Among the world leaders and Big Tech bosses meeting at Bletchley Park, England, last week to debate the regulation of artificial intelligence, was one less well-known chief executive.
Arthur Mensch, founder of Paris-based start-up Mistral AI, was representing the only European company present on the second day of the UK’s AI Safety Summit, when about 30 executives and politicians gathered for a more intimate discussion than that held by the previous day’s 100 attendees.
Mensch told the Financial Times he believed his six-month-old company was invited alongside the Silicon Valley luminaries from Microsoft, Google, Meta and OpenAI, because of its “technical expertise”.
“We have been pioneers of the technology,” he said, pointing to his own work creating advanced AI models at Google DeepMind and that of his co-founders, Guillaume Lample and Timothee Lacroix, at Meta.
“We compete with everybody,” he added, despite the company’s youth and small size next to Microsoft-backed OpenAI or Google.
That expertise — which investors say only a few dozen people worldwide can claim — has put the fledgling start-up at the centre of the current investor frenzy for AI.
Silicon Valley heavyweights including General Catalyst and Andreessen Horowitz are participating in an investment of as much as €400mn in a deal that could value Mistral at between €1.5bn and €2bn, including the new capital, according to people with direct knowledge of the negotiations.
The terms of Mistral’s latest funding are in flux as it grapples with what one investor called a “bunfight” among prospective backers.
Mistral, Andreessen Horowitz and General Catalyst declined to comment on the discussions. Some details of the financing were previously reported by The Information and Business Insider.
The interest in Mistral stems from its work in the fast-moving world of generative AI, where so-called large language models (LLMs) are capable of creating humanlike prose and code in seconds.
Most investors have focused on Silicon Valley-based groups: OpenAI is exploring an employee stock sale at a valuation of $86bn while Anthropic has recently received investment commitments from Google and Amazon, which could total $6bn.
Mistral’s emergence is all the more eye-catching in the European market, where €1bn-plus companies are more scarce than in the US. The French government has also been cheerleading Mistral as a symbol of President Emmanuel Macron’s ambition for Europe to foster its own homegrown AI players so as not to get left behind as it has on key technologies such as semiconductors and internet platforms.
Mistral’s existing venture capital investors say that despite the high expectations set by its record-breaking €105mn seed round in June, when it was just four weeks old, the company is making rapid progress.
Jeannette zu Fürstenberg, co-founder of La Famiglia, an early Mistral investor that merged with US-based General Catalyst last month, believes the French start-up’s emergence can be a “real Eureka moment for Europe”. “Europe is so good at driving research but so bad at capturing the commercial upside,” she said.
Antoine Moyroud, a partner at Lightspeed Venture Partners, which led Mistral’s first fundraising round, said he was “increasingly excited about the business” because of the speed with which it was moving — releasing its first AI model in September, three months ahead of schedule.
“They have outperformed our internal expectations,” he said, despite “pretty aggressive forecasts”, adding: “The ability to show flexibility and execution speed is important in a world which is moving so fast.”
Multibillion-dollar price tags for young AI companies with immature businesses have prompted some tech investors to draw parallels with the dotcom bubble.
“VCs all want a piece of the next big AI thing and they are willing to swallow huge valuations for that,” said Mike Volpi, a partner at Index Ventures. Some venture capitalists were “too optimistic about the timeframes” for AI’s transformative effects, he said, which was driving some to make “overvalued investments in the short term”.
But Mensch, while refusing to discuss Mistral’s latest funding talks, insisted that AI companies’ skyrocketing valuations simply reflect that “this is a revolution that is certainly comparable to the internet”.
“The valuation of something that may have close to infinite return, because it changes so much of society that our jobs will never look the same, is very hard to estimate,” he said.
Mistral is not yet making any money but Mensch expected that to change “before the end of the year”, as it prepares to release a new platform for customers to access its AI models.
Pia d’Iribarne, partner at New Wave, a Paris-based investor in Mistral that counts French billionaire Xavier Niel among its backers, said the “fundamentals are there” for big AI businesses to be built.
“That doesn’t mean that there are not some crazy valuations and crazy investor behaviours right now in AI,” she said. “But the high valuations for LLM companies are more linked to the price of compute and scarcity of teams, than that behaviour.”
Mistral’s edge over its larger and better-funded rivals was in its efficiency, Mensch argued. It launched its first AI model with a team of just 10 people, spending less than $500,000 on training costs, he said, in contrast to the tens of millions that rivals spent.
“We are happy to be the most capital-efficient [LLM] company,” he said.
Another differentiator, argued zu Fürstenberg, was Mistral’s “open source” approach, in which the model was published publicly. This gives corporate customers more control over their data and visibility over its use, as well as attracting developer talent.
Despite its lower overheads, Mensch admitted: “We do need more capital.” That was primarily to spend on the chips necessary to train LLMs, he said. “Graphics processing units are going to be our main cost for a while.”
Venture investors are betting that there is room for only a handful of companies to profit from developing the models that underpin chatbots. But developing and training those models is hugely costly, requiring massive computing power and deep technical expertise.
Even as new rivals such as Elon Musk’s xAI emerge, Lightspeed’s Moyroud argued that Mistral might be the last new LLM developer with a credible chance at challenging OpenAI.
“There is a growing sense today that the ship has sailed, in part because of the capital intensity,” he said. “If someone was to go out to market [today] with similar ambitions and a similar quality team, I’m not sure they would get funded.”
Additional reporting by Madhumita Murgia in San Francisco
Read the full article here