Part IThe Story
The oath came first. Before Arthur Mensch could say anything about electricity, or vassals, or the trillion euros a year he believed Europe was about to send across the Atlantic, the acting chair of the National Assembly's commission of inquiry into digital vulnerabilities reminded him of Article 6 of the ordinance of 17 November 1958. That article is the procedural spine of every French parliamentary investigation. She asked him to declare any public or private interest that might color his testimony, and then she administered the oath. It was May 12, 2026. The commission's chairman, Philippe Latombe, had sent his apologies and would arrive partway through. Beside Mensch sat Audrey Herblin-Stoupé, Mistral AI's director of public affairs and communication. Before giving him the floor, the acting chair offered a sentence that was half compliment and half burden: among generative AI tools, Mistral was the exception, a European open model when every major competitor was extra-European, and a great deal of hope rested on him.
He was thirty-three. He had a baby face that Politico had once found worth mentioning, and he spoke in the flat, unhurried French of a man more at home with a whiteboard than a lectern. The first substantive thing he did under oath was fix a date. "Nous avons fondé Mistral AI le 28 avril 2023." Until then the date had circulated in several forms. Biographies said May. The Financial Times, on June 13, 2023, called the company "four weeks old," which pointed somewhere in mid-May. Under oath it became April 28, with about fifteen people in France and England, nearly all of them researchers. The founders had looked at the large American players they knew from the inside and seen an oligopoly forming.
Then he talked for nearly an hour and a half, and most of it was not about software.
"We transform electrons into tokens," he said, and kept saying it, in slightly different forms, until it became the line every summary would lead with. One gigawatt of data-center capacity cost roughly $50 billion over five years and produced roughly $20 billion a year in tokens. Mistral itself now spent the equivalent of 10 percent of its payroll on AI for its own employees. Apply that ratio to Europe's payroll and you got about a trillion euros a year. If those tokens were imported, that was a trillion euros of additional trade deficit, funding someone else's next round of research. He estimated AI would require something like a kilowatt per person within five years: 40 gigawatts to build in France, 400 across Europe, around €20 trillion of investment. France had an average electricity surplus of roughly nine gigawatts. American hyperscalers were deploying a trillion dollars a year and would lock up that surplus if Europe did not organize its own demand first. "It all comes down to the next two years," he told the deputies. His closing warning, as recorded in the English summaries, was this: if Europe combined AI strength with electrical capacity it could regain a sustainable market share, and "we absolutely must do it, because otherwise we will become a vassal state."
The vocabulary had changed. Three years earlier, Mensch had been a research scientist in DeepMind's Paris office, best known in his field as one of twenty-two authors on a paper about how much data a language model should eat. Now he was testifying about gigawatts and trade balances, about public procurement as 50 percent of European GDP, about helium supplies for semiconductor fabs. He was the son of a physics teacher, and he was explaining to the French Republic, under oath, that intelligence was a form of electricity.
He did not resolve the contradictions of his position in that room, and he has not resolved them since. He runs a company whose most celebrated early act was giving its product away, and which now sells to the Ministry of the Armed Forces. It was founded to resist American concentration, it was financed in its first years largely by American venture capital, and its most important distribution partner is Microsoft. It is championed as a national cause and was built by a man whose professional formation happened at Google. He is a mathematician who says he cannot sell, and he has raised more equity than any technology company in European history.
By the Numbers
Mistral AI, April 2023 – September 2026
28 Apr 2023Founding date, as stated by Mensch under oath before the National Assembly
€105MSeed round raised by a four-week-old company with no product (June 2023)
€21B+Post-money valuation after the Samsung-led €3 billion Series D (September 2026)
11%Stake taken by ASML for €1.3 billion, making it Mistral's top shareholder (September 2025)
1,000Employees declared under oath in May 2026, up from roughly 25 in the first year
€1BRevenue target for end-2026, with another €1 billion committed to R&D that year
1 GWMistral's own compute target for 2029, which Mensch himself called 'not enough'
A Physics Teacher's Son Learns to Count
The outline of the origin is short, and he has kept it that way. Arthur Mensch was born on July 17, 1992, in Sèvres, the porcelain town on the Seine just southwest of Paris. He grew up next door in Ville-d'Avray, in the Hauts-de-Seine, the department where France's managerial class tends to raise its children. His father was a businessman. His mother taught physics. Le Point, in one of the few interviews where the family is mentioned at all, gives no more than that, and he has offered little else. A 2024 portrait in Les Échos carried the headline "Arthur Mensch, père discret derrière Mistral AI." The phrase was read as a report that he had become a father that year. It could equally describe a man who fathered a company and kept his own name out of the story. Both readings fit.
He entered the École Polytechnique in 2011. It is the military-run engineering school whose motto, Pour la Patrie, les Sciences et la Gloire, still sits over its graduates like an unpaid debt. He went on to Télécom Paris and then to the Mathématiques, Vision, Apprentissage master's at the École normale supérieure Paris-Saclay. The MVA is the program French machine learning uses to sort its most mathematical students from everyone else.
The doctorate is where the instincts became visible. From 2015 to 2018 he worked in the Parietal team, a joint unit of Inria and the CEA's NeuroSpin center at Saclay. NeuroSpin houses some of the most powerful MRI magnets in Europe. His supervisors were Bertrand Thirion, Gaël Varoquaux, and Julien Mairal. The problem was functional MRI, the three-dimensional scans that record which parts of a living brain draw blood while a subject performs a task. These datasets are enormous and noisy, and the signal is faint and smeared across hundreds of thousands of voxels. In practice you have far more numbers than meaning. Mensch's work concerned how to process such data at scale: stochastic optimization, matrix factorizations too large to fit in memory, and ways to learn representations of human cognition by pooling many separate imaging studies. He presented on massive matrix factorization at ICML in 2016 and on multi-study cognitive representations at NeurIPS in 2017. He defended his thesis, Learning representations for functional fMRI data, on September 28, 2018.
The code from those years is still public. On his GitHub page the most popular repository is
modl, "Randomized online matrix factorization," written in Python, with 140 stars. Below it sit didyprog (Differentiable Dynamic Programming), cogspaces (multi-study decoding models for task fMRI data), online_sinkhorn, g-softmax, and a Jupyter notebook called deep-fmri. These are the working tools of a young researcher whose subject, underneath the neuroscience, was efficiency. How do you extract the most structure from the least computation? How do you look at a brain without storing all of it?He spent the next two years, 2018 to 2020, as a postdoctoral fellow in the applied mathematics department at the École normale supérieure on the rue d'Ulm, working on optimal transport and stochastic optimization. He also spent several months at NYU's Courant Institute with Joan Bruna on multi-agent reinforcement learning. Optimal transport is an old problem dressed in modern clothes. Gaspard Monge posed it in 1781: given a pile of earth and a hole of the same volume, what is the cheapest way to move one into the other? The modern version asks how to match one distribution to another at minimum cost. It is a mathematics of allocation, of getting supply to demand with nothing wasted.
