, France's national audit office, puts the shortfall at 8.06 billion euros in 2025. Applying the old accounting convention, the Court even estimates that the cumulative cost would reach historic highs.

Remember that figure: eight billion euros a year. Since 2023 it has been the most expensive tax expenditure of the French state. More expensive than any other tax break. Eight billion that do not reach the public purse each year, in a country that is desperately looking for budget savings and closing hospital beds.

And where does that money go? That is where the scandal really begins.

Act 3: who collects the eight billion

The CIR was sold to the French public as support for innovation. Innovation by everyone: the start-up in Montpellier, the industrial SME (small or medium-sized enterprise) in the Vendée, the biotech laboratory in Lyon. In the speeches, the CIR helps the small players who are inventing the future.

In practice, it is the opposite.

The top fifty beneficiary companies alone account for nearly 45% of the scheme. The top 200 capture two thirds of it. According to the Commission nationale d'évaluation des politiques d'innovation, the CNEPI (the national body that evaluates innovation policy), which sits within France Stratégie, the government's policy planning agency, the largest 10% of beneficiaries receive 77% of the total amount. And according to the most recent detailed report available, large companies, 492 of them, received 48% of the CIR on their own.

Meanwhile SMEs, which make up around 80% of beneficiaries by number of claims, take only around a third of the amounts. More than 10,000 companies, half of all beneficiaries, share a third of the claim. Crumbs for the many, the big slice for a handful of giants.

That concentration is not an accident. It is written into the very design of the scheme. The ceiling on eligible spending is calculated company by company, and can therefore be stacked within a group. In practice, large groups spread their R&D spending across several subsidiaries in order to limit the application of the reduced 5% rate and maximise the full 30% rate. That is pure tax engineering, available only to those with armies of tax specialists to practise it. An SME does not have that capacity. A global giant does.

Let me be precise, because I refuse to take the easy road: a large share of those big beneficiaries are French groups. Sanofi, Stellantis, the defence and aerospace players. The CIR is not a tap reserved for Americans. But the mechanism it puts in place is exactly the one that favours the biggest and the most international players, those with the means to optimise. And among them, the subsidiaries of the American giants figure prominently, because the United States is by far the leading foreign provider of jobs in France, with more than 504,000 jobs, ahead of Germany and Switzerland.

Back to my lesson from 1998. In every rupture, there are those who own the asset and those who rent it. The CIR, as currently designed, is a machine for transferring public money to those who already own everything: the scale, the tax specialists, the dominant positions. It was meant to be a springboard for the small players who invent. It has become an annuity for the big players who dominate. It is the exact opposite of its founding intention, and that inversion did not happen by chance. It happened because the 2008 reform turned an incentive scheme into a volume scheme, and volume, by definition, benefits whoever has the biggest volume.

Eight billion euros a year. The question is not whether research deserves support, it does. The question is whether we are supporting research or fattening margins. The figures answer: we are fattening margins, and increasingly the margins of players whose centre of gravity is not in France.

Act 4: the windfall effect, or the intellectual con at the heart of the scheme

Here now is the concept that every company director and every citizen should have in mind. I will define it simply.

A windfall effect is when the state pays a company to do something it would have done anyway, without the subsidy. The public money triggers no new decision. It merely inflates a margin that already existed.

The 2021 CNEPI report, commissioned by France Stratégie, is damning on this point. For micro-businesses and SMEs, the CIR has a real and measurable effect: one million euros of CIR generates around 1.165 patents in small companies, against only 0.464 in large ones. A return two and a half times higher for the small ones. When you look at high-quality patents, the so-called triadic patents filed in Europe, the United States and Japan at once, the gap widens further.

For large companies, on the other hand, the report identifies a devastating threshold effect: they benefit from a high subsidy rate on investments that would have been made regardless of the CIR. In other words, on the most expensive part of the scheme, the part captured by the giants, public money does not create additional research. It retrospectively subsidises research that would have happened anyway.

This is not me speaking in a fit of militancy. It is written in black and white in a written question to the French National Assembly: the CIR is mainly used by large groups, within which it represents a powerful windfall effect that does not translate into any significant increase in research effort. The legislator itself knows this. And it changes nothing.

A global giant is, by nature, obliged to do R&D in order to survive. Salesforce would do research with or without the CIR, because without research, Salesforce dies. The French subsidy does not change its decision to invest. It just improves its bottom line. We are subsidising the profitability of companies that have no need of us in order to exist.

