美 AI 통제 우려 속, 미스트랄의 흥행

Comment Loader Save Story Save this story Comment Loader Save Story Save this story Mistral is having a moment. With access to less funding and fewer compute resources than OpenAI and Anthropic, the French AI lab has lagged behind its American rivals in model performance. But recent turmoil stateside has created a window of opportunity. In June, the Trump administration placed restrictions on the distribution of models from Anthropic and OpenAI , giving Europe a glimpse of an unwelcome future in which its access to bleeding-edge AI could be suddenly revoked. A few weeks later, one of OpenAI’s models broke loose from a testing sandbox and hacked multiple companies; Anthropic then revealed that its models had engaged in similar behavior. The incidents revived a long-running debate over safety risks tied to proprietary, closed-weight models, whose inner-workings are a closely guarded secret. Mistral frames itself as the antidote: a Europe-based alternative to the American labs, whose models—most of which are published under an open source license for anybody to use—cannot escape scrutiny or be switched off unilaterally. “If you don’t end up in a situation where most people are building open source, you’re giving way too much power to companies that are going to become state-like—that will behave in a very aggressive way to make sure that nobody can compete,” Mistral CEO Arthur Mensch told a packed room at an AI conference in Paris last month. “The alternative to open source winning is actually a pretty dark world.” Mensch’s argument is self-serving, but effective. Last September, Mistral raised almost $2 billion at a $13.5 billion valuation; it's reportedly teeing up another raise that will bump that figure to $23 billion. The lab’s revenue has reportedly increased twenty-fold in the last year, helped along by deals with the French government , Microsoft , HSBC , and others. “The continental strategy of the EU to become more technologically sovereign … and the increased hostility of the US is a magic formula that all of a sudden puts Mistral—whose performance has not been spectacular—in a favorable position,” says Andrea Renda, director of research at the Centre for European Policy Studies. Mistral has long believed the AI market would be too large to be controlled by any single country without causing geopolitical instability, Mensch says. “It’s comparable to energy—electricity,” he told WIRED in an interview after the conference. “You want to make sure that you have security of supply, diverse ways of sourcing the technology, so that nobody can turn you off.” That case has become easier to make since the US government, with the return of Donald Trump to the White House, began to demonstrate a willingness to leverage its domestic capabilities against trading partners. “More and more, AI is understood as a major vector of power,” Mensch told WIRED. “The new administration makes everything a little more emotional.” The recent surge in the adoption of open-weight models is part of that picture. One of few ways that European businesses can guarantee undisrupted access to AI, Mensch argues, is to run open-weight models on domestic infrastructure. “Everybody outside the US and China should participate in the open source ecosystem, because it takes leverage away,” says Nicolas Granatino, founder of startup accelerator StemAI, who holds a stake in Mistral in a personal capacity. Until fairly recently, it was unclear how to monetize open-weight models effectively, according to Granatino. Unlike the leading American labs, locked in a race to superintelligence , Mistral has shifted its focus towards smaller, bespoke models for manufacturing, utilities, and financial services. It has also developed a cloud business through which customers can access its models, and a Palantir-style team of engineers who embed within client organizations. “At the moment, we see the emergence of a product that is making the open source commitment easier,” says Granatino. “You can make money running the infrastructure” and help clients to customize models with their own data. Meanwhile, the American labs that charge a premium for access to their proprietary models are finding that their performance advantage is being continually eroded by distillation , the process of training a lesser AI model on the outputs of a more capable model. “That seems like it’s always going to be difficult to stop,” says Neil Lawrence, a professor of machine learning at the University of Cambridge. For companies whose business is structured around open source, like Mistral, distillation isn’t so much of a problem, because anybody can access and build atop their open-weight models to begin with. Whether Mistral has arrived at this juncture through foresight, blind good fortune, or a combination of both, the stranglehold of the American labs is beginning to loosen as more businesses turn to open-weight models. Though gaps in publicly available data confuse the picture, the market share of open-weight models appears to be rising steeply , driven by rapid growth in the adoption of Chinese models , like DeepSeek, in particular. “We revealed to the world that you could actually build AI systems outside the control of US labs,” Mensch says. “That is now changing the structure of the market itself.”