His mother taught physics. He studied how to move mass efficiently. Much later he would describe his company's work as converting electrons into tokens. Nobody planned the rhyme.
The Giants Were Undertrained
In late 2020 Mensch joined DeepMind's Paris office. When he arrived, he later told the hosts of No Priors, the language model team "was 10 people at the time." It was a small group inside a large lab inside one of the largest companies on earth, and it was working on a technology whose commercial importance almost nobody outside the field yet understood.
He went to work on retrieval almost immediately. The resulting paper, published in December 2021, introduced RETRO, a model that improved its predictions by fetching passages from a database of several trillion words while it ran. Mensch was the second author. His closest collaborator was Sebastian Borgeaud, whom he called "my friend Seb Borgeaud, who is still at DeepMind." It is a small, warm phrase, and the qualifier at the end says more than it means to. Borgeaud was the friend who stayed. The idea behind RETRO was elegant and slightly subversive: "The point there was to use very large databases during pre-training so that we didn't force knowledge into the model itself," Mensch said. The model "would have access to an external memory anyway." Do not make the engine memorize the library. Give it a card.
He also worked, early on, on sparse mixture-of-experts architectures. These are models that activate only a fraction of their parameters for any given input, routing each token to a small committee of specialist subnetworks instead of the whole parliament. He came to it, he explained, through his postdoc. "Optimal transport is a setting where you have, I guess, tokens, you need to assign them to devices and you need to make sure that there's some good assignment in between the two of them so that the devices don't see too many tokens." It was Monge's problem again, with tokens in place of earth and GPUs in place of the hole. In 2022 he co-signed a study of scaling laws for these "routed" models.
Then came the paper that made his name in the field.
"Training Compute-Optimal Large Language Models" appeared in March 2022 and was later presented at NeurIPS. Inside the lab it was known by the name of the model it introduced: Chinchilla. Jordan Hoffmann, Sebastian Borgeaud, and Arthur Mensch were the first three of twenty-two authors. The team trained more than 400 language models, ranging from 70 million to more than 16 billion parameters, to answer one question: for a fixed budget of computation, how should you divide your money between making the model bigger and feeding it more text?
In 2019 and 2020 the reflexive answer had been to make it bigger. Parameter counts were how labs boasted, and size was what the debate was about. Chinchilla showed that the giants of the day had been, in a precise technical sense, undertrained. They were too large for the amount of data they had seen. At constant compute, model size and training data should grow together. A smaller model fed more data could match or beat a larger one starved of it. As Anjney Midha later summarized it on the a16z podcast, the paper demonstrated "that datasets were actually more important than just the sheer size of the model."
It is hard to overstate how much of what followed was already contained in that result. A company whose founding proposition was that it could compete with American labs on a fraction of their capital had, in effect, a mathematical proof that capital spent badly was capital wasted. Chinchilla did not say small beats big. It said waste loses. For a researcher whose doctorate had been about extracting signal from fMRI noise without drowning in it, this was less a discovery than a vindication.
The paper also carried a lesson he would only appreciate later. Being right about the economics of a technology is a different thing from owning the economics of that technology. Chinchilla's insight belonged to the whole field within months. Everyone retrained. Google owned the paper, and in a sense so did everyone who read it. Mensch, the third author, owned a line on his CV.
Waiting for the Hour
"I guess we were waiting for the hour," Mensch told Elad Gil in March 2024, "and the hour came with GPT to some extent."
The "we" was three people. Guillaume Lample and Timothée Lacroix had come out of the same French pipeline that produced Mensch. The three originally met during their studies at the École Polytechnique. Lample and Lacroix had gone to Meta, where they were among the researchers behind the original LLaMA models, Meta's first serious entry into large language models and the family that would soon define open-weight AI. Lample would become Mistral's chief scientist and Lacroix its chief technology officer. Contemporaneous accounts sketch them in complementary shorthand: Lample the large-scale language modeler, Lacroix the engineering depth, Mensch the optimizer. They were three Polytechnique men who had each gone to work for an American giant and each watched it move somewhere they had not expected.
"The various companies we were in," Mensch told No Priors, "move into directions that we hadn't anticipated when we joined the company." He did not spell out the directions, and he did not need to. By early 2023 the frontier of AI was consolidating behind APIs. Models were becoming services, weights were becoming secrets, and the field that had once published everything was beginning to publish press releases. ChatGPT had arrived in late November 2022 and turned a research curiosity into a consumer phenomenon. It also turned a research community into a set of competing corporate fortresses.
The three of them had been doing machine learning for about ten years. "We did know where to start from and how to make a good model with a limited amount of compute and money," Mensch said. Then, with the precision of someone who distrusts his own rhetoric: "Well, not so limited, but at least more limited than where we were coming from."
The plan, as he told Gil, was to "go and try to do speedrun a bit because we weren't starting the first." A speedrun is a gamer's term: finishing a known game as fast as possible, exploiting every shortcut a first-time player would miss. You can only speedrun a level you have already played. Mensch had co-written the field's best map of how compute should be spent. Lample and Lacroix had built LLaMA. They were not trying to invent the game. They were trying to finish it faster.
The company was named for the mistral, the cold, violent wind that comes down the Rhône valley from the north and blows into the Mediterranean. The word comes from Occitan, the old language of the south, a language the French state spent centuries suppressing in favor of the Parisian standard. That a Parisian company of Polytechnique graduates would take its name from the vocabulary of France's internal periphery is the kind of detail nobody remarks on. The company's own gloss was speed and scale.
The fundraise came almost immediately. It was the first sign that this would not be a typical European startup. On June 13, 2023, the Financial Times ran a headline that made the European venture world do a double take: "Four-week-old AI start-up raises record €105mn in European push." The investors included Lightspeed Venture Partners, Eric Schmidt, Xavier Niel, and JCDecaux. The FT estimated the valuation at €240 million. Rodolphe Saadé, the French shipping magnate, was among the backers. According to later accounts, the pitch rested on a seven-page strategic memo arguing that foundation models would hold the primary value in the AI ecosystem and that a European provider could claim a leading position among them.
Arjun Kharpal, a CNBC technology correspondent, happened to be in Paris that month. Three years later he still remembered how strange it was. "No product. No marketing. And not much info about who they were." The company's own website records its first employee as joining on June 5, 2023, about a week before the round was reported. They were raising a European record on the strength of a memo, three CVs, and a speedrun.
From the beginning, the company had what Politico would describe as "strong ties with the French political establishment, boasting a former junior digital affairs minister as its co-founder." In September 2023, Mensch joined the generative AI expert committee established by Prime Minister Élisabeth Borne. The wind from the periphery had found its way to Matignon within five months.
A Magnet Link and No Guardrails
The first model did not arrive with a keynote.