Act 5: the coup de grace, the CIR does not even attract what it claims to attract

We were sold the CIR as a magnet for foreign R&D. Thanks to its generous tax regime, France was supposed to have become the honeypot of laboratories the world over.

That is false. And it is the CNEPI, again, that demonstrates it.

The CIR did not manage to halt the decline in the attractiveness of France as a location for the R&D of foreign multinationals. Their share of national research spending actually fell from 22.6% in 2005 to 19.6% in 2016, with a particularly marked loss of interest from the big American technology groups, which invested massively elsewhere in Europe.

Read that carefully, because that sentence dismantles two lies at once.

First lie, the official one: the CIR supposedly attracts foreign R&D. False, the foreign share has fallen.

Second lie, the one I could have committed myself out of laziness: the Americans supposedly come to France for the CIR first. Also false. The interviews conducted by the CNEPI with company directors show that public aid is not the determining factor in where R&D is located. What really counts is access to talent and to innovation ecosystems.

And this is where the case for the prosecution becomes unanswerable. The American giants do not come to France for the CIR. They come for our brains. And once they are here, capturing our brains, they collect the CIR on top, as a bonus. We are not attracting them with the money. We are giving them the money as a premium, when they would have come anyway to hunt for our engineers. It is the worst of both worlds.

Act 6: the second tap, the energy we are selling off cheap

The CIR is only the first stage of the rocket. The second, brought into full view by the Choose France summit this June 2026, is even more strategic. It is energy.

On 1 June 2026, at the Palace of Versailles, Emmanuel Macron announced a record figure: 93 billion euros of foreign investment, more than 15,000 jobs expected, across 71 announcements. This single vintage exceeds the sum of the eight previous editions of Choose France, which together totalled 87 billion since 2018.

The centrepiece is SoftBank, the Japanese technology investment giant: 45 billion euros earmarked for data centres in the Hauts-de-France region by 2031. Add Brookfield, the Canadian firm, for digital infrastructure. The Emirati fund MGX with Bpifrance, France's public investment bank, 7.5 billion for a computing site. Amazon, logistics sites. Salesforce, 2 billion dollars for its cloud.

A data centre is a warehouse full of servers that runs the cloud and artificial intelligence. Those machines have one characteristic: they devour electricity. And this is where France plays its trump card, the one no other major European country holds.

France has low-carbon electricity that is abundant and cheap, thanks to its nuclear fleet. It is the structural French differentiator against Germany, which has been paying far more for its electricity since it closed its power stations. That nuclear electricity, paid for and built by decades of French public investment, has become the magnet drawing data centres from all over the world.

The figures on that appetite are dizzying. The electricity consumption of French data centres jumped by 38% in three years, according to the Arcep, France's electronic communications regulator. AI-related uses, which accounted for only 2% of demand in 2021, could reach nearly 90% of it by 2030. RTE, the operator of the French electricity transmission grid, is facing around 14 gigawatts of connection requests for those projects alone. If all of them went ahead, their consumption would climb to around 80 terawatt hours a year, eight times more than today. RTE has in fact made this its number one grid investment priority, with a programme costed at 100 billion euros over ten years.

One hundred billion euros of public money to adapt the French electricity grid in order to supply, as a priority, the servers of foreign giants. You read that correctly.

Act 7: the vice closing on Mistral, or how to strangle your own champion

This is where the case for the prosecution reaches its most painful and its most unassailable point, because it is a matter of physical arithmetic.

The electricity on a grid, at any given moment, is a finite resource. Every gigawatt of connection capacity granted to an American hyperscaler, one of the giant global cloud operators, or to a foreign fund, is a gigawatt that will not go to a French player. The grid does not duplicate itself by magic. In the short term it is a zero-sum game.

And while SoftBank is putting down 45 billion euros and capturing massive capacity in the Hauts-de-France, look at the scale on which our national champion is fighting.

Mistral AI, the only French company capable of competing with OpenAI and the American models, had to raise 830 million dollars in debt in March 2026 to fund its first data centre on French soil, equipped with 13,800 Nvidia graphics processing units, the chips used to train AI models. The capacity of that first site: 44 megawatts. Mistral's target for 2027: 200 megawatts across Europe. For 2030: one gigawatt, the equivalent of the output of a single nuclear power station.

Compare the two. One gigawatt targeted by Mistral for 2030. Fourteen gigawatts of connection requests already on RTE's desk, the bulk of them for foreign players. The French champion is fighting to scrape together, megawatt by megawatt, a fraction of what the foreign giants obtain in a single announcement at Versailles.

