When AI researchers Arthur Mensch, Guillaume Lample and Timothée Lacroix came together a year and a half ago to create a new startup called Mistral AI in Paris, it caused quite a stir.
The trio, who worked in the science and engineering teams of Meta and DeepMind, founded the company with the promise to build large-language models (LLMs) from Europe that could compete with those developed by huge US players like OpenAI, which are the backbone of widely used tools like ChatGPT.
Their strength? Their brains. “The real differentiator is whether or not you have talent capable of building state-of-the-art LLMs,” Jonathan Userovici, general partner at US VC Headline, an investor in Mistral, tells Sifted.
“Mistral built the company at the right time with people who had the right expertise [...] A year and a half ago, there were probably fewer than 100 of these experts in the world.”
The pitch worked. Mistral raised a €105m seed round just weeks after launching; a €385m Series A and a €600m Series B shortly followed, featuring high-profile US investors like General Catalyst, Lightspeed and Andreessen Horowitz.
Mistral is now valued at €5.8bn — making it Europe’s most valuable AI startup.
But it’s not only about building high-quality LLMs — the company also has to sell them. “Our vision is to become a leading actor in the field, while developing a very valuable business around integrating these models in the European industry and beyond,” reads the startup’s pitch memo from its seed round.
Mistral’s team of just over 100 people started monetising products at the beginning of 2024; the company is on track to finish the year with €30m annual recurring revenue (ARR), two sources with direct knowledge of the matter, who wished to remain anonymous to protect relationships, tell Sifted. Mistral declined to comment.
The startup is up against some deep-pocketed, well-established tech behemoths. OpenAI, which was founded in 2015, is reported to be nearing $3.7bn (€3.5bn) in revenues this year, and to be anticipating to reach $11.6bn in 2025. It has a team of 1,800 employees.
Meta is also investing billions in building its Llama LLM, while Google has developed Gemini. US-based Anthropic, which was founded by OpenAI alumni and also develops LLMs, has raised more than $10bn from tech giants like Google and AWS.
Meanwhile, Germany’s Aleph Alpha, one of Europe’s other key AI players, recently pivoted away from building its own LLM to developing a “generative AI operating system” to sell to customers.
The company, which missed its revenue goals in 2023, told Sifted on stage at Sifted Summit in October that “it’s hard to build a working business model” around building LLMs, and that Big Tech companies throwing money at the technology was creating an uneven playing field for startups.
“Our view is that there will be a couple of players that will dominate the market,” says Harald Nieder, general partner at Swiss VC Redalpine, an investor in Mistral. “That is absolutely where we see Mistral.”
What does Mistral sell?
Mistral launched with the promise to be an open-source competitor to companies like OpenAI — meaning the startup releases LLMs that developers can freely download the code for and use for their applications.
In February, the company released its first flagship commercial product — an LLM called Mistral Large — which can be accessed through a cloud-based API for a fee, in a similar way to other closed-source models like OpenAI or Gemini.
Since then, Mistral has fleshed out its commercial offering. The startup still offers a range of models that are free to access, but it also has premium models — including Mistral Large — targeting enterprises, which vary in size and tasks they specialise in.
These can be used through a cloud-based API and are charged on a pay-as-you-go basis based on the number of prompts and completed tasks. Alternatively, enterprises can licence the models and deploy them directly on their premises, rather than going through the cloud — meaning deployment is less ‘off-the-shelf’ than with an API, but they can access the source code and have more control over the model.
Mistral also monetises a range of services to support customers as they deploy its models for different use cases — like creating a customer support assistant or a chatbot.
The company says that this approach means applications are more customised and efficient, but also safer: businesses can build using their own data and keep the technology on their premises.
Redalpine’s Nieder says this is where Mistral’s real value lies. “People in the end won’t bother about the underlying model,” he says. “What they really want is for it to be really integrated into their workflow, with their own data, with an on-premise solution where they can rely on the data not going away.
“The application is what really counts for the users.”
Closed-source models like OpenAI’s ChatGPT, Google’s Gemini and Anthropic’s Claude offer cloud-based access rather than on-premises; Meta’s Llama, which is open-source, can similarly to Mistral’s models be deployed on enterprises’ own infrastructure.
Who is Mistral selling to?
