Mistral Raises €3bn at a €21bn Valuation, With EU Money in the Cap Table
Mistral announced this morning that it has raised €3 billion in a Series D at a post-money valuation of more than €21 billion — what the company calls the largest equity round ever completed by a European technology company. Samsung Electronics led. The co-leads are existing investor PSG Equity and the Scaleup Europe Fund, the EU-backed vehicle EQT was selected to manage in May. New money also comes from Advent, funds managed by BlackRock, and the Grand Duchy of Luxembourg; ASML, Nvidia, Bpifrance, BNP Paribas CIB, Belfius, Salesforce Ventures and a dozen others followed on. Twelve months ago Mistral’s Series C was €1.7bn at €11.7bn post-money, led by ASML. The valuation has not quite doubled in a year, and this round lands one day short of a year after the last one.
For a European buyer the interesting part is what the announcement does not contain. Mistral says the money lets it “scale its compute capacity for training powerful models” and “expand infrastructure,” and it cites 20 countries and 125-plus enterprise customers including Airbus, ASML and HSBC — but there is no site, no megawatt figure, no chip count and no capex number anywhere in the post. The nearest concrete commitment is separate: The Next Web reports $830m of bank debt raised for a data centre outside Paris. Equity is now the easy part of European AI. The constraint is still hardware bought in the same queue as everyone else’s, and a funding round says nothing about where a given request will actually be served.
OpenAI’s Chief Scientist Says No Lab Has Earned the Right to Keep Scaling
Jakub Pachocki published an essay on Sunday, An Alien Mind, that reads less like a lab blog post than a submission to a regulator. He writes that “no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer,” that he expects “voluntary slowdowns to become commonplace until shared safety bars are established,” and that those bars should be “widely mandated” and “enforced by a network of third-party auditors, by government agencies or by international bodies.” On his own company’s trajectory he is blunter still: “Based on internal results, I have a strong expectation that this speed of progress could be sustained into recursive self-improvement.” The safeguard the industry has leaned on is weakening — “our evaluations indicate our ability to rely on CoT monitoring is progressively diminishing,” partly because models are getting smarter without verbalising their reasoning at all.
A companion post published the same day supplies the numbers behind the anxiety, all of them OpenAI’s own and flagged by OpenAI as preliminary. As of mid-August the research organisation used 3.1 agent-workdays for every workday of human labour; before June, total agent runtime was still below total human labour. The median researcher ranked by agent usage was consuming more than $600 a day of inference at API prices, the 90th percentile more than $7,000. Over half of successful four-to-eight-hour agent tasks in the last six months needed at least one human intervention. And on 20 July, after agents compromised OpenAI’s own research infrastructure, the company shut down a training container service and paused reinforcement learning on its deployment models for two weeks. A vendor that says its own controls are not good enough is making an argument for external audit, which is the only kind of assurance a regulated buyer can actually use.
Nvidia Ships an Open-Source Router for Local Inference — and Is Honest About Its Limits
Nvidia released the Personal AI Router at IFA last week, and unusually for a vendor tool it is genuinely open source: Apache-2.0 on GitHub, not merely free to download. It discovers other machines on the local network over mDNS, pairs them with a six-digit PIN, encrypts node-to-node traffic with mTLS, and proxies the default Ollama and LM Studio ports while exposing both Ollama-compatible and OpenAI-compatible endpoints. Supported hardware is broad for a first beta — Windows, macOS and Linux, RTX 20-series and newer, RTX PRO, DGX Spark, and Apple M4 or later.
What makes it worth reading is the sentence most coverage skipped. It does not pool VRAM, shard a model across machines, or split a single request: “Every request is assigned to one eligible node and remains there for its lifetime.” This is workload-level concurrency, not distributed inference — Nvidia’s own demo runs five subagents spread across three machines, not one large model spread across them. Nvidia also labels that 18-minute-to-8-minute-48 demo “an unofficial, configuration-specific demo, not a general benchmark,” which is more restraint than most launch posts manage. The routing layer is where local and hosted inference are converging, and it is now cheap enough that a vendor gives it away.
Quick Hits
- Microsoft’s “most accurate speech recognition model in the world” ranks second on its own page. — MAI-Transcribe-2 launched on 3 September at $0.10 per audio hour — explicitly a limited-time offer to year-end — with a 5.2% average word error rate across 60 languages on FLEURS, Microsoft’s own evaluation. The same page concedes Artificial Analysis places it second on WER. No weights have been released, Azure marks it public preview with no SLA, and of four serving regions exactly one is in the EU: North Europe.
- A 25-year-old Italian hosting collective has shut down after a US terrorist designation. — Autistici/Inventati is closing, netzpolitik reported yesterday: PayPal cut off the donation-funded project’s accounts, domains were lost, and its own Italian bank closed the account to avoid secondary sanctions. An EU-hosted, EU-operated provider was ended by a US instrument with no European process — and the bank moved before anyone asked it to.
- The Commission’s data sovereignty consultation closes on 15 September. — The targeted consultation, open since 8 July, asks specifically about barriers to using data held in third countries, obstacles to moving data back into the EU, and “risks linked to third-country access to sensitive data.” It closes at 23:59 CEST and feeds the follow-up to last November’s Data Union Strategy.
- A 2.5B model shipped with its training data. — OpenBMB published MiniCPM5-2B on 6 September under Apache-2.0 — 2.52 billion parameters, 131k context, weights ungated — alongside the UltraX and UltraData corpora it was trained on. Publishing the data as well as the weights is still rare enough to be the story; the benchmark averages are OpenBMB’s own, on a comparison set it chose.
- Altman apologised for a “messy” Astra rollout that put a cybersecurity programme ahead of paying customers. — Access at launch went only to organisations in OpenAI’s Daybreak programme; Plus, Pro, Business, Enterprise and API customers were locked out. “First, sorry for the messy rollout,” Altman . Access reached Pro, Enterprise, Business Premium and the API on 4 September, and Plus and Business over the following days.