On September 27, 2023, Mistral AI posted a BitTorrent magnet link. A magnet link is a string of characters that tells a peer-to-peer client how to find a file distributed among strangers' computers. The file was Mistral 7B, a language model with 7.3 billion parameters, released under the Apache 2.0 license. In practice that meant anyone could download it, run it, modify it, and sell products built on it without asking permission or paying a cent. Mistral released both a base model and an "instruct" version tuned to follow chat-style prompts. The company noted that the fine-tuned model was intended for demonstration only and did not have guardrails or moderation built in. A tidier blog post and a Hugging Face upload followed.
Mistral's claims for it were pointed. Mistral 7B, the company said, outperformed Meta's LLaMA 2 13B on every benchmark it tested and was on par with LLaMA 34B on many. A model half the size of one competitor and a fifth the size of another was beating them. This was Chinchilla carried out as product strategy. Later, in a long interview with the French YouTube channel Underscore_, Mensch would list the ingredients plainly. They knew the importance of high-quality data. They knew how to train efficiently. They controlled their compute costs. And they knew exactly which open-source models they had to beat.
The torrent was a provocation, and it was understood as one. Every other frontier release that year came with a product page, a safety card, and a carefully staged rollout. Mistral offered a string of hexadecimal characters and an invitation to the internet. In December it did it again. On December 9, 2023, Mixtral 8x7B appeared via another torrent link posted on Twitter, two days before the blog post explaining it. Mixtral was a sparse mixture-of-experts model, the architecture Mensch had approached at DeepMind through the mathematics of optimal transport. Eight expert subnetworks were available and only a fraction were active per token. Mistral claimed it beat both LLaMA 70B and OpenAI's GPT-3.5 on most benchmarks. Gizmodo's headline caught the tone: "Buzzy Startup Just Dumps AI Model That Beats GPT-3.5 Into a Torrent Link."
The day after the Mixtral link went out, on December 10, 2023, Mistral announced a second round: €385 million, led by Andreessen Horowitz, with BNP Paribas and Salesforce participating, at a valuation above €2 billion. The company was seven months old. Anjney Midha, the a16z partner who led the investment and later joined Mistral's board, belongs to a particular modern type: the venture partner as collector of model labs. By his firm's own listing he would eventually sit on the boards of Black Forest Labs, Luma AI, Mistral AI, Sesame AI, LMArena, OpenRouter, and Periodic Labs. That is a portfolio that amounts to a thesis about the open, fragmented future of AI. On the podcast Midha recorded with Mensch that December, the cold open captured the founder's argument: "I think the battle is for the neutrality of the technology." And then, on why open release was not a radical choice at all: "If you look at the history of software, the only way we did software collaboratively is really open source, so why change the recipe?"
Weeks earlier, there had been a photograph. In November 2023 the British government convened the world's first AI Safety Summit at Bletchley Park, the country house where Alan Turing broke German codes during the Second World War. The official photograph shows US Vice President Kamala Harris, European Commission President Ursula von der Leyen, Prime Minister Rishi Sunak, Turing laureate Yoshua Bengio, OpenAI's Sam Altman, and DeepMind co-founder Demis Hassabis. On the top riser, second from the left, stands a young man from a company that was then about six months old. In the MIT Technology Review write-up that later named him among its Innovators Under 35, the caption is simply: Arthur Mensch standing on the top riser, second from the left. He had been at DeepMind eight months earlier. Now he was in the same photograph as its co-founder, standing a little higher.
The open releases had a shadow, and it was not long in coming. In March 2024 Patronus AI tested leading models on a hundred prompts designed to elicit text from books under U.S. copyright. OpenAI's GPT-4 reproduced copyrighted text verbatim in 44 percent of responses, Mixtral in 22 percent, LLaMA 2 in 10 percent, and Anthropic's Claude 2 in 8 percent. Mistral was not the worst offender. It was not clean either. Releasing weights to the world means releasing everything the weights contain, and guardrails left out of a demonstration model stay out of every copy a stranger downloads.
An Engine Devoid of Usage
The European Union was writing its AI Act through the second half of 2023. Late in the drafting, the question of foundation models became the hinge of the whole negotiation. Should the rules attach to the underlying models, the engines, or only to the applications built on them?
Mistral lobbied hard for the second answer, and France was one of the leading voices in Brussels calling for what Politico later called "innovation-friendly" regulations. Mensch's line, repeated in various forms, was compact: "We cannot regulate an engine devoid of usage." Regulate the car, not the combustion. Regulate the use, not the mathematics. It was a principled argument, and it was also, transparently, an argument that would let a company whose business depended on distributing base models keep distributing them. He has never pretended otherwise. In interviews he consistently says regulation should target applications, not base models. He has called the focus on extinction risk a route to regulatory capture, by which he means a way for the incumbents who could afford compliance to pull up the ladder behind them.
It worked, more or less. The final AI Act treated general-purpose models more lightly than early drafts had, and von der Leyen made "AI Innovation" a priority of her next mandate. Politico drew the connection explicitly: "Mensch's influence in Brussels was evident in the negotiations for the AI Act." The young man on the top riser at Bletchley had learned how to stand in a room with ministers and walk out having moved the text.
Then, on February 26, 2024, he did something that cost him most of the goodwill he had just banked.
That morning Mistral released three things at once. Mistral Large was its most capable model to date and, unlike everything before it, proprietary: no torrent and no weights. Mistral Small came out alongside it. Le Chat, the company's consumer chatbot, launched the same day; the name means "the cat." And it announced a partnership with Microsoft. Mistral's models, including Mistral Large, would be available on Microsoft's Azure cloud, and Microsoft would invest $16 million.
Sixteen million dollars was trivial against the €385 million Mistral had raised ten weeks earlier, and the symbolism was anything but. Microsoft was OpenAI's principal backer. The company built in the name of European independence from American platforms had just put its flagship, closed model on the most important American platform of all. French lawmakers, Politico reported, were angered. Mensch's defense was the one he would give for years: without the partnership, Mistral would struggle to access the computational resources it needed to power its made-in-Europe technology. The engine needed electricity, and in February 2024 the electricity was in Redmond.
Weeks later, under a congratulatory LinkedIn post by Jean-Philippe Vert, a French machine learning researcher who had co-written papers with Mensch "not so long ago," a stranger left a five-word comment. "More like a made at Google story if you ask me."
Whoever wrote it had put his finger on something the French establishment preferred not to say aloud. Politico said it eventually, in its 2025 list of the most powerful people in Europe: Mensch's critics saw him as "something of a prodigal son that they don't fully trust." He was Polytechnique and Paris-Saclay, yes, "but he got his professional chops working for Google, a company that is decidedly not French." The question looming over him, Politico wrote, was whether he would prove "a truly European champion" or be "dismissed as a French cog in a much grander 'made-in-the-USA' machine."
The Prodigal Son
Three months after the Microsoft announcement, and a couple of weeks after Mistral's first birthday, Mensch appeared before the French Senate's Economic Affairs Commission. For ninety minutes he answered questions from senators while the commission's president, Dominique Estrosi Sassone, moderated. She referred to the company's first two rounds as "figures that are dizzying, to say the least." Public Sénat built its headline on a line he gave them, in the plainest French possible: "On est Français et on va le rester." We are French and we will stay French.
