AI 2026: Mistral Will Rise As Compute is Unleashed

The French startup has delivered leading AI models under compute constraints. Backing from the French government, Nvidia, and ASML will give it access to much more compute.

Mistral cofounder and CEO Arthur Mensch in 2023. Screenshot from YouTube. Source.

This Bismarck Brief is part of our upcoming AI 2026 report, featuring in-depth analysis of the state of the artificial intelligence sector and technology, together with a comprehensive profile of key players. This Brief is publicly available; the rest are available to our paid subscribers only. Last week, we investigated DeepSeek and evaluated its research capabilities and collaboration with China’s rising semiconductor industry. Subscribe now to never miss a report.

Mistral is a French artificial intelligence startup headquartered in Paris with offices in London and Palo Alto. Founded in April 2023 by Frenchmen Arthur Mensch, Guillaume Lample, and Timothée Lacroix, the company makes large language models (LLMs) and sells access to them via its interface, La Plateforme. As of September 2025, the company is valued at $13.8 billion after a funding round in which it raised $2 billion.¹ In total, the company has raised over $3 billion. This currently makes Mistral the highest-valued AI startup in Europe, and its models have been well-received for their cost-efficiency. Despite its small size relative to leading U.S. AI companies like OpenAI, Anthropic, or xAI, Mistral has secured the support of Nvidia as well as the enthusiastic backing of the French government through the personal involvement of French President Emmanuel Macron. In September 2025, Europe’s most valuable technology company and a key link in the global semiconductor supply chain, the Dutch semiconductor equipment manufacturer ASML, took an 11% stake in Mistral for €1.3 billion, becoming the startup’s largest shareholder.²

Mistral initially raised around $100 million purely on the reputation of its three technical founders: CEO Arthur Mensch was previously an employee at Google’s DeepMind, while CTO Timothée Lacroix and Chief Scientist Guillaume Lample both worked at Meta. All three founders are in their early 30s. After fundraising, the company released and open-sourced a groundbreaking small model, Mistral 7B, which matched or outperformed Meta’s larger open-sourced Llama 2 models on widely-used benchmarks.³ The company followed up by offering access to Mistral Medium, a model that outperformed OpenAI’s GPT 3.5 on a number of standard benchmarks. The company has continued releasing a number of larger open-sourced and high-performing models, as well as Mistral Large, the company’s largest model, released in February 2024 and marketed as a competitor to OpenAI’s GPT-4.⁴

As of September 2025, Mistral reportedly had 350 employees, rapid growth compared to just 35 in early 2024.⁵ The company has recruited a sizable fraction of the team from Meta that built the early Llama models. Mistral was considerably constrained in its first year by a shortage of the high-end graphics processing units (GPUs) designed by Nvidia and manufactured by Taiwan’s TSMC, which are the preferred chips to develop and run AI models due to their great computing power. This did not however stop Mistral from developing globally competitive models in the meantime. The company has already gathered a series of allies and backers in both the U.S. and Europe who together can unlock more funding and access to chips for Mistral. As its “compute” resources grow, Mistral’s models are likely to become still more competitive while increasingly enjoying a home-field advantage among corporate and government customers in Europe, where Mistral can persuasively argue it is a patriotic alternative to U.S. or Chinese tech companies.

---

Become a paid subscriber and get a new in-depth investigation of a key live player, institution, or industry in your inbox every Wednesday:

---

Mistral Develops Cost-Effective AI Models Using French Talent

Mistral app logo on an iPhone in 2025. Photo by Solen Feyissa. Source.

Mistral released its first model in September 2023, a few months after the company’s founding. From the beginning, the company used a strategy of open-sourcing many of its models, allowing anyone to download their products without paying a fee to Mistral. This made the company popular with software developers and likely helped the company win some early customers for the models it restricted to paying customers. Mistral has effectively marketed itself using its idiosyncrasies, such as marketing its products under French names like Le Plateforme and Le Chat, and emphasizing its “hacker” credentials early on by releasing new models with simple torrent links and using a logo made in Microsoft Word.⁶ Mistral 7B, the company’s first model, was a large language model that proved highly popular with consumers looking for a free open-source model that was small enough to run on consumer hardware: it was downloaded over 3.4 million times from the AI site HuggingFace as of May 2024.⁷

Mistral followed up by releasing Mixtral 8x7B in December 2023, a sparse mixture-of-experts (MoE) model that was likely one of the world’s leading open-source models at the time of its release. In mixture-of-experts models, each token—numerical representations of text processed by the model—is routed to a subset of the whole model, according to the context: this allows for a model with a large number of total parameters to be relatively cheap to “inference” i.e. generate outputs for users, since only a small fraction of the whole model is actually used per token.

