Mistral raises €3B, valuation tops €21B

Good morning 👋 A chipmaker just wrote the biggest check in European AI history, and the reason behind it matters more than the number does.

In today's issue:

• 🔭 Mistral raises €3B, valuation tops €21B
• 🧠 DeepMind maps every possible DNA letter change
• 🔬 Agent memory rot, credit for typing over clicking, and catching a bad coding run before it burns compute
• 🔗 OpenAI's journalism push, Runway's ARR claim, and a chip startup backed by Alphabet

🔭 THE ONE THING

💰 Mistral raises €3B, valuation tops €21B

Mistral closed a €3 billion Series D at a post-money valuation above €21B (~$24B), nearly double the €11.7B mark it hit a year ago. Samsung Electronics led the round, flanked by EQT's Scaleup Europe fund, PSG Equity, and a cap table that reads like a sovereignty play: Luxembourg's government, BlackRock, ASML, NVIDIA, Salesforce Ventures. A chipmaker leading the biggest AI check outside the US says more about the strategy than the valuation does. Mistral calls this the largest equity round ever raised by a European tech company. Fair claim, and also a hedge: raise enough that nobody asks whether the revenue is there yet.

🧠 MODELS & RELEASES

• 🧬 Google DeepMind shipped AlphaGenome Atlas: a 1-petabyte map predicting the molecular effects of all 9 billion possible single-letter DNA changes in the human genome, over 30x the size of the AlphaFold Database. It's free for academic use, and the real value sits in the 98% of the genome that isn't a gene, the part nobody could score at this scale before. AlphaGenome Atlas

🔬 RESEARCH HIGHLIGHTS

• Agent memory doesn't survive a model swap. A new controlled study puts a name on something most builders have felt but not measured: upgrade the model under a long-running agent and its memory store can quietly rot. The new model reads old notes differently, mixed embedding versions break retrieval, and repair often isn't possible without the original evidence trail still intact. If your agent has persistent memory, a model upgrade is a migration project now, not a config flag. Paper

• CUA-Universe stops rewarding agents for clicking when they should type. Computer-use benchmarks like OSWorld score agents that click through GUIs, even when a terminal command would finish the task in a tenth of the steps. CUA-Universe builds environments that grade hybrid GUI-plus-CLI behavior, so an agent that opens a shell instead of hunting through menus finally gets credit for it. Overdue fix. Paper

• Flagging a bad coding agent before it burns compute. Coding agents fail the expensive way: they act confidently wrong, and you only find out after the sandbox has already run the retry. This paper's "Speculative Uncertainty" method pulls a failure signal straight from the agent's own output tokens, before execution, so you can catch the run likely to go sideways instead of paying for it first. Paper

🔗 QUICK LINKS

• OpenAI is pushing its journalism push further, from classroom tools for students to training and partnerships for newsrooms. OpenAI
• OpenAI's chief scientist Jakub Pachocki wants real safeguards and international coordination before the next capability jump, not after. OpenAI
• Runway says ARR hit $200M in September and it's telling Bloomberg $350M is the target by year-end. That's the company's number, not an audited one. Bloomberg
• Celero Communications raised $275M at a $3B+ valuation for chips that link AI data centers over long distances, with Alphabet's CapitalG backing the round. Bloomberg

See you tomorrow.

Pradeep Perugu