The rest of the testimony was more interesting than the headline, because it was more honest about the bind. Mistral had needed massive early funding, Mensch told the senators, and the structure of French venture capital had forced it to look to the United States for that money. French labor rules made it hard to recruit top talent quickly, which was pushing the company to expand in America. He asked for more public research funding, a more generous crédit d'impôt recherche, and more flexibility in labor law. His argument was not that Europe should wall itself off from America. It was that Europe should stop making it necessary for its best companies to become partly American.
He also laid out, more clearly than he had anywhere before, what the open releases had been for.
— Arthur Mensch, testimony before the French Senate Economic Affairs Commission, May 2024The reason we launched Mistral was to say that we knew how to make this technology in France, that we could recruit a talented team very quickly, and then, rapidly propose foundational models... We put what was about to be a particularly oligopolistic field back on a path where there was an opportunity perhaps for a little more decentralization, thanks to the model we were proposing.
The models had been given away "free of charge, so that developers, who are the first people we're talking to, the people who make applications, can take hold of this technology, deploy it wherever they want — not necessarily in the three American public clouds — and modify it as they see fit." The cloud providers, he said, were "both competitors and partners." He described the alternative bluntly: the natural course of things would be "to leave the keys to this technology to the four companies that have captured most of the value of tech in the last 20 years and which, thanks to their dominant position, can largely capture most of the added value of the new industrial revolution that's coming in the next 20 years."
Under this logic, the Microsoft deal was not a betrayal. It was one channel among several, and the point of having several was that no single one could hold you hostage. Openness and sovereignty, which looked like opposing values, were for Mensch the same strategy at different scales. Release the weights, and no single cloud can capture the developers who use them. Put your models on every cloud, and no single cloud can capture you. Two years later, before the National Assembly, he would compress the idea into a sentence: don't think of sovereignty as isolationism, think of it as leverage.
The money kept arriving at a pace that made the senators' "dizzying" sound quaint. In June 2024, General Catalyst led a €600 million Series B at a valuation of €5.8 billion. By valuation that made Mistral the fourth-largest AI company in the world and the largest outside the San Francisco Bay Area. MIT Technology Review noted the company had just 65 employees. Time put Mensch on its TIME100 AI list in September 2024, the only Frenchman included. In May 2025 the Republic made him a Knight of the National Order of Merit. Later that year Bloomberg reported that the three founders, worth roughly $1.1 billion each, had become the first AI billionaires in France.
In June 2025 he stood on a stage in Paris with Emmanuel Macron and Jensen Huang, the founder and CEO of NVIDIA, the man who sells the picks and shovels to every side of the AI gold rush. Maurice Lévy was there too, and so were the launch partners for Mistral's new compute effort: Black Forest Labs, BNP Paribas, Kyutai, Mirakl, Orange, Schneider Electric, SLB, SNCF, Thales, Veolia. Mensch posted a photograph with a caption that, for him, came close to exultation: "Today, we are paving the way for Europe's leadership in AI, and it all starts from Paris!" The most-liked comment beneath it, with fifty-eight reactions, came from a stranger who had read the caption and gone straight to the payroll. "Very curious to see how you can attract very good AI engineers given the taxation and the average wage in France."
Je ne savais pas vendre
Inside the company, the early months looked less like a national mission than a well-funded graduate lab running short on graphics cards.
In April 2024, Mensch told Harry Stebbings on 20VC that Mistral was working with roughly 1,500 Nvidia A100 GPUs, a small fraction of what its competitors had. Delays from compute providers had hurt. Raising $2 billion as a seed round had simply not been possible in 2023, so the scale of everything was capped. The team had started at around twenty-five people. Its operating philosophy, as he summarized it, was that a small team could outrun a large one if properly decoupled: groups that share infrastructure and code but work independently. It was a lesson he said he had taken from DeepMind. Feedback inside the company was "almost fully transparent." There had been a mistake, and he named it. They had launched sales before the product was mature, and that had caused friction inside the organization.
Sales was the thing he kept coming back to, in the tone of a man confessing a weakness he had decided not to be ashamed of. On January 14, 2026, he returned to the École Polytechnique at the invitation of the student entrepreneurship club, the Binet start-up. He structured his talk around the school's motto — Patrie, Sciences, Gloire — and told the students something that founder mythology rarely admits. When he started Mistral, he said, the only thing he knew how to do was code and train models. The romantic picture of entrepreneurship, sitting in a café writing code, was wrong. Going fast and far meant working with people who gave you leverage to do something else. The job of a founder, whether CEO or CTO, very quickly stopped being technical. It was organizing, managing outside relationships, and selling.
— Arthur Mensch, speaking to students at the École Polytechnique, January 2026Je ne savais pas vendre et je pense que je ne sais toujours pas vendre.
I didn't know how to sell, and I think I still don't. He had known great salespeople and hired great salespeople. Selling looked like a thankless task, he said, but it was the one that gave the most leverage, because it created revenue and revenue was what let you grow. At Mistral, by his account, the CTO and the chief scientist had become sellers too. Their job was to understand what the customer wanted and fold it into the implementation.
The company that emerged from this discovery was not the one the torrents had announced. The open models kept coming: Mixtral 8x22B in April 2024 (another torrent, posted on April 10), Mathstral, Codestral Mamba, Pixtral, and the Ministral and Magistral families. But the licenses grew more complicated. Codestral, the first code model, launched on May 29, 2024, under a "Mistral Non-Production License" that forbade commercial use. Mistral Large 2, released July 24, 2024, came with open weights under a research license, with commercial use available for a fee. Mistral Medium 3, in May 2025, was proprietary and sold as an enterprise model for on-premise deployment. The June 2025 Magistral reasoning release, which Reuters described as Europe's first AI reasoning model, came in pairs: Magistral Small open under Apache 2.0, Magistral Medium closed. The open base came free. The best version, adapted to your data, behind your firewall, and serviced by Mistral engineers who would sometimes sit inside your company, came with an invoice.
The customer list filled in with a very French ruling-class geography. In January 2025, Agence France-Presse signed a deal letting Le Chat query its entire archive going back to 1983 and feeding more than 2,000 articles a day into the assistant. On February 6, 2025, Le Chat reached iOS and Android with a Pro tier at $14.99 a month. In April 2025, the shipping group CMA CGM signed a €100 million partnership focused on customer service. HSBC began equipping its 200,000 employees with Mistral's assistants and agents. By the time of his National Assembly testimony, the clients Mensch named under oath included the state's digital directorate (DINUM), the Caisse des dépôts, France Travail, the Ministry of the Armed Forces, Stellantis, TotalEnergies, BNP Paribas, and the Grand Duchy of Luxembourg. He said about 30 percent of revenue came from France, 70 percent from outside it, and three-quarters from Europe. In February 2026 came a partnership with Accenture. Later, the company would report more than 125 enterprise customers using its products for mission-critical work.
