Mistral Doubled to 21 Billion Euros. Samsung Led. Brussels Co-Led.

Yesterday OpenAI measured 3.1 agent-days per researcher and asked the industry to slow down. This morning Europe wrote a check.

Mistral closed a 3 billion euro Series D at a post-money valuation of more than 21 billion euros. The company calls it the largest equity fundraising round a European technology firm has ever completed. Samsung Electronics led. The co-leads are the part worth pausing on: existing backer PSG Equity, and the Scaleup Europe Fund, the roughly 5 billion euro vehicle Brussels stood up so the continent’s biggest tech companies would not have to raise abroad.

Twelve months ago ASML led a 1.7 billion euro Series C at 11.7 billion euros. The valuation nearly doubled in a year. Both of the last two rounds were led by manufacturers, not by software funds. ASML makes the machines that make the chips. Samsung makes the chips, the memory, and a great deal else.

That is the whole day, in one sentence. The intern is still a ratio. The check is cash.

The table, not the slogan

Mistral’s own post is a sovereignty pitch with a cap table attached. New money: Advent, BlackRock-managed funds, the Grand Duchy of Luxembourg. Existing money that came back: a16z, ASML, Nvidia, Bpifrance, General Catalyst, Index, Lightspeed, Salesforce Ventures, and a long European list. Customers named in the same breath: Airbus, ASML, HSBC. Operations in 20 countries. More than 125 enterprises.

The Register puts the scale in the only comparison that matters: OpenAI’s last round carried 122 billion dollars of committed capital at 852 billion dollars. Mistral is a European beast and a global minnow. The argument is not that it outspends either US lab. The argument is that a government or a factory does not want its intelligence loop sitting in someone else’s cloud when the next export control lands.

The Next Web has the clean read on the co-lead. Scaleup Europe is EQT-managed public-backed capital sitting at the front of Europe’s biggest private tech round. That is not a vibe. That is industrial policy buying a seat.

The Journal has CFO Johan Bergqvist putting annual recurring revenue slightly above 1 billion dollars, with an IPO still described as a possibility rather than a date. Treat that as a reported number, not a filing.

Sovereignty with Nvidia in the cap table

Mistral’s line, quoted in the Register: during the first wave the question was who could build the most powerful model. Organizations now ask how to use one without surrendering the infrastructure and the intelligence loop. Four control claims: data stays inside the wall, models are customizable, compute is private and predictable, production is auditable. Open weights are the product. Owned halls are the insurance.

Arthur Mensch told CNBC the long-term plan is to fully rely on capacity Mistral builds itself, with owned compute growing around 100 percent over five years. Data Center Dynamics has the existing pile: 830 million dollars of debt earlier this year for 13,800 Nvidia GB300s at Eclairion’s Bruyères-le-Châtel site south of Paris, a July Microsoft compute deal across Europe on Vera Rubin GPUs, Digital Realty Paris South, and a 1.2 billion euro plan at EcoDataCenter in Borlänge, Sweden.

Nvidia is an investor. Nvidia is also the silicon in the halls. That is not a contradiction if you are honest about the next five years. It is a contradiction if you sell “sovereign” as “US-free.” Mistral is selling control of the loop, not a domestic process node. Samsung in the lead seat is the memory-and-foundry version of the same bet ASML made last September: the industrial base wants a lab that will run inside a factory, not a chatbot that phones home.

If you are going to take that pitch seriously on a box you actually own, 64GB of DDR5 and a 2TB NVMe still matter more than another SaaS seat. Open weights are only open if you can load them.

A worm that rings, then walks the contact list

While Paris was counting euros, a California shop published the other Tuesday number.

Calif says it built WeWorm, a zero-click worm that spreads through WeChat voice calls on iOS and Android. Attacker calls victim. Victim does not have to answer. Exploitation takes seconds. Full control of the WeChat account: read, send, call, act as the owner. Demo chain: Pixel 10a calls iPhone 17e while it is still ringing, compromised iPhone calls another Pixel. Friend-list required, which is a speed bump, not a wall. Compromise one contact and the trust graph does the rest.

The timeline is the AI story. An internal model found a VoIP-stack memory-corruption bug sometime in July. First Android RCE in about two days. iOS RCE a few days later. Polished cross-platform worm in a week. Calif’s line: a worm at this scale used to take a larger team months. AI already does most of the work. Tencent patched clients (Android 8.0.77, iOS 8.0.76) and mitigated server-side; Calif confirmed the exploit dead for all users on August 28 and published with the Times today. Technical details withheld for a conference. WannaCry is the named accident they do not want repeated.

This is not a lab claiming it can write exploits in a gated eval. This is a disclosed, patched, cross-platform messenger worm assembled with model assistance on a consumer everything-app used by more than a billion people. The constraint that used to be “who has the team” is now “who has the judgment not to ship the half-finished copy.”

If you run agents against anything that touches a phone, the boring controls still win: hardware keys beat SMS, and an AIO cooler on the box you actually patch is more useful than a new policy PDF.

Anthropic looked at 6 billion dollars of chip-squeezing and said no

The third Tuesday print is a deal that did not happen.

Bloomberg, via The Next Web: Anthropic walked away from buying Israeli startup Decart after due diligence. Talks had been reported in August at around 6 billion dollars. Both sides declined to comment. Collaboration is still on the table.

The misread since August was that Anthropic wanted Decart’s world models. The reported rationale was efficiency: software that makes existing chips work harder on training and inference, so Anthropic’s already-contracted compute absorbs more demand. Decart raised 300 million dollars in May at close to 4 billion dollars. Six billion would have been a 50 percent markup in four months, and it would have been an order of magnitude larger than Anthropic’s notable prior buy (about 400 million dollars for a team of fewer than ten). Globes adds an Israeli-side claim that founders would not move the center of gravity from Tel Aviv to San Francisco. Bloomberg’s version is quieter: diligence finished, buyer left.

A lab staring at a late-year listing does not want a messy 6 billion dollar premium in the S-1. Buying a way to squeeze more tokens from the same racks is rational. Paying up for it weeks before public-market scrutiny is a different conversation. Watch for a commercial license instead of a close.

Mistral just raised 3 billion euros to own more of the rack. Anthropic just declined to buy a company whose pitch was making the rack go further. Same week, two answers to the same cost curve. OpenAI’s intern is still running 3.1 agent-days per human on that curve. None of those are a safety bar. One of them is a check Europe can actually cash.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.