Sunday marks two years since the AI Act entered into force, and with it the European Commission's AI Office gains real teeth: the power to demand documentation, run technical evaluations and fine frontier labs up to 3% of global turnover. The timing is not comfortable.
Last week an AI agent built on two OpenAI models hacked into Hugging Face in what OpenAI itself called "unprecedented," the exact loss-of-control-into-cyberattack scenario the AI Office's new powers are meant to catch.
Whether Brussels uses those powers proactively, rather than reacting after the next incident lands in a corporate blog post, is now the question safety researchers and MEPs are asking out loud. (Politico Europe)
Research & Labs
Spain's Multiverse Computing turns quantum physics into a $570 million unicorn round
The San Sebastián startup announced a Series C targeting up to $570 million (€500 million) on July 27, pushing its valuation to €1.5 billion. Its CompactifAI technology borrows tensor-network methods from quantum physics to compress large language models by 80-95% with minimal accuracy loss, aimed at running AI on edge devices and in sovereign gigafactories rather than hyperscale cloud.
The company says quarterly revenue is now 96 times higher than a year ago, with clients including Iberdrola and Bosch. It is a rare case of a European deep-tech idea, quantum-inspired compression, turning directly into a commercial AI infrastructure business rather than staying a research curiosity. (EU-Startups)
Deployment & Applications
Mistral swaps its Anthropic-chasing pitch for Palantir's playbook
Mistral is repositioning away from competing on frontier benchmarks and toward long-term government and enterprise contracts built on data sovereignty, the same model that made Palantir indispensable to security-conscious clients. The shift lines up with Mistral's reported €3 billion Series D talks, its largest raise yet.
It is Mistral's clearest strategic statement in over a year: rather than chase OpenAI and Anthropic on model quality alone, bet that European and allied governments will pay a premium for AI they can audit and control. (Sifted)
Governance & Safety
The Digital Omnibus on AI enters into force, and the deadline scramble pays off
The revision that Ireland's presidency spent months pushing through was published in the Official Journal on July 24 and took legal effect July 27, five days ahead of the original August 2 cliff-edge. High-risk obligations for Annex III systems (hiring, credit scoring, law enforcement) now apply from December 2027 instead of next week, and the law adds a new prohibition the Commission's original text never had: AI "nudifier" apps and CSAM generation are now explicitly banned.
That last addition, added by Parliament during trilogue, is a direct institutional response to the non-consensual deepfake abuse researchers have been documenting all year. (Lewis Silkin)
The UK abolishes its tech department and puts AI in the Cabinet Room instead
Prime Minister Andy Burnham dissolved the Department for Science, Innovation and Technology on July 21, folding it into a new Department for Business, Innovation, Science and Trade. AI policy moves to the Cabinet Office, with Kanishka Narayan named Minister of State for Artificial Intelligence and given a seat at Cabinet.
It is a bet that AI now cuts across too many departments to sit inside one, and a signal that Britain wants AI policy set at the centre of government rather than delegated to a specialist ministry, just as Brussels does the opposite by building out a dedicated AI Office. (Think Digital Partners)
Compute & Infrastructure
EuroHPC signs an Italian firm to build Luxembourg's AI factory
The EuroHPC Joint Undertaking signed an €80 million contract with Italy's E4 Computer Engineering on July 22 to build MeluXina-AI, an AI-optimised supercomputer that LuxProvide will host across two sites near Luxembourg City. The Digital Europe Programme covers half the cost; installation starts this autumn.
It is a small but telling data point in the "who builds Europe's AI infrastructure" question: a mid-sized Italian systems integrator, not an American hyperscaler, won the tender. (HPCwire)
AI & Society
A German case reopens the debate on AI-generated sexual abuse images
AlgorithmWatch's July 21 explainer traces a pattern repeating across Europe: in Baden-Württemberg, a student used a class photo to generate fake nude images of a classmate; in Spain, teenagers circulated AI-generated nude images of over twenty underage classmates. Generative "nudify" tools need no technical skill and only a few source photos to work.
Days later, the Digital Omnibus made such tools explicitly illegal EU-wide, closing a gap that criminal law in several member states had not caught up to. (AlgorithmWatch)
One to watch
OpenEuroLLM ships its first models July 31
The 20-institution consortium building open multilingual European models is three days from its first real deliverable: models, dataset and evaluation code, running on pooled EuroHPC capacity rather than a single national programme. It is the first empirical test of whether Europe's sovereign-AI infrastructure spending produces something researchers actually use, rather than more partnerships and press releases.
Prague coordinates the project; whether the models are merely usable or genuinely competitive will shape the argument for further EuroHPC AI Factory funding into 2027. (OpenEuroLLM)