The technique is widely used for cutting the costs of actually serving models to users, and it can also cut costs of model “training” i.e. initial development depending on the quality of the training process. Although Mistral’s various MoE models have typically scored very well on benchmarks that are widely used to evaluate language models, especially for their low cost of inference, mixture-of-experts was not pioneered by Mistral and the company has not advanced the technique to the same degree as, for example, the Chinese company DeepSeek did, nor has Mistral published algorithmic and architectural advances as impressive as for example DeepSeek’s “multi-head latent attention.”

Mistral’s first year, however, saw the company operating under extreme compute constraints relative to other leading AI labs. The dominant paradigm for making further AI progress is currently the “scaling hypothesis.” In layman’s terms, this states that the key factor to make AI models better and smarter is simply applying ever more computational power (“compute”) to ever-larger data sets, from more and better chips. In an April 2024 interview, CEO Arthur Mensch stated the company only had access to 1500 H100 Nvidia GPUs, “just a few percent” of the compute available to leading U.S. AI labs.⁸ This seems to have been as much due to the massive rise in global GPU demand leading to sudden bottlenecks in supply as due to a shortage of funding: Mensch mentioned in an August 2024 interview that the company was just about to gain access to compute that it had raised money for in December 2023.⁹

Mistral was even substantially more compute-constrained than DeepSeek, which by this time already had a cluster of over 2000 H800 GPUs—almost equivalent to the H100—plus a cluster of 10,000 older-generation A100 GPUs. Given that the H100 is roughly 2.5 times more powerful than the A100, DeepSeek likely had roughly four times more compute available than Mistral did at the time. OpenAI’s GPT-4 was reportedly trained on a cluster of 20,000 A100 GPUs, equivalent to the compute power of approximately 8000 H100s—five times more compute than Mistral had access to almost a year into the company’s life.¹⁰ This makes Mistral’s achievements relatively more impressive and indicates the high quality of its researchers and software engineers.

Mistral has continued to release notable models, including Mistral Medium, a GPT-3.5 level closed-source model released in December 2023, and Mistral Large, initially released in February 2024; upgrades were released in July 2024 and again in November 2024. The original version of Mistral Large was somewhere between GPT-3.5 and the original version of GPT-4 in performance level and it was also notably less censored than competitor models from OpenAI and Anthropic, willing to discuss more controversial topics. Alongside Mistral Large, Mistral also launched its free chatbot, Le Chat, a consumer-facing product similar to ChatGPT.

As of September 2025, it has released dedicated coding, speech recognition, reasoning, image analysis, and optical character recognition models, as well as a range of multimodal models for more general use cases.¹¹ It continues to release smaller open-source models, but the company’s largest and most advanced models are almost all only available to paying customers. In October 2024, Mistral released a number of models called Les Ministraux, aimed at “edge computing” use cases such as phones and autonomous robotics.¹² In August 2025, the company released Mistral Medium 3.1, its most sophisticated model to date, with multimodal capabilities. Mistral claims the model offers around 90% of the performance of Anthropic’s Claude Sonnet 3.7 at a significantly lower cost of inference.¹³

Progress in AI has been driven by researchers with academic backgrounds in mathematics and computer science and, much like DeepSeek has mobilized Chinese talent, Mistral is mobilizing European and especially French talent. Arthur Mensch himself studied for his PhD at the French Institute for Research in Computer Science and Automation before working as a postdoc at New York University and the École normale supérieure, one of France’s most prestigious, elite grandes écoles or “great schools.” He worked in DeepMind’s Paris office for almost three years before leaving to found Mistral, playing a prominent role in the development of the Gemini, Chinchilla, and Flamingo models.

Mensch’s work on Chinchilla was especially important for the history of subsequent large language models, because the March 2022 technical paper that accompanied the model set out that OpenAI’s original scaling laws were incorrect in some critical respects, and language models were being trained on insufficient quantities of data for a given quantity of compute. The paper, on which Mensch was third author, has been cited over 2500 times, and “Chinchilla-optimal” is now widely used as a term for describing models trained in accordance with these laws.¹⁴

Mensch is an AI optimist who consistently downplays the existential risks of AI in his interviews, and disagrees with the idea that current-generation large language models are conscious, a key difference with figures such as Geoffrey Hinton.¹⁵ In this regard he seems more aligned with the thinking of Meta Chief AI Scientist Yann LeCun, another Frenchman, who is likely personally acquainted with Mensch’s technical cofounders and may be acquainted with Mensch himself. Mensch appears to have grown frustrated with the corporate bureaucracy at DeepMind, both the technical bureaucracy and the AI safety bureaucracy, and set up Mistral to create a faster-moving organization: he has discussed the benefits of keeping teams small, at no more than five people.¹⁶ He is also motivated by wanting to prevent an AI “oligopoly” dominated by U.S. companies and to create a European AI company that can compete globally.¹⁷

