Europe spent this week testing how much of its AI stack it can take back, and the answers came in uneven. A Munich court reached across the Atlantic to rule on a training run that happened in the United States, and Finnish cancer diagnostics got EU money to scale.
The hardest case is Palantir. Governments from Madrid to Paris are trying to prise the American data firm out of policing, intelligence and health systems, and finding that a decade of embedded workflows does not unwind on a press release. (Politico Europe)
Research & Labs
Google DeepMind ships a robot brain that watches itself work
Gemini Robotics ER 2, released July 30, sits above motor control: it reads a live video feed, judges whether a task is going well, and hands off to lower-level models. It also lets several robots share semantic context in one workspace, which is the part factory integrators have been waiting on.
The published numbers are honest about the gap, with progress classification at 57.4% accuracy and moment-finding at 91.3%. European developers reach the model through the Gemini API, which is the part worth sitting with: the reasoning layer for Europe's robots is being rented, not built. (Google DeepMind)
European robotics raised more in six months than in the previous two years
Half-year figures published August 4 show European robotics startups raising more equity than in 2024 and 2025 combined, led by Germany, Switzerland and France. NEURA Robotics accounts for much of it, on the back of the billion-euro Series C we tracked through May and June and its fourth training gym this year, in Aachen.
The capital is real and the hardware is European. What sits on top of it, as the DeepMind release shows, mostly is not. (Sifted)
Deployment & Applications
The EIB puts €20 million behind Finnish cancer pathology
Aiforia, a 2013 spinout from the University of Helsinki, secured €20 million in venture debt from the European Investment Bank on August 3 under the InvestEU programme. Its deep learning models read ultra-high-resolution biopsy images to help pathologists find cancer cells faster.
The structure is the interesting part. The money arrives in three tranches, starting at €5 million, with each drawdown conditional on revenue targets written into the agreement rather than paid out as a grant. (European Investment Bank)
Lovable buys a Swedish team rather than a product
Lovable took on the three founders behind Nalvin, a Swedish AI agent platform for automating routine business work, in a deal announced July 31 with no disclosed terms. It is an acqui-hire, the fourth founding team the Stockholm company has absorbed since launch.
Lovable is valued at $6.6 billion and reportedly negotiating a round that would double that. In a European market short of senior AI engineers, buying a team beats hiring one. (tech.eu)
Governance & Safety
A Munich court says training abroad is no shield
On July 31 the Munich Regional Court ruled that Suno infringed copyright by training its music generator on works represented by GEMA, Germany's licensing body. The court found the system had memorised and reproduced six songs, took jurisdiction over training done in the United States, and rejected a fair use defence under US law.
Suno must disclose revenue tied to the infringement and faces damages, with a €250,000 penalty per further breach. It is the first European ruling to hold that where you run the training does not decide whose law applies. (The Next Web)
Washington finalises an AI framework it will not publish
The Trump administration completed its voluntary evaluation framework for frontier models on August 3, built on a June executive order that offers federal reviewers up to 30 days of pre-release access to models firms choose to share. The official announcing it gave no detail on what the framework actually contains.
The contrast with Brussels is now concrete. The same labs face published legal obligations on one side of the Atlantic and an undisclosed voluntary arrangement on the other, and only one of those can be audited by anyone outside the room. (CBS News)
Compute & Infrastructure
An Italian oil major opens exascale compute to startups
Eni opened its combined HPC6 and HPC7 systems to outside companies and research institutions on July 28, the first time a private industrial firm has offered exascale-class capacity as an external service. Together they deliver roughly one exaflop sustained, placing HPC7 sixth globally and second in Europe on the June TOP500 list.
Every other confirmed exascale machine belongs to a government lab. A European energy company reaching that scale and renting it out is a different route to sovereign compute than the EuroHPC procurement track. (Tech Times)
AI & Society
The nudification ban does not reach the people it hurts most
AlgorithmWatch published an investigation on July 28 into sex workers targeted by non-consensual sexualisation tools, days after the AI Omnibus made such tools illegal across the EU. Every worker interviewed reported losing income once they became afraid to advertise in their usual channels.
The gap is enforcement, not law. Police told victims that AI-generated images could not be proven to depict them, and platform takedowns demand identity documents that would expose a sex worker to worse.
Bellingcat researcher Kolina Koltai puts the market at 483 million downloads on the Apple store alone, and AlgorithmWatch tracked reported accounts returning after temporary suspension. That is the sort of evidence a regulator now has legal grounds to act on. (AlgorithmWatch)
One to watch
OpenEuroLLM's models are not due until December
Last week we flagged July 31 as the date the 20-institution consortium would ship its first models. That was wrong: the July 31 deliverable was the evaluation code package, and the first models and initial dataset are scheduled for December 31.
So the test of whether pooled EuroHPC capacity produces models researchers actually use is five months away, not five days. Prague still coordinates the project, and the December window now sits on top of the 2027 funding argument it was meant to inform. (OpenEuroLLM)