Mistral tells Sifted that almost 1m developers have used the startup’s free and premium models. The company declined to specify how many paying customers it has secured so far.
Marjorie Janiewicz, Mistral’s US general manager, tells Sifted that the startup is seeing the most traction in the US and Europe from regulated companies and organisations — the public sector, financial services, insurance companies and hedge funds — but also among digital-native tech startups.
In Europe, it has signed deals with French insurer AXA and telco Orange, as well as fintech Qonto and online marketplace Mirakl. In the US, where Mistral started hiring a team six months ago, customers include legaltech startup Harvey and Q&A website Quora.
“We’re seeing an acceleration in the business when it comes to major clients like Fortune 500 companies in the US and beyond,” says Janiewicz.
In July, BNP Paribas announced a partnership with Mistral to start putting into production a range of use cases that it had been testing for several months — such as customer-facing chatbots, virtual assistants to help banking advisors prepare for client meetings and services to transcribe and analyse phone calls with customers.
The French bank is trialling several LLMs including OpenAI’s models, says Sophie Heller, chief operating officer at BNP Paribas’s Commercial Personal Banking and Services division. “It’s not an exclusive partnership,” she tells Sifted. “We’re taking the best choice depending on the use case.”
Mistral was particularly suited to use cases requiring sensitive data, Heller says. “We ban the use of ChatGPT internally, because we don’t want confidential data to leave [the bank’s premises],” she says.
“Mistral gives us the possibility of using the models in our own infrastructure.”
Can Mistral win the LLM market?
“There will be a small group of players [building LLMs],” says Nieder. “And if it’s a space dominated by a few players, you need to choose where you will differentiate yourself.”
“The differentiation that Mistral is choosing is to not have a closed-model approach.”
Janiewicz says that Mistral’s focus on integrating models into enterprises’ workflows thanks to its open-source approach is how the company hopes to differentiate itself as the market selects its champions.
“Customers need support [...] and until now they didn’t have a partner that could support them [as they adopt AI],” she says. “So, I am very confident that we can compete.”
It sets the startup apart from OpenAI, which has an enterprise offering but resonates even more with consumers thanks to its flagship tool ChatGPT — which the US company says registers 250m users per week.
Earlier this year, Mistral released a ChatGPT-style conversational platform called ‘Le Chat’ — which the startup updated last week with functions like web search with citations and image generation, making Le Chat even more similar to OpenAI’s tool.
Weekly users of Le Chat increased from 100k-300k after the updates, according to one source with knowledge of the matter. Mistral declined to comment.
The startup seems to be planning to increase efforts to target the consumer market. For now, Le Chat is free and Janiewicz says that it is “premature” to discuss monetising the platform; a premium version of the tool is planned in the next few months, according to the same source. Mistral declined to comment.
But Mistral is not only competing against OpenAI. Meta’s Llama, Google’s Gemini, Anthropic and more recently Chinese company Alibaba’s open-source LLM Qwen are all competing for market share.
When they’re not spun from Big Tech’s research labs, some of these companies have strong financial support from major tech firms. Microsoft has invested $13bn in OpenAI, while AWS has poured $8bn in Anthropic.
Although not at the same scale, Mistral is backed by some big players, including IBM, Nvidia and Microsoft — which reportedly invested €15m in the startup. Janiewicz says that the company has also secured distribution deals with all three major hyperscalers in the US — Microsoft’s Azure, AWS and Google Cloud.
And Mistral seems keen to get closer to its US competition: in six months, the company’s team there has grown to 26 people, according to a Sifted analysis, which represents a third of its total workforce.
In his 2024 ‘State of AI’ report, Nathan Benaich, the founder of AI VC Air Street Capital, which is not an investor in Mistral, describes the startup as “the undisputed European foundation model champion”.
“[Mistral] have proven their ability, independently from Big Tech and with a small team, to develop models that are on par or better than the market’s top LLMs, with much fewer resources,” says Userovici.
“Yes, there is a hype, but there are good reasons: there is a real shot that it becomes a leader.”
Mistral is currently valued at 186x its revenues — showing investors are bullish about the startup’s future. The next few months will show if they bet on the right horse.
Update, December 2, 2024: This article has been updated to reflect that BNP Paribas is not using Google's Gemini model for AI applications.