Guillaume Lample received a master’s degree from Carnegie Mellon University after attending the École polytechnique, another grande école specializing in science and engineering. He was briefly a research intern at Jane Street Capital and did a number of internships at Meta while studying for his PhD at the Sorbonne, working full-time at Meta for a little over three years, including on the first generation of Llama models. Timothée Lacroix also attended the École normale supérieure before working at Meta while studying for a PhD. He also worked on the first generation of Llama models. Lacroix, in a recent interview, talks about Lample’s mastery of data cleaning for model training as a key factor in the company’s success.¹⁸

Mistral’s technical staff are disproportionately French or have a European educational background. A connection to Meta is also common, due to a history of U.S. tech companies setting up offices in Paris to recruit elite French mathematical and computer science talent. Facebook AI Research, run by Yann LeCun, announced that it was setting up a Paris office in June 2015, giving the company deeper roots in France than other big tech AI labs.¹⁹ DeepMind followed suit three years later.²⁰ Of the 18 listed authors on the Mistral 7B paper, which included many of the company’s founders and earliest employees, 13 were French or had attended an elite French educational institution, including the company’s cofounders. Most studied at doctoral or master’s levels in France, such as Thomas X. Wang or Lelio Levaud, Head of Engineering, who has a master’s degree in computer science from the École polytechnique.

Of the five authors who did not have a French background of some kind, Albert Qiaochu Jiang, lead author on the papers for Mistral 7B and Mixtral 8x7B, has a PhD in Computer Science from Cambridge University, his master’s degree is from Oxford University, and he now leads the Mistral reasoning team. Chris Bamford, formerly at Meta, studied for his PhD at Queen Mary, University of London. Devendra Singh Chaplot was educated in India and at Carnegie Mellon University before working at Facebook’s AI research group FAIR, which is likely where he met the Mistral founding team. Gianna Lengyel, the company’s head of business development, is Italian. Diego de las Casas was educated in Brazil before working as a Research Engineer at DeepMind.

As the team has expanded over time, the preponderance of French talent has remained. Out of 99 authors listed on the June 2025 technical paper that accompanied Magistral, the company’s first reasoning model—comparable to DeepSeek’s R1—55 have predominantly French educational backgrounds, typically with master’s and doctoral degrees from French institutions.²¹ Of the remainder, seven were educated in the United Kingdom, and most of the rest of the staff, 25 in total, were largely educated in the United States, almost all of which are now working in Mistral’s U.S. office in Palo Alto. Of Mistral’s American staff listed on the Magistral paper, at least 11 were educated in India up to at least high school level, with some also receiving their undergraduate educations there before moving to the U.S. for graduate studies. This includes the lead author on the Magistral paper, Abhinav Rastogi, who grew up in India before studying at Stanford University and working for some years at Google and DeepMind.

Emmanuel Macron Has Enlisted Mistral to Become a European Strategic Champion

French President Emmanuel Macron embraces Indian Prime Minister Narendra Modi at the Paris AI Action Summit in 2025. Source.

Mistral has received the enthusiastic endorsement and support of the French state, evidently through the personal intercession of currently 47-year-old French President Emmanuel Macron and his office. One of Mistral’s “non-executive” co-founders is Cédric O, who served as Secretary of State for the Digital Sector in the French government from 2019 to 2022. O is a core member of Macron’s political clique, who served as treasurer of Macron’s initial 2017 presidential campaign and was reportedly one of the first five people to join Macron’s new political party.²² Like Macron, O sees Europe as ossified in technology and business and believes incubating successful European technology companies is a way to strengthen Europe’s “strategic autonomy.”²³ In his previous government role, O was a go-between for the French government and global technology executives, and has since acted as an advocate and lobbyist for Mistral in European government circles, including in Brussels. O is currently 40 years old.

The French state-owned investment bank Bpifrance has participated in multiple of Mistral’s funding rounds, including the June 2023 $100 million seed round and a $600 million round a year later.²⁴ Bpifrance is also partnering with Nvidia, Mistral, and the Emirati government’s new dedicated AI investment firm MGX to build a 1.4-gigawatt GPU cluster near Paris.²⁵ This cluster will presumably be used by Mistral for model training and inference, although it does not appear to be planned exclusively for the company’s use, and the size of the cluster will be smaller than some of the larger AI clusters currently under construction in the U.S., such as Meta’s planned 5 GW Hyperion cluster.²⁶

