Invest1 publisher3 min readPublished
Mistral contests a frontier pause from 23.5 points behind the top-ranked US model
Dario Amodei's essay asks frontier labs to pace capability gains and asks for antitrust room to coordinate on safety. Europe's challengers, holding 11% of global AI venture funding, say the leaders wrote it for themselves.
The Investor · Invest desk
What happened
- Anthropic's Dario Amodei published an essay on September 12, "We Must Pace the Frontier", arguing labs should slow advances in AI capability while continuing to train models and release products.
- Alongside the pacing, he asked for an antitrust exemption so labs can work together on safety, with external evaluators given employee-level access as the first of three steps.
- Prosus put Europe at 11% of global AI venture funding as of 2023 against 77% for the United States, and said the gap widens at later funding stages.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- constraint Pacing as drafted stops at capability gains and leaves training and product launches running, so what still holds Mistral back is Europe's 5% share of global AI data-centre compute.
- decision Governments funding AI factories under the AI Continent plan have to choose whether that money buys compute capacity or a seat at threshold-setting, and the two budgets compete.
- precedent An antitrust exemption for safety coordination would make a standing forum of frontier labs lawful, and its membership rules get written while the best European entrant sits 23.5 points off the top.
Epoch AI's capability index puts GPT-6 Astra at 166 points and Mistral Medium 3.5, released in April, at 142.5 [13]. The difference is 23.5 points [1]. That score was built on 11% of global AI venture funding, the share Prosus gave for 2023 against 77% for the United States [10], and on 5% of the world's AI data-centre compute, the figure Christine Lagarde gave in an address on September 14 [12]. So Mistral's flagship reaches 86% of the leader's number [2] on a seventh of the venture money [3] and a fifteenth of the compute [4]. Lagarde also counted 59 notable AI models from the United States in 2025, 35 from China, and one each from France and the UK [11]. Epoch does not estimate the compute cost of closing 23.5 points, and Lagarde's speech does not either.
Amodei's essay, published September 12 under the title "We Must Pace the Frontier", slows capability advance while product releases and training continue [3]. The second of his three steps is coordination among laboratories in democratic countries [4], which on its face includes France. The competitive stake is in the other ask: an antitrust exemption so labs can work together on safety matters [2].
Mistral's representatives believe established players are "pushing for regulation designed to favour them over competitors", Reuters reported [5]. Raphael Auphan, chief operating officer of Proton, said the initiatives are "totally self-serving" and serve US companies' interest in keeping others dependent on their services [6]. Black Forest Labs argues that "arbitrary thresholds" can impede innovation [7]. Clement Delangue of Hugging Face thinks Europe should advance, and he supports Amodei's call for independent assessment [8].
The dependency argument has the stronger evidence. Generative AI usage among working-age people in Norway ran at 48.6% in Microsoft's Global AI Diffusion report [15], and the continent serving that demand holds 5% of the compute [12]. The lock-in argument is thinner, because a cap on capability that leaves training running freezes 166 while 142.5 keeps moving [3][13]. In my view the term that actually costs Mistral money is the evaluator requirement, since external assessors with employee-level access [4] are a fixed compliance charge, and a lab drawing on 11% of the venture pool carries it less easily than one drawing on 77% [10]. If a point on Epoch's index scales with compute spend, then 23.5 points is several training generations and the 86% is not a near miss [1][2]. And if pacing stays voluntary and unenforced, nothing binds anyone, and the 15-to-1 compute ratio settles the next release [4].
Sam Altman endorsed the embedded evaluators and said pacing should not be understood as stopping development [9]; Elon Musk also backed the proposal [1]. Forbes reported the issue spilled into the market on September 14, when Intel fell about 7% to around $96, AMD about 6% to $487 and Nvidia about 3% to $212 [14].
What to watch
- Whether any antitrust authority grants the safety-coordination exemption Amodei asked for, and which labs qualify for it.
- An update to the European funding share, which in the record still dates to Prosus's 2023 count of 11%.
- Mistral's next model score measured against the 166 points at the top of Epoch's ranking.