In October 2025 Mensch wrote, on LinkedIn, the most corporate sentence of his public life. It was also one of the most revealing: "Getting AI systems that work in the enterprise requires focus and a scientific mindset." The product being announced, Mistral AI Studio, existed to move "prototype use-cases to production use-cases delivering ROI." You tackle progressively, he wrote, "the edge cases that any AI software is bound to run into at first." The doctoral student who had spent three years wringing signal from noisy brain scans was now wringing return on investment from noisy enterprise processes. It was the same problem with a different budget.
In May 2026, the cat was renamed. Le Chat became Vibe, an agent for long-horizon work that combined the chatbot with Mistral's coding tools. The name had been a pun; the new one is a mood. Nobody mourned the cat in public.
Electrons Into Tokens
In September 2025 the most important technology company in Europe bought a piece of Mistral.
ASML, the Dutch maker of the lithography machines without which no advanced semiconductor on earth can be manufactured, invested €1.3 billion for roughly 11 percent of the company and became its largest shareholder. It anchored a €1.7 billion Series C at an €11.7 billion post-money valuation. Politico's headline caught the mixture of hope and doubt that would hang over the deal: "ASML-Mistral is Europe's dream tech tie-up. Can it deliver?" It was a strange marriage on paper: a model lab and a maker of extreme-ultraviolet machines. In practice it marked the moment Mistral stopped describing itself as a software company and started describing itself as a link in an industrial chain that ran from photons hitting silicon to tokens hitting a screen.
The chain filled in quickly. On December 2, 2025, Mistral released Mistral Large 3, a sparse mixture-of-experts model with 675 billion total parameters and 41 billion active per token. It came out under Apache 2.0, fully open, the largest Mistral had ever given away. The architecture Mensch had studied as a problem of optimal transport at DeepMind was now the backbone of the company's flagship. In February 2026 Mistral made its first acquisition, Koyeb, a Paris startup whose serverless infrastructure TechCrunch said would "back its cloud ambitions." In March 2026 it raised $830 million in debt to build data centers near Paris and in Sweden, and announced a partnership with NVIDIA on open frontier models. In May it acquired Emmi AI, an Austrian company building simulation models for industrial engineering, which promised to cut industrial simulations from hours to seconds. In June, on CNBC's Tech Download, recorded in Paris with the same Arjun Kharpal who had puzzled over the four-week-old company three years earlier, Mensch disclosed for the first time that Mistral was exploring designing its own chips.
— Arthur Mensch, before the National Assembly commission of inquiry, May 12, 2026Une bulle, c'est quand la demande est surestimée. Or le problème aujourd'hui est plutôt celui de l'offre.
A bubble is when demand is overestimated. The problem today is rather one of supply. The scarcity he described under oath was material: chips, memory, motherboards, the helium semiconductor fabs consume, electrons. "Today the demand is actually way above the supply," he told CNBC a month later. In the United States the binding constraint was the grid itself. "The United States is saturating its electricity grid to convert it into artificial intelligence," he told the deputies. The sentence sounded like a complaint and was really a business plan. France has nuclear power and a surplus. France has electrons. The company that converts them first owns the conversion.
He had been rehearsing the metaphor for months. "L'intelligence artificielle, c'est un peu un travail d'électricien," he told the Polytechnique students in January. Artificial intelligence is a bit like electrician's work. Our role is to transform megawatts into intelligence. By May it had become doctrine. Digital services, the cloud, and AI were "vraiment la même chose," really the same thing. Traditional hosting was becoming a low-margin commodity, he argued, and "you can't build a scalable business starting from a position where margins are low." Along the chain from electron to token, he estimated, about 10 percent of the value sat in the electron and about 90 percent everywhere else. Europe was exporting the 10 and importing the 90.
He was as blunt about Europe's habits as about America's appetite. Twenty-seven unsynchronized national regimes, GDPR, the AI Act, and copyright law stacked on top of one another added up, in his phrase, to a system in which "regulation favors the big players." Mistral, a thousand people strong, had five employees in compliance. Any claim that regulation would defend Europe, he said, was the road to hell paved with good intentions. The American narrative that dismissed European rules, and that Europeans had internalized, was "a form of colonialism." Public procurement accounted for half of EU GDP, and "the United States and China have used it massively since the 1940s — we need to stop being afraid to use it." The man who had lobbied Brussels to keep its hands off base models was now asking European states to put their thumbs on the scale.
On defense he drew a line that put him, implicitly, at odds with at least one American rival. Mistral works with the French Ministry of the Armed Forces and with France's allies, he said, but "we don't claim to have the democratic legitimacy to explain to the French armed forces what they can do." Mistral would advise on reliability. It would not veto use. On cyber, he dismissed what he called the "fear marketing" of an American competitor. Offensive capabilities in models were rising, he said, "in a linear, predictable way, for everyone at the same time."
The money in 2026 made the earlier rounds look like rehearsals. On July 21, 2026, Microsoft, whose $16 million had caused the 2024 uproar, signed a multibillion-dollar agreement under which it would use part of Mistral's expanded European computing capacity. The models went into Microsoft Foundry and Copilot Studio. Two years earlier, Mistral had needed Microsoft's compute. Now Microsoft was buying Mistral's. In September 2026, Samsung Electronics led a €3 billion Series D, with EQT's Scaleup Europe Fund and PSG Equity as co-leads, at a post-money valuation above €21 billion. TechCrunch called it "the largest equity fundraising round ever completed by a European technology company." The partnership sent Mistral's models into Samsung's semiconductor operations. "AI is reshaping how we build complex technologies, from silicon to software," Mensch said in the release. A company that had begun as three researchers with a magnet link was now helping design the chips its models would run on, backed by the firm that makes the machines that make the chips.
The company had set itself a target of one gigawatt of capacity by 2029. Under oath, Mensch called it "not enough." He also said that being bought would be a failure. The aim, according to accounts of the hearing, was a public listing, not an exit.
The Commoditized Engineer
At Polytechnique in January 2026 he told the students something that sat uneasily with the school's motto. "La génération d'ingénieurs qui est la mienne a à peu près réussi à commoditiser son propre métier." My generation of engineers has more or less succeeded in commoditizing its own profession. What they had been trained for — synthesis, calculation, reasoning, software development — was exactly what they had spent the past decade teaching machines to do better and better. The technology now demanded that people climb one more level of abstraction. AI, he said, promoted a new kind of brain, a far more creative one.
Then he added the condition that made the talk memorable. You had to learn to use AI, and you also had to know how to do entirely without it. To learn, to understand, to develop the capacity to reason, work without a computer remained, in his telling, indispensable. In the interviews he gives most often, he calls deskilling the biggest risk of the technology he sells. Business Insider's headline on one of them reduced it to "AI's biggest threat is people getting lazy." The CEO of one of the world's leading model labs was telling the next generation of French engineers to close their laptops.