Mistral has also won a number of government contracts, including a recent award from the Prime Minister’s Office to upgrade a chatbot used by French civil servants.²⁷ The firm is also likely to be working with the French military, which operates its own AI research unit: Minister of the Armed Forces Sébastien Lecornu stated publicly in January 2025 that cooperation between Mistral and the military was guaranteed, and the company’s website also discusses ongoing collaboration with the Ministry of Defense’s AI agency.²⁸ The government has explicitly endorsed Mistral as a matter of marketing: Mistral CEO Arthur Mensch was a member of the government’s Commission on Artificial Intelligence that wrote a 130-page report in 2024 on France’s AI ambitions, and is also the face of a government-led marketing campaign for entrepreneurs and investors to “choose France” to do business in.²⁹ There are reports that Macron has gone so far as to have personal calls with key French-speaking researchers working for leading U.S. AI companies to try and convince them to come found or work for new French AI companies, presumably including Mistral.³⁰ Mistral executives met Macron shortly after the company’s founding.³¹

The support of the French government for Mistral has been crucial in moderating the European Union’s overzealous regulatory ambitions for software and data collection. Alongside Meta, Mistral has refused to voluntarily comply with portions of the EU’s 2024 AI Act that do not take legal effect until 2027, and its lobbying played a key role in watering down certain provisions of the act that sought to regulate large foundation models.³² In particular, the AI Act as passed exempts all models trained with fewer FLOPs—floating-point operations per second, a measure of computing power—than OpenAI’s GPT-4 from its most onerous provisions: all of Mistral’s current and many of its future models, therefore, are likely to be outside the Act’s scope. Mistral executives have met with European Commission leadership, but it is the political heft of France as the EU’s second-largest economy and only nuclear power that carries the most weight in Brussels negotiations.³³

As in the U.S. or China, the bulk of funding for Mistral’s AI ambitions will come from the private sector. Mistral has already proven an ability to raise significant funds from the U.S. venture capital sector: its 2024 funding round was led by U.S. venture capital firm General Catalyst, with a number of other U.S. firms participating, including the prestigious Andreessen Horowitz and Lightspeed Venture Partners. The French bank BNP Paribas and German media conglomerate Bertelsmann were notable European investors, alongside a number of corporate investors including Microsoft, Salesforce, Samsung, IBM, and Nvidia.³⁴ Notably, after OpenAI, Mistral was only the second company whose models Microsoft had offered.³⁵ Mistral has highlighted that Bpifrance, Andreessen Horowitz, and Nvidia also all participated in its September 2025 funding round alongside ASML.³⁶

Nvidia’s relationship to Mistral is key because Nvidia ultimately controls the supply of the most powerful computer chips needed to develop and run AI models. Nvidia founder and CEO Jensen Huang appeared on stage with President Macron and Arthur Mensch in 2025 to celebrate the partnership between the three to launch the new data center near Paris.³⁷ Huang related that he had asked Macron to call large companies in France to assist Mistral.³⁸ Huang likely has a more than an incidental financial interest in Mistral and the development of a European AI sector. A successful strategy for Nvidia would be to back smaller customers with equity investments and preferential chip allocations in order to avoid excessive dependence on a handful of Nvidia’s biggest customers and so risk a large drop of revenue if one or more of them shift the majority of their chip usage to custom chips developed in-house with the assistance of companies like AMD or Broadcom.

But with the investment of ASML into Mistral, the startup has arguably found a partner with high free cash flow that can fund expensive AI expansion ambitions, much in the same way the high profits of companies like Microsoft, Amazon, Google, and Meta have patronized the expansions of OpenAI, Anthropic, and others. ASML is one of the best-placed companies to patronize Mistral, since it is the most valuable technology company in Europe with a market capitalization of over $300 billion as of September 2025, generated $9 billion in free cash flow in its last fiscal year, and monopolizes the global production of advanced extreme ultraviolet photolithography (EUV) machines that are used by chip foundries like TSMC to manufacture the most powerful chips.³⁹ This means ASML will financially benefit from continued growth in semiconductors and AI and then be able to reinvest portions of its earnings into Mistral to ultimately capture more value from the other end of the supply chain, as an investor.

Notably, although it is a Dutch company, the CEO of ASML since 2024 is a Frenchman, Christophe Fouquet, and, in December 2024, the company appointed Bruno Le Maire as a special advisor to its executive board: the 56-year-old Le Maire served as Macron’s Economy and Finance Minister from 2017 until September 2024. In 2024, the company was reported to be contemplating an expansion in France, rather than the Netherlands, reportedly due to concerns at the prospect of losing access to foreign talent due to tightening Dutch immigration and tax laws.⁴⁰ It therefore appears as though the ASML investment may have been coordinated with the help of the French government too.

Mistral’s Future Brightens As Compute Constraints Lessen

Paris skyline in 2021. Photo by Atahan Güç. Source.