Mistral's chief technology officer says it plainly: computing power and its scarcity are a source of concern for them. The sizing of AI infrastructure now depends as much on access to electricity as on chip availability. And what are we doing with that electricity, our only real competitive advantage? We are rolling it out as a red carpet in front of those who will crush Mistral tomorrow.

Because that is exactly what is at stake. The giant data centres we are powering today will train the American AI models that will dominate the world market. Once that dominance is in place, what room will be left for a French or European alternative? With our energy and our subsidies, we are funding the crushing of our own capacity to exist in AI.

We need to name what economists call the opportunity cost: the value of what you could have done with a resource, and will not do because you committed it elsewhere. One hundred billion euros of grid investment to connect foreign data centres as a priority is one hundred billion that does not, as a priority, strengthen our industry, our electric transport, our own AI sector. Fourteen gigawatts of capacity directed towards servers whose profits will flow back to foreign head offices is that much capacity not going to French reindustrialisation, at the very moment when RTE is forecasting a massive rise in national electricity consumption by 2035 under an ambitious reindustrialisation scenario.

We are pre-allocating our energy future, our most strategic resource for the next thirty years, to uses that do not consolidate our sovereignty but dilute it. And we are doing it at the precise moment when that energy is becoming the limiting factor in the global AI race. French low-carbon electricity is, in 2026, the equivalent of what oil was in the twentieth century: a geostrategic asset. You do not sell off a geostrategic asset. You reserve it, you negotiate with it, you turn it into a lever of power. France, for its part, is handing it out with a fanfare at Versailles.

Act 8: data sovereignty, the final lock

And even if Mistral succeeded, even if the energy followed, there would remain the deepest lock of all: data.

The three leading cloud providers in the world are American, and their dominance is not an impression, it is a number. In the third quarter of 2025, according to the Synergy Research Group, Amazon Web Services held 29% of the global market, Microsoft Azure 20% and Google Cloud 13%, meaning 63% between the three of them in a market worth 107 billion dollars. In France the lock is even tighter: a Markess study noted as early as 2022 that AWS alone accounts for around 46% of market share there, leaving only around 20% to all French, European and other providers combined. Our largest companies are already tied in: BNP Paribas, Renault and EDF have made Azure their reference cloud, while Canal+, Air France-KLM, Accor and Michelin are among the French customers of AWS. The state itself is no exception. The problem is not only commercial, it is legal. Data hosted with an American player, even when physically stored in France, remains exposed to extraterritorial American law, which allows Washington to demand access to it. Our health data, our industrial secrets, our administrative files can therefore fall under a jurisdiction that is not our own.

This is not a lawyer's hypothesis, it is a documented affair of state. In 2019, France entrusted the hosting of the Health Data Hub, the repository that centralises the health data of 67 million French citizens, to Microsoft Azure. Without a tender. The choice immediately triggered controversy, and the Commission nationale de l'informatique et des libertés, the CNIL (France's data protection authority), refused to authorise the wholesale transfer of national health insurance data to that platform, citing the risk of access by the American authorities. The case dragged on for seven years. It took a 2024 law imposing "guaranteed sovereign" hosting for sensitive state data, then a selection procedure, for the government to finally entrust that data, in April 2026, to Scaleway, the French subsidiary of the Iliad group.

That episode says two things at once. First, the scale of the dependency: it took seven years to repatriate the health data of an entire nation. Second, the decisive lever, the SecNumCloud label issued by the French cybersecurity agency, which requires that a host be subject to no extra-European legislation. That single criterion automatically rules out Microsoft, Amazon and Google, wherever their servers are located. Data sovereignty is therefore not a pious wish. It is a standard that exists, that works, and that simply needs to be applied. Provided you actually want to, and do not wait seven years.

Data, in the age of AI, is not a by-product. It is the fuel. Models train on it in order to improve, and whoever controls the data flows controls the raw material of tomorrow's AI. Yet we are letting that raw material flow towards foreign infrastructure, while funding, with our energy and our research tax credit, the players who will refine it.

That is the complete chain of vassalage. We train the engineers. We pay part of their salary through the CIR when they serve the American giants. We supply the nuclear electricity that powers their data centres. And our data itself escapes our control.

Brains, energy, public money, data. We are giving away all four. What are we keeping?

Act 9: the honest objection, and why it is not enough

I now have to give the defence its turn, because a case for the prosecution that does not confront the objections is nothing but a sermon.