Four months later, under oath, he described what had happened inside his own company. Engineering productivity had doubled in six months. "Les ingénieurs chez Mistral n'écrivent plus de lignes de code." Mistral engineers no longer write lines of code. "Vous n'êtes plus un artisan, vous êtes un manager." You are no longer a craftsman; you are a manager. Accounts of the hearing record him describing a solo developer working with agents as ten to twenty times faster, a gain that shrank in a team of five and collapsed inside a large organization, where coordination ate it. He described the economic shift underneath all of it as something that "moves intelligence from labor toward capital and toward machines." Asked about unemployment, he did not rule out a rise.
He has never shared the mysticism of some of his American counterparts. "I don't believe in AGI," he has said. On the Big Technology podcast in January 2026 he put the alternative in terms an engineer would accept and a prophet would not: "There's no such thing as one system that is going to be solving all the problems of the world. You don't have any human able to solve every task in the world." He has compared a model to an intern that improves with feedback. He has told students that the frontier is no longer text but simulators, materials, robots, biology, and that "if you want to push the capabilities of a model in materials science or quantum physics, you will have to get up early, go get data." His personal investments follow the same line: Braintrust for evaluation, Linkup for connecting agents to the web, Parallel for hospital administration, Biorce for clinical trials, Lemni for agent production, Mirelo for generated sound. These are the plumbing of AI after the model, the unglamorous layers where the edge cases live.
The man who wrote that he could not sell has, by every external measure, sold more of a European technology company than anyone before him. The man who made his name proving that models had been fed too little data now sells the infrastructure to feed them more. He gave away the engines in 2023 to keep any single company from owning them. In 2026 he was asking his own state to buy from him on purpose. He is a researcher whose engineers no longer write code, warning a generation of engineers not to stop writing it.
Before the deputies he had said that the question was not so much whether Europe could compete as how. He would not say what he was. Patriot, prodigal, champion, cog: he let others choose. On the subject of himself he has always been the least forthcoming source in the room.
His GitHub page is still there. Paris, France. A personal domain, amens.ch. Five hundred ninety-four followers; he follows three. At the top, the repositories from the Saclay years: randomized online matrix factorization, 140 stars; differentiable dynamic programming; multi-study decoding models for task fMRI. Nothing about Mistral. Beneath the follower count, in the line where a profile names where its owner works, he has typed three letters.
TBA.
Part IIThe Playbook
What follows are twelve principles drawn from Mensch's first three years as a founder. They do not describe a temperament. They describe a set of choices about where to spend scarce resources (compute, capital, credibility, attention) when your competitors have more of all four. Several of them pull against each other. That tension is the playbook. Mensch has spent three years running a company built on openness and sovereignty, on research and sales, on European independence and American capital, and he has held these opposites together by treating each as a lever, not an identity.
The principles
- 1.Find where the giants are overspending.
- 2.Speedrun the level you've already played.
- 3.Raise for the ambition, not for the ecosystem.
- 4.Make the release the marketing.
- 5.Open the base, sell the edge.
- 6.Treat every giant as a channel, never as a home.
- 7.Define sovereignty as leverage, not isolation.
- 8.Choose where the rules bite.
- 9.Hire the skill you lack, then make everyone use it.
- 10.Keep teams small and decoupled.
- 11.Move down the stack toward the scarce input.
- 12.Never lose the ability to work without the tool.
M
The Capital Stack, 2023–2026
Each round changed both the size of the company and the nationality of its dependencies.
Apr 2023
Mistral AI founded on April 28 by Mensch, Lample, and Lacroix; about fifteen people in France and England.
Jun 2023
€105M seed (Lightspeed, Eric Schmidt, Xavier Niel, JCDecaux) at roughly €240M valuation. No product.
Dec 2023
€385M led by Andreessen Horowitz, with BNP Paribas and Salesforce; valuation above €2B.
Feb 2024
Microsoft invests $16M; Mistral Large goes to Azure. French lawmakers object.
Jun 2024
€600M Series B led by General Catalyst at €5.8B. Largest AI company outside the Bay Area by valuation.
Sep 2025
ASML invests €1.3B for ~11%, anchoring a €1.7B Series C at €11.7B.
Mar 2026
$830M in debt to build data centers near Paris and in Sweden.
Jul 2026
Multibillion-dollar Microsoft agreement; Microsoft uses Mistral's European compute.
Sep 2026
€3B Series D led by Samsung Electronics, with EQT and PSG Equity; valuation above €21B.
Principle 1
Find where the giants are overspending
Chinchilla is usually read as a paper about scaling laws. It is better read as a paper about waste. Its central finding was that the largest language models of 2020 and 2021 had misallocated their compute: too many parameters, too little data. The giants had not been wrong to spend. They had been wrong about where to spend. That distinction is the whole basis of Mistral's existence. A company with roughly 1,500 A100s in early 2024 could not out-spend Google or OpenAI. It could out-allocate them.
Mistral 7B made the argument concrete. A 7.3-billion-parameter model that, by Mistral's benchmarks, beat a 13-billion-parameter competitor everywhere and matched a 34-billion one on much of the field was a commercial demonstration of a research result. Later the mixture-of-experts architecture extended the logic: Mistral Large 3 holds 675 billion parameters but activates only 41 billion per token. The company calls these "compute multipliers," gains drawn from algorithms and data quality instead of raw spend.
The point generalizes beyond AI. Incumbents with abundant resources develop habits that only make sense when resources are abundant. Their spending becomes a signal of seriousness as much as a means of production. A challenger who can find the habit that has become a ritual can compete in the gap between what the incumbent spends and what the result actually requires.
Tactic: Before competing with a better-funded rival, identify one specific input it is overspending on out of habit, and build your entire cost structure around not making that mistake.
Principle 2
Speedrun the level you've already played
"We weren't starting the first," Mensch told Elad Gil. The admission is the strategy. Mistral's founders did not need to discover that large language models worked, or how to train them, or which open models they had to beat. They had written the scaling paper and built LLaMA. What they lacked was not knowledge but a vehicle, and the vehicle was the only thing they had to build from scratch.
Hence the pace. The company was founded on April 28, 2023, raised a European record seed round within weeks, shipped a competitive open model five months after founding, and shipped a mixture-of-experts model that beat GPT-3.5 on most benchmarks seven months in. That is not the timeline of invention. It is the timeline of execution against a known map. The speedrun metaphor implies two disciplines: knowing the route cold, and refusing every detour. In its first year Mistral built text models and nothing else.
The broader lesson concerns timing. The founders say they had been "waiting for the hour." They had the expertise for years and lacked the moment. ChatGPT supplied the moment, by converting a technical consensus into a market. The right time to leave an incumbent is not when you know enough. It is when the world suddenly understands what you already know.
Tactic: If you are leaving a large organization to compete with it, start only once the market has validated the category, then use your inside knowledge to skip every step that a first mover had to discover.
Principle 3
Raise for the ambition, not for the ecosystem
The €105 million seed was a scandal by European standards, which was precisely its value. A normal French seed round would have bought a normal French startup: a few dozen people, a modest cluster, a polite trajectory toward acquisition. Mensch told the Senate plainly that the structure of French venture capital could not provide what a foundation-model company needed, so he went to American funds. He did not apologize for this. He testified about it, asking the state to fix the conditions that had forced his hand.