Mistral’s most notable customers so far are predominantly leading European firms. Its executives have repeatedly emphasized their focus on the European enterprise market, in which firms, especially those in highly regulated industries, might be unwilling or unable to use U.S. models. European regulations in the future could function as something of a barrier to entry in Europe that favors Mistral, while many large European firms, led by aging executives, are likely to follow the cues given by the French government, EU bureaucracy, and other European governments, if they signal that Mistral is the preferred AI vendor in Europe.

Political tensions between the second Trump administration and European governments have already reportedly driven European companies towards Mistral’s models as a high-quality alternative to U.S. companies like OpenAI and Anthropic: in May 2025, Arthur Mensch noted that the company had tripled its business in the last 100 days, with most of the growth coming in Europe and outside the U.S.⁴¹ Annual revenue is now reportedly over $100 million, although this is currently still far lower than, for example, Anthropic’s reported annual recurring revenue of around $5 billion.⁴²

Mensch has also noted how European firms are more likely to be storing their data “on-premises” or in private clouds, rather than in Amazon Web Services or Microsoft Azure. In the initial pitch memo that helped the company raise money in its seed round, the founders argued that most of the value in large language models would accrue at the level of generative models themselves, given the difficulty and expense of training high-quality models, and further argued that the closed-source, paid-access via API model of OpenAI would be inadequate for firms with critical data that they would not wish to expose to a closed-source model accessed via a public cloud.⁴³

On its website, Mistral lists Stellantis, the French-Italian automaker and world’s fifth-largest car company, as a key client, with Stellantis developing in-car AI assistants for drivers using Mistral models, while also deploying Mistral’s technology for analysis of complex component databases.⁴⁴ Other clients include BNP Paribas, French IT company Capgemini, German defense technology company Helsing, the Singaporean Ministry of Defense, German software company SAP, and French medical technology company Synapse.⁴⁵ Reportedly around half of Mistral’s business in total comes from Europe, with French logistics and shipping firm CMA CGM reportedly one of its biggest customers, signing a contract with Mistral worth €100 million.⁴⁶

Mistral has reportedly begun to “forward-deploy” its software engineers to work closely with its customers on customized software, on a quasi-consultancy model somewhat akin to that formerly extensively used by Palantir.⁴⁷ This indicates Mistral is targeting large bureaucratic organizations as customers, which tend to be very complex with high levels of technical debt in need of custom solutions. While this model slows down Mistral’s sales cycle, this business model is likely to embed a certain level of dependence on Mistral in its customers over the long term.

So far, Mistral seems to have been remarkably cost-efficient: it had not yet spent all of its $113 million seed funding over a year after the company’s foundation, despite the high costs of compute required to train and inference the company’s models.⁴⁸ For comparison, OpenAI has raised almost $60 billion in external funding and was most recently valued at $500 billion, while spending billions of dollars per year.⁴⁹ Mistral’s greatest constraint has been lack of funds and lack of access to enough GPUs, as the costs of building the largest and most high-performing models continue to escalate. Next-generation successor models to GPT-4 are estimated to cost over $1 billion just to train; future models beyond that are likely to see costs escalating into the tens of billions and tens of gigawatts of power, even before inference costs are accounted for.

The financial costs are likely to be much more of a problem for Mistral than the energy costs. Although average industrial electricity costs in France are roughly double those in the United States, Mistral is unlikely to be energy-constrained even if it does acquire the GPUs to begin training very large models over the next five years. The French nuclear fleet only runs at a capacity factor of around 75%, in part due to lengthy maintenance outages at older plants, but mostly because France simply has excess power supply capacity relative to demand, especially during the summer months when demand is lower and solar farms in France and neighboring countries such as Spain are producing their highest volumes of electricity.⁵⁰ The French grid, therefore, almost certainly can accommodate quite a sizable increase in power demand from Mistral’s GPU clusters without building any new nuclear plants: it can simply produce more power from its existing reactors. Furthermore, there is nothing stopping Mistral simply from building or renting additional GPU clusters in the U.S.

But Mistral’s access to financial and computational resources is likely to grow substantially in the future. The company’s growing business in Europe will attract further investment from both U.S. and European investors. It has now inked a strategic partnership with ASML, a company with robust cash flows and a high market capitalization. Mistral has already recently begun to rent out its own existing GPUs and AI training capabilities, with the most likely clients presumably being less well-established European startups, under the brand Mistral Compute.⁵¹ This indicates confidence that their compute resources will significantly rise in the future.