First objection: these investments create jobs. True, but let us look at which ones. 93 billion euros for 15,000 jobs works out at one job for roughly 6 million euros invested. For conventional industrial projects, that ratio would be judged catastrophic. The reason lies in the nature of data centres: for an equivalent floor area, a logistics warehouse employs 300 to 400 people, a data centre between 15 and 20. One researcher calculated a ratio of one job per 24 million euros invested in a data centre. And the promises are often inflated: a senator for the Essonne noted that in Seine-Saint-Denis, out of 1,600 posts announced by the developers, only 500 were actually created.

The operators reply, quite rightly, that these jobs are good ones: durable, local, impossible to offshore, 80% of them on permanent contracts (CDI, the French open-ended employment contract). That is accurate. You cannot splice an optical fibre remotely. But you do not offset the loss of 700 industrial jobs by creating a few dozen maintenance jobs. In La Courneuve, where a factory employed 700 people, the data centre that took its place employs a few dozen.

Second objection: by setting up here, Salesforce creates a French ecosystem of customers, partners and developers, and therefore local value. That is also true. I am not claiming that these giants leave nothing behind. I am asking the question of the net balance. When you add up the cheap electricity, the CIR collected, the engineers captured, the intellectual property that leaves and the data that escapes, and you subtract the few thousand jobs and the local ecosystem, does the balance really tip in favour of France? I do not have mathematical proof of the contrary. But nobody, at Bercy or at the Élysée, has proof of the opposite either. And that is precisely the problem: we announce billions as victories without ever calculating what the nation owns at the end of it all. An investment is not a blessing because of its size. It is a blessing because of what it leaves behind once the cameras have gone and the ribbons have been cut. And nobody talks about that balance, because it is less glorious than a record figure displayed on a lectern at Versailles.

Third objection, the most serious one: even our champion depends on the Americans. The AI campus at Bruyères-le-Châtel, presented as a spearhead of French sovereignty, rests on a joint venture bringing together Nvidia, Mistral, Bpifrance and the Emirati fund MGX. The international partners bring the capital and the chips, the French players guarantee the local anchoring and compliance with the GDPR, the European data protection regulation known in France as the RGPD. Even Mistral runs on Nvidia processors, an American company. Sovereignty is partial, and that has to be said. But that objection does not argue for giving up. It argues for urgency. If even our best asset is half dependent, then every megawatt, every euro of CIR, every location decision we hand to the giants instead of reserving it for our own players makes an already critical gap worse.

Act 10: this is not fate, it is a choice

I come now to my conclusion, and it rests on one distinction that I want to engrave in the mind of every company director and every policymaker who reads me.

France is not being plundered. You do not plunder a country that sells willingly. We invoice our electricity, we collect tax, we sign every agreement in full awareness. The word "plunder" would be false, and one false word brings down a true demonstration.

Nor is France a colony. A colony has no choice and cannot legislate against the occupier. We, on the other hand, could decree a moratorium on data centres tomorrow morning, and that is in fact one of the scenarios studied by the Ademe, France's environment and energy agency. We can refocus the CIR on SMEs, something proposed to Parliament every year. We can reserve priority connection capacity for our champions. We have every lever.

What we are is what the head of Mistral dared to name in front of the MPs: vassals. And vassalage, unlike plunder or colonisation, is a voluntary status. The vassal has sworn allegiance. He chose. And what one choice has made, another choice can unmake.

In two decades we have built a perfect harness in the service of the interests of the American giants. The CIR reimburses them for the research our brains produce. Our nuclear power feeds the machines that will train the AI models designed to crush our own. Our clouds depend on their infrastructure. And we celebrate all of it, every June, under the gilding of Versailles, as a victory for French attractiveness.

Attractiveness is not an end in itself. Attracting billions means nothing if the wealth created leaves again straight away. The real question was never "how much did they promise". The real question, the only one, is: "what does France own at the end of it all". And on that question, today, the account does not add up.

There is still time. The levers are in our hands. But we have to stop confusing the volume of announcements with the creation of value, and the red carpet with strategy. A nation that sells off its only competitive advantage, its energy, while subsidising the profitability of those who will dominate it, is not doing attractiveness. It is organising its own dependency.

I never lose. Either I win, or I learn. France, right now, is learning an expensive lesson. It is up to France to decide whether it draws the conclusions, or whether it goes back to Versailles next year to applaud the next record.

What company directors can take from this, concretely

This dossier is not only a piece of macroeconomic analysis. It carries a lesson that is directly actionable for every director of a small, medium or mid-sized company reading me.