The financing pattern since then shows the same refusal to size ambition to local expectation. Each round has been calibrated to the next technical threshold: models, then compute, then data centers, then chips. Each has been followed by a public argument that Europe should make the next one easier to raise at home. Accounts of the 2026 testimony record that less than 30 percent of Mistral's capital is American, and that this reflects the absence of a European alternative, not a preference.
The principle is not "raise as much as possible." It is that the capital required by the problem is an objective quantity, and a founder who scales the round to the local market's comfort is choosing the market's comfort over the problem.
Tactic: Calculate the capital your technical goal actually requires before you talk to any investor, and if your home market cannot supply it, raise abroad and spend some of your credibility lobbying to change that.
Principle 4
Make the release the marketing
The magnet link of September 27, 2023, was the most efficient piece of marketing in the AI industry that year. It cost nothing to post. It signaled confidence, since only a team sure of its model would hand it to strangers without a launch event. It signaled ideology, since Mistral was obviously the open lab. And it put the product directly into the hands of the audience Mensch most wanted, the developers who, as he told the Senate, "are the first people we're talking to." Trust followed adoption. On 20VC he described open releases as a way to build community vouching, and brand trust as essential for enterprise adoption.
Notice what the torrent did not do. It did not require Mistral to build distribution, a consumer interface, a sales team, or a cloud. Thousands of strangers became the distribution. Hugging Face became the storefront. The developers who fine-tuned and deployed Mistral 7B on their own hardware became an unpaid evangelism force, and every derivative model carried the name.
The risk is real, and the Patronus copyright study shows it. A model released without guardrails takes all of its contents into the world, including the parts that embarrass you. Mensch accepted that trade, and has argued since that the transparency of open models is itself a guarantee of safety.
Tactic: If your product can be given away in a form that demonstrates its quality on first use, let the act of releasing it be your launch, and spend the money you saved on what comes after adoption.
Principle 5
Open the base, sell the edge
Read across Mistral's licensing history and you see something less ideological than its reputation: a company constantly re-drawing the line between what it gives away and what it charges for.
L
Mistral's Licensing Line, Redrawn
The open/closed boundary has moved with nearly every major release.
| Model | Released | License | Posture |
|---|---|---|---|
| Mistral 7B | Sep 2023 | Apache 2.0 | Open |
| Mixtral 8x7B | Dec 2023 | Apache 2.0 | Open |
| Mistral Large | Feb 2024 | Proprietary | Closed |
| Codestral 22B | May 2024 | Mistral Non-Production License | Restricted |
| Mistral Large 2 | Jul 2024 | Mistral Research License | Restricted |
| Mistral Medium 3 | May 2025 | Proprietary (on-premise) | Closed |
| Magistral Small / Medium | Jun 2025 | Apache 2.0 / Proprietary | Split |
| Mistral Large 3 | Dec 2025 | Apache 2.0 | Open |
The pattern resolves into a hybrid. Open models build adoption, trust, and an ecosystem. Proprietary models, enterprise deployment, post-training on a customer's own data, and Mistral engineers embedded in client organizations generate revenue. Mensch's 2025 announcement of AI Studio made the commercial logic explicit. What the enterprise pays for is not the base model but the loop that adapts it to "proprietary knowledge and data, and then the signals produced during AI-human interactions." On 20VC, Mensch predicted that value would accrue at the application layer, where specific data and user feedback create differentiation.
The critics are right that "open weights" is not the same as open source. The line between them is where the business lives.
Tactic: Give away the part of your product whose value grows with the number of users, and charge for the part whose value grows with the specificity of one customer's problem.
Principle 6
Treat every giant as a channel, never as a home
The February 2024 Microsoft deal looked to French lawmakers like capitulation. Two and a half years later it looks more like the first move in a longer game. Mensch's explanation to the Senate was structural: Mistral had made its models available through multiple cloud providers precisely so that no single one could capture it. The clouds were "both competitors and partners." In July 2026 the relationship partly inverted, as Microsoft agreed to use part of Mistral's European computing capacity.
The principle depends on multiplicity. One giant partner is dependence. Several giant partners — Microsoft, NVIDIA, ASML, Samsung, Accenture — are a portfolio, each limiting the leverage of the others. Mistral's multi-cloud strategy, described in its earliest positioning, was designed to reduce dependency on Microsoft or anyone else. The deals with giants are channels to reach customers, compute, or capital. None of them is where the company lives.
The cost is reputational, and it is real. Politico's "French cog in a much grander 'made-in-the-USA' machine" is the kind of phrase that can stick for years. Mensch absorbed it by continuing to argue, publicly and under oath, for the European conditions that would make such deals less necessary.
Tactic: Never sign an exclusive dependency with a dominant platform; sign with several, publicly, so that each one knows you have alternatives.
Principle 7
Define sovereignty as leverage, not isolation
The most intellectually useful move Mensch made in his 2026 testimony was to redefine a word that had become a slogan. "Don't think of sovereignty as isolationism but as leverage." Asked whether Europe could compete in AI without depending on non-European actors, he answered without hedging: "obviously no." The real question was whether it could compete while gaining leverage over its suppliers, as an exporter of technology and as a negotiator with something to trade. To that he said yes.
The distinction dissolves what looked like Mistral's central hypocrisy. A sovereign European AI company built on Nvidia chips, financed partly by American and Korean capital, and distributed on Azure is not a contradiction if sovereignty means having enough of the stack that no one supplier can dictate terms. An entity that imports 100 percent of its digital infrastructure has no bargaining position. One that owns the models, part of the compute, and some of the customer relationships does.
This is a framework any company can use when it depends on a more powerful partner. The goal is not independence, which is usually impossible. The goal is to own enough of the value chain that the partner needs you as much as you need it.
Tactic: List every supplier you could not replace within a year, and for each one build or acquire one capability that makes them need you back.
Principle 8
Choose where the rules bite
Mensch's regulatory record looks inconsistent until you notice the pattern. In 2023 he argued against regulating base models: "We cannot regulate an engine devoid of usage." In 2024 he asked the Senate for more labor flexibility and a bigger research tax credit. In 2026 he told deputies that stacked European regulation "favors the big players," and in the same hearing demanded that Europe use public procurement, half of EU GDP, as an industrial weapon.
The common thread is location. Mensch has never argued for an absence of rules. He argues for putting them where they harm challengers least and help them most. Rules on applications leave the model layer competitive. Rules on procurement direct demand toward domestic suppliers. Rules that require five compliance staff to absorb are trivial for a hyperscaler and heavy for a thousand-person company. His defense posture — advising on reliability without claiming the democratic legitimacy to veto use — places responsibility with the elected state, not with the vendor.
This is not cynicism. It is an accurate understanding that regulatory design is competitive design. Every rule shifts advantage, so a founder who treats regulation as weather instead of terrain has already ceded the ground.
Tactic: For every regulation that affects your market, work out which size of company it advantages, and spend your policy energy moving the burden onto uses instead of capabilities.