Although it is unlikely to receive major direct financial support from the French government, another potential source of large funding for Mistral, ASML, and other European companies in the future is perhaps surprisingly the EU itself, which is less constrained in funding big-budget projects than national legislatures due to its undemocratic nature: France has been a key supporter of the so-called “European Chips Act,” a proposed EU act for €42 billion in funding for semiconductors and AI, and there is no reason in principle that further future funding projects could not be pursued.⁵²

There is also no reason in principle why Mistral could not partner with a U.S. big tech company like Microsoft or Apple. Microsoft’s relationship with OpenAI has been slowly weakening in the aftermath of the attempted boardroom removal against Sam Altman. Microsoft could plausibly view Mistral as an important alternative LLM provider, especially if OpenAI becomes more of a direct competitor to Microsoft and other large tech companies. Media reports that surfaced in August 2025 indicated that senior executives at Apple had seriously contemplated buying both Mistral and AI search company Perplexity, but had to date not pulled the trigger on the acquisition in part due to their perception that Mistral is not a top-tier AI lab.

In many ways such a purchase would make sense: Mistral needs money, while Apple is cash-rich but has developed no cutting-edge LLMs, or even anything close to a cutting-edge LLM, and is widely perceived as having missed the boat in AI due to an old and conservative senior executive team. Additionally, the acquisition may well have faced serious political barriers given Mistral’s perceived importance to the French tech ecosystem, and that of Europe more broadly.⁵³ Such an outright acquisition by a U.S. let alone Chinese would negate the government’s strategic justification for backing Mistral. Fundamentally Apple’s evaluation of Mistral’s research work is likely incorrect and the unrealized acquisition a missed opportunity, despite such political considerations.

With its pool of French, European, and American research and engineering talent, Mistral has already proved it can produce leading models under challenging compute constraints. As these constraints lessen, the default projection is that Mistral’s models will become that much better and may even result in new algorithmic breakthroughs. In its current configuration, Mistral appears to be a live player with the support of multiple other live players including the French Presidency and Nvidia’s Jensen Huang, successfully building up a European software alternative to U.S. big tech companies in a cutting-edge domain, in a way that Europe has at best very rarely succeeded at before. Mistral stands a good chance not just of remaining a leading global AI company, but becoming the first European big tech company.

This Bismarck Brief is part of our upcoming AI 2026 report, featuring in-depth analysis of the state of the artificial intelligence sector and technology, together with a comprehensive profile of key players. This Brief is publicly available; the rest are available to our paid subscribers only. Last week, we investigated DeepSeek and evaluated its research capabilities and collaboration with China’s rising semiconductor industry. Subscribe now to never miss a report.

---

Become a paid subscriber and get a new in-depth investigation of a key live player, institution, or industry in your inbox every Wednesday:

1. Ruxandra Iordache. “AI Firm Mistral Valued at $14 Billion as Chip Giant ASML Takes Major Stake.” CNBC, 9 Sept. 2025, www.cnbc.com/2025/09/09/ai-firm-mistral-valued-at-14-billion-as-asml-takes-major-stake.html.

2. “ASML, Mistral AI Enter Strategic Partnership.” ASML, 9 Sept. 2025, www.asml.com/en/news/press-releases/2025/asml-mistral-ai-enter-strategic-partnership

3. Mistral AI. “Mistral 7B.” Mistral.ai, 27 Sept. 2023, mistral.ai/news/announcing-mistral-7b/

4. AI, Mistral. “Au Large.” Mistral.ai, 26 Feb. 2024, mistral.ai/news/mistral-large/.

5. See here at 11:50: link; Bradshaw, Tim. “ASML and Mistral Agree €1.3bn Blockbuster European AI Deal.” Financial Times, 9 Sept. 2025, www.ft.com/content/98e78f6b-0ebf-4546-b25f-bf7621e26c8b.

6. Goldman, Sharon. “Mistral AI Bucks Release Trend by Dropping Torrent Link to New Open Source LLM.” Venturebeat, VentureBeat, 8 Dec. 2023, venturebeat.com/ai/mistral-ai-bucks-release-trend-by-dropping-torrent-link-to-new-open-source-llm.

7. “@Singhsidhukuldeep on Hugging Face: “Mistral 7B Might Be One of the Most Popular Open-Source LLMs out There, with A….”” https://huggingface.co/posts/singhsidhukuldeep/337455617984321

8. 20VC with Harry Stebbings. “Arthur Mensch: Open vs Closed - Who Wins and Mistral’s Position | E1146.” YouTube, 29 Apr. 2024, www.youtube.com/watch?v=e7Y84vpWhkU.