National vassalage is the sum of thousands of company-level vassalages. Every time a French company hands over the entirety of its data, its AI and its infrastructure to a player whose head office and intellectual property sit elsewhere, it reproduces at its own scale the pattern I have just described. It rents its computing capacity, it feeds a model it does not control, and it makes itself dependent on a technology it owns nothing of.

The opposite strategy exists. It consists of using AI as a multiplier of your own expertise, not as a dependency. Of keeping control of your data. Of building value that you own, instead of renting it. That is exactly the work I do with the RAPID method, designed for directors who want to deploy AI without becoming anyone's vassal.

If this dossier made you as angry as it made me while writing it, let us turn that anger into strategy. For your teams, for your ecosystem, for your region.

AI strategy for company directors: rapid.appsvelocity.com AI keynotes: patrickdecarvalho.com/fr/conferences

I never lose. Either I win, or I learn.

FAQ

What is the French research tax credit?

The research tax credit, or CIR (Crédit Impôt Recherche), is a corporate tax reduction granted to companies that incur research and development spending in France. It covers 30% of eligible spending below 100 million euros, and 5% above that. It is France's largest tax expenditure: 8.06 billion euros in 2025, against 4.45 billion in 2008, the year of its reform.

Can the subsidiary of a foreign group claim the CIR?

Yes. The eligibility criterion is the place where the research is carried out, not the nationality of the parent company. A French subsidiary of an American group doing R&D on French territory is entitled to it on the same basis as a French company. No public statistics, however, break the amounts down by nationality of the parent company: this dossier therefore puts no figure on that share.

Does the CIR really attract foreign research to France?

The available evaluation work does not show that it does. France's share of global research fell from 22.6% in 2005 to 19.6% in 2016, over the very period in which the scheme tripled in size. The France Stratégie report and the analyses by the French Senate conclude that there is a massive windfall effect on large groups, meaning public funding of research that would have taken place anyway.

Who really collects the eight billion?

Concentration is the striking fact. The largest 10% of beneficiaries receive 77% of the total amount, and large companies capture 48% of the scheme on their own. Conversely, 80% of beneficiaries by number of claims share only a minority fraction of the envelope.

Why has energy become a sovereignty issue for AI?

Because a data centre, the warehouse of servers that runs AI models, consumes a huge amount of electricity and creates very few jobs per euro invested. The electricity consumption of French data centres rose by 38% in three years according to the Arcep. France attracts these sites with low-carbon, cheap electricity, without any counterpart on the ownership of the models trained there.

What is the Cloud Act and how does it concern French data?

The Cloud Act is an American law that allows the United States authorities to require an American company to give access to data it holds, including data stored outside American territory. Since Amazon Web Services, Microsoft Azure and Google Cloud together account for 63% of the global market, the question is not a theoretical hypothesis but a live contractual matter.

In practice, what can the director of an SME do?

National vassalage is the sum of thousands of company-level vassalages. At the scale of an SME, the answer comes down to three decisions: know where your data lives and under which law, use AI as a multiplier of your expertise rather than as a dependency, and build value that you own instead of renting it.

Sources

Choose France 2026 (announcements, amounts, jobs)

Cost, history and concentration of the CIR

Windfall effect and attractiveness for foreign R&D

The CIR and foreign companies, the Salesforce case

Energy, data centres and employment

Mistral AI, sovereignty and energy

Cloud market shares (American dominance)

Data sovereignty, the Health Data Hub case

Methodological note

No public statistics break down the amount of the CIR by the nationality of the beneficiaries' parent company. This dossier therefore states that subsidiaries of foreign groups, American ones included, are eligible for the CIR and benefit from it in respect of the R&D they conduct in France (an established fact), without putting forward any figure broken down by nationality (unpublished data).

The infrastructure and logistics investments announced at Choose France are not, as such, eligible for the CIR: the CIR concerns research and development spending only. This dossier treats those two mechanisms as two distinct levers of the same dynamic of dependency, without conflating them.

Want the rest?

Join the list. You'll get new articles and the behind-the-scenes, spam-free.

Join the list

Also worth reading

Patrick de Carvalho August 27, 2026

The AI Act: four questions to answer before 2 December

Patrick de Carvalho August 25, 2026

Hugging Face: 13 billion dollars for the choke point of global AI

Patrick de Carvalho August 21, 2026

8.3 billion virtual humans: I read the study everyone is commenting on without having opened it