Principle 9
Hire the skill you lack, then make everyone use it
"Je ne savais pas vendre et je pense que je ne sais toujours pas vendre." This is unusual candor from the CEO of a company valued above €21 billion. What matters is what he did with the admission. He hired great salespeople. Then he made selling part of the technical leadership's job: the CTO and the chief scientist, by his account, became sellers too, responsible for understanding what customers wanted and folding it into the implementation.
He also named the mistake that preceded the fix. Launching sales motions before the product was mature created friction. The lesson he gave students was that product and go-to-market should be staged. The scientists should be exposed directly to user problems, and the salespeople should understand the technical nuances. At the Polytechnique talk he called selling the task that gives the most leverage, because it creates revenue and revenue creates growth.
The general point concerns technical founders specifically. The skill you were trained in is rarely the skill that constrains your company after the first year. Admitting the gap is only useful if it changes who you hire and how you redistribute responsibility.
Tactic: Identify the one non-technical skill your company's growth depends on, hire someone better at it than you will ever be, and then require your technical leaders to spend real time in front of customers.
Principle 10
Keep teams small and decoupled
Mistral began with roughly twenty-five people and an operating philosophy imported from DeepMind: organize into sufficiently uncoupled teams that share infrastructure and code but work independently. Mensch summarized it on 20VC: a small, properly decoupled team moves faster than a large one. It was 65 people when it was valued as the fourth-largest AI company in the world. It had 1,000 when Mensch testified in 2026.
The 2026 testimony added a twist that makes the principle sharper. AI agents, he said, had doubled engineering productivity in six months. But the gains were largest for an individual working alone with agents and shrank as team size grew, until coordination overhead inside large organizations swallowed most of them. If that observation is right, it makes decoupling more valuable, not less. The bottleneck is no longer production. It is communication between people.
Transparency is the other half. Mensch described feedback inside Mistral as "almost fully transparent," which he credited with letting the company scale without breaking its culture. Decoupled teams only work if information flows freely between them without requiring meetings to move it.
Tactic: Design your organization so that each team can ship without waiting on another, share infrastructure instead of approvals, and treat any new coordination meeting as a cost to be justified.
Principle 11
Move down the stack toward the scarce input
Mistral began as a model company. By 2026 it was building data centers near Paris and in Sweden, had acquired a cloud infrastructure startup, counted the world's leading lithography company as its top shareholder, and was exploring its own chips. The logic is in Mensch's own economics: along the chain from electron to token, about 10 percent of the value sits in the electron, about 90 percent elsewhere. Value accrues to whoever controls the conversion, and the conversion is constrained by whatever is scarcest.
In 2023 the scarce input was talent and algorithmic knowledge, and Mistral had both. By 2026 it was physical: chips, memory, helium, grid capacity. "A bubble is when demand is overestimated. The problem today is rather one of supply." A company that stays at the layer where it started will find its margins captured by whoever owns the bottleneck below it.
This is the inverse of the usual advice to focus. Mensch focused ruthlessly in year one, on text models and nothing else. He then moved down the stack deliberately, each step following the constraint. Focus is not a fixed layer. It is a fixed attention to wherever the binding constraint has moved.
Tactic: Every year, identify the scarcest input your business depends on, and if someone else controls it, make a concrete plan to own part of it before they price you out.
Principle 12
Never lose the ability to work without the tool
The CEO of a model company told Polytechnique students that learning to work without a computer remained indispensable for developing reasoning. He calls deskilling, the erosion of human expertise through overreliance on AI, the technology's most serious risk. Inside his own company, engineers no longer write lines of code and have become managers of agents.
These positions are compatible, and together they form a theory of how to manage a powerful tool. To direct agents you must understand what they are doing. To understand what they are doing you must have done it yourself, unaided, enough times to know what correct looks like. "You are no longer a craftsman; you are a manager" is only good news if the manager was once a craftsman. Mensch's skepticism of AGI fits the same view. Systems specialize, people direct them, and the value lies in "identifying the problem to solve," one level of abstraction above the code.
X
Conventional Wisdom vs. Mensch
Where his positions depart from the industry default.
| Conventional wisdom | Mensch's position |
|---|---|
| Bigger models win. | Better-allocated compute wins; the giants were undertrained. |
| Protect your weights; they are the moat. | Give away the base; sell adaptation, deployment, and control. |
| Sovereignty means independence. | Sovereignty means leverage over suppliers you cannot replace. |
| Regulate frontier models for safety. | Regulate uses; base-model rules favor incumbents. |
| AI is heading toward a general intelligence. | "There's no such thing as one system that is going to be solving all the problems of the world." |
| We are in an AI bubble. | Demand is not overestimated; supply of chips and electrons is the constraint. |
| Adopt AI and let skills follow. | Learn to work without AI in order to direct it well. |
Tactic: Before you automate a task, make sure at least one person on the team can still do it by hand and can recognize a wrong answer when the machine produces one.
Free playbook
Get The Business Model Playbook
58 business models, one visual page each: how the money flows, the metrics that matter, and who runs it. Free when you join the Faster Than Normal email.
Free. No spam. Unsubscribe anytime.
Part IIIQuotes and Maxims
In their words
— Arthur Mensch, to MIT Technology Review, 2024We think that it's very important for Europe to have a voice in how [AI] technology is going to shape our societies.
— Arthur Mensch, a16z Podcast, 'Safety in Numbers: Keeping AI Open'I think the battle is for the neutrality of the technology.
— Arthur Mensch, Big Technology Podcast, January 2026There's no such thing as one system that is going to be solving all the problems of the world. You don't have any human able to solve every task in the world. You of course need some amount of specialisation to solve problems.
— Arthur Mensch, speaking at the École Polytechnique, January 2026La génération d'ingénieurs qui est la mienne a à peu près réussi à commoditiser son propre métier.
— Arthur Mensch, No Priors podcast, 2023We did know where to start from and how to make a good model with a limited amount of compute and money. Well, not so limited, but at least more limited than where we were coming from.
Maxims
- Waste is the incumbent's weakness. Find the input your rival overspends on out of habit, and compete in that gap.
- Speedrun what you already know. Leave to build only once the market has caught up to your expertise, then skip every step a first mover had to discover.
- Size the round to the problem. If your home market cannot fund the ambition, raise abroad and lobby to fix the market.
- Let the release be the launch. A product handed to strangers who can judge it on first use is its own best advertisement.
- Give away what scales with users; sell what scales with specificity. The open base builds trust, and the adapted edge pays the bills.
- Many giants make no master. Partner with several dominant platforms at once so none can dictate terms.
- Sovereignty is a bargaining position. You rarely need independence; you need enough of the stack that your suppliers need you back.
- Regulation is terrain. Every rule advantages some size of company, so fight to put the burden on uses, not capabilities.
- The founder's job stops being technical fast. Hire the salespeople you cannot be, and send your scientists to meet customers.
- Follow the bottleneck down the stack. When the scarce input changes, move toward it before someone else prices you out.