9. Henshall, Will. “Mistral AI CEO Arthur Mensch on Microsoft, Regulation, and Europe’s AI Ecosystem.” TIME, Time, 4 Aug. 2024, time.com/7007040/mistral-ai-ceo-arthur-mensch-interview/

10. Patel, Dylan. “100,000 H100 Clusters: Power, Network Topology, Ethernet vs InfiniBand, Reliability, Failures, Checkpointing.” SemiAnalysis, 17 June 2024, semianalysis.com/2024/06/17/100000-h100-clusters-power-network/

11. Mistral AI, “Models Overview”, docs.mistral.ai/getting-started/models/models_overview/

12. “Un Ministral, Des Ministraux | Mistral AI.” Mistral.ai, October 16, 2024, mistral.ai/news/ministraux. https://mistral.ai/news/ministraux

13. “Medium Is the New Large. | Mistral AI.” Mistral.ai, May 7, 2025, mistral.ai/news/mistral-medium-3.

14. Hoffmann, Jordan, et al. “Training Compute-Optimal Large Language Models.” ArXiv:2203.15556 [Cs], 29 Mar. 2022, arxiv.org/abs/2203.15556.

15. See here: link.

16. 20VC with Harry Stebbings. “Arthur Mensch: Open vs Closed - Who Wins and Mistral’s Position | E1146.” YouTube, 29 Apr. 2024, www.youtube.com/watch?v=e7Y84vpWhkU.

17. Mistral Ai Strategic Memo. June 2023. https://promorphcapital.com/wp-content/uploads/2023/06/mistral.ai-strategic-memo.pdf

18. Applied Machine Learning Days. “Mistral AI: Frontier AI in Your Hands | Keynote 2 | Timothée Lacroix.” YouTube, 7 May 2024, www.youtube.com/watch?v=kDVOPKezeXs.

19. Romain Dillet. “Facebook Opens New AI Research Center in Paris | TechCrunch.” TechCrunch, 2 June 2015, techcrunch.com/2015/06/02/facebook-opens-new-ai-research-center-in-paris/

20. David, Eric. “.Google’s DeepMind opens new AI lab in Paris”, SiliconANGLE, 29 Mar. 2018, siliconangle.com/2018/03/29/google-deepmind-opens-new-ai-lab-paris

21. Mistral-AI, et al. “Magistral.” ArXiv.org, 2025, arxiv.org/abs/2506.10910.2025.

22. Daphné Leprince-Ringuet. “Brunch with Mistral AI’s Cédric O: “Europe Could Be Marginalised” | Sifted.” Sifted, 17 July 2025, sifted.eu/articles/brunch-with-cedric-o

23. Ibid.

24. “Bpifrance Supports French Companies in the Artificial Intelligence Revolution - Bpifrance.com.” Bpifrance.com, 30 June 2023, www.bpifrance.com/2023/06/30/bpifrance-supports-french-companies-in-the-artificial-intelligence-revolution https://legaltechnology.com/2024/06/11/year-old-llm-vendor-mistral-ai-raises-510m-and-is-valued-at-4-9bn/

25. Swinhoe, Dan. “MGX, Bpifrance, Nvidia, and Mistral AI Plan 1.4GW Paris Data Center Campus.” Datacenterdynamics.com, 20 May 2025, www.datacenterdynamics.com/en/news/mgx-bpifrance-nvidia-and-mistral-ai-plan-14gw-paris-data-center-campus/.

26. Zeff, Maxwell. “Mark Zuckerberg Says Meta Is Building a 5GW AI Data Center | TechCrunch.” TechCrunch, 14 July 2025, techcrunch.com/2025/07/14/mark-zuckerberg-says-meta-is-building-a-5gw-ai-data-center/.

27. Beurnez, Victoria. “L’État Signe Un Contrat Avec Mistral Pour Booster Son “Chatbot” Albert.” Acteurs Publics, 12 June 2025, acteurspublics.fr/articles/letat-signe-un-contrat-avec-mistral-pour-booster-son-chatbot

28. “20 Janvier - Mistral AI va Travailler Avec l’Agence Du Ministère Des Armées.” Maddyness - Le Média Pour Comprendre l’Économie de Demain, 24 Jan. 2025, www.maddyness.com/breve/20-janvier-mistral-ai-va-travailler-avec-lagence-du-ministere-des-armees

29. “Marquez Les Esprits. Choisir La France. - Accueil.” Choosefrance.fr, 2025, choosefrance.fr/en.

30. Anonymous sources.

31. Petitjean, Olivier. “AI Act : Le Troublant Lobbying Des “Champions” Européens, Mistral AI et Aleph Alpha.” Observatoire Des Multinationales, 2019, multinationales.org/fr/enquetes/intelligence-artificielle-lobbying-et-conflits-d-interets/mistralai-alephalpha-gafam-ai-europe

32. Drew, Rocket. “Meta, Mistral Will Not Sign EU’s AI Pledge.” The Information, 25 Sept. 2024, www.theinformation.com/briefings/meta-mistral-will-not-sign-eus-ai-pledge

33. Petitjean, Olivier. “AI Act : Le Troublant Lobbying Des “Champions” Européens, Mistral AI et Aleph Alpha.” Observatoire Des Multinationales, 2019, multinationales.org/fr/enquetes/intelligence-artificielle-lobbying-et-conflits-d-interets/mistralai-alephalpha-gafam-ai-europe

34. Hill, Caroline. “Year-Old LLM Vendor Mistral AI Raises £510m Series B and Is Valued at £4.9bn - Legal IT Insider.” Legal IT Insider, 11 June 2024, legaltechnology.com/2024/06/11/year-old-llm-vendor-mistral-ai-raises-510m-and-is-valued-at-4-9bn/

35. Gilchrist, Karen. “Microsoft Invests in Europe’s Mistral AI to Expand beyond OpenAI.” CNBC, 26 Feb. 2024, www.cnbc.com/2024/02/26/microsoft-invests-in-europes-mistral-ai-to-expand-beyond-openai.html

36. “Mistral AI Raises 1.7B€ to Accelerate Technological Progress with AI.” Mistral.ai, 2025, mistral.ai/news/mistral-ai-raises-1-7-b-to-accelerate-technological-progress-with-ai.

37. Alexandre Piquard. “At VivaTech, Macron Hails “Historic” Partnership between Mistral AI and Nvidia.” Le Monde.fr, Le Monde, 12 June 2025, www.lemonde.fr/en/economy/article/2025/06/12/at-vivatech-emmanuel-macron-hails-historic-partnership-between-mistral-ai-and-nvidia_6742267_19.html.

38. “French President Rallies behind Mistral-Nvidia Cloud Partnership.” PYMNTS.com, 11 June 2025, www.pymnts.com/artificial-intelligence-2/2025/french-president-rallies-behind-mistral-nvidia-cloud-p

39. Data from Koyfin.

40. Jacob, Sarah, and Diederik Baazil. “Netherlands Working on Plan to Deter ASML Expansion Abroad.” Bloomberg.com, Bloomberg, 6 Mar. 2024, www.bloomberg.com/news/articles/2024-03-06/asml-plan-to-expand-abroad-worries-government-telegraaf-says.

41. Ibid.

42. Sri Muppidi. “Anthropic Revenue Pace Nears $5 Billion in Run-up to Mega Round.” The Information, Aug. 2025, www.theinformation.com/articles/anthropic-revenue-pace-nears-5-billion-run-mega-round

43. Mistral Ai Strategic Memo. June 2023. https://promorphcapital.com/wp-content/uploads/2023/06/mistral.ai-strategic-memo.pdf

44. “Stellantis and Mistral AI.” Mistral.ai, 2024, mistral.ai/customers/stellantis

45. “Customer Stories | Mistral AI.” Mistral.ai, 2025, mistral.ai/customers

46. Bradshaw, Tim. “ASML and Mistral Agree €1.3bn Blockbuster European AI Deal.”, Financial Times, 9 Sept. 2025, www.ft.com/content/98e78f6b-0ebf-4546-b25f-bf7621e26c8b.

47. Bradshaw, Tim.. “Europe’s Mistral Benefits from Search for Artificial Intelligence Alternatives.” Financial Times, 8 June 2025, www.ft.com/content/65f79839-d637-48a7-a0f2-3fab8952b315.

48. Henshall, Will. “Mistral AI CEO Arthur Mensch on Microsoft, Regulation, and Europe’s AI Ecosystem.” TIME, Time, 4 Aug. 2024, time.com/7007040/mistral-ai-ceo-arthur-mensch-interview/

49. Hu, Krystal. “OpenAI Closes $6.6 Billion Funding Haul with Investment from Microsoft and Nvidia.” Reuters, 2 Oct. 2024, www.reuters.com/technology/artificial-intelligence/openai-closes-66-billion-funding-haul-valuation-157-billion-with-investment-2024-10-02/.

50. EDF, “2024 Facts and Figures”, https://www.edf.fr/sites/groupe/files/2025-04/annual-results-2024-facts-and-figures-en-2025-04-09.pdf

51. “Mistral Compute - Less Setup, More Ship.” Mistral.ai, 2025, mistral.ai/products/mistral-compute.

52. Robinson, Dan. “Euro Semi Firms Push for “Chips Act 2.0” to Expand beyond Manufacturing.” Theregister.com, The Register, 20 Mar. 2025, www.theregister.com/2025/03/20/european_chips_act_2/.

53. Huang, Kalley. “Why Mistral Is a Takeover Candidate.” The Information, 17 Oct. 2024, www.theinformation.com/articles/why-mistral-is-a-takeover-candidate