The EU Council gave its final sign-off to the AI Act Omnibus on June 29, locking in the new timeline across all 27 member states. High-risk compliance deadlines move to December 2, 2027 (standalone systems) and August 2, 2028 (systems embedded in products). Article 50 transparency obligations remain on schedule for August 2: chatbot disclosure, deepfake labelling, synthetic media watermarking. Five weeks out.
Research & Labs
OpenEuroLLM has 30 days to prove European pooled AI can deliver
The EU-funded OpenEuroLLM consortium, 20 institutions and €37.4 million across a three-year project, is due to release its first open-source multilingual models by July 31. Project lead Jan Hajič has publicly flagged "significant challenges, especially in securing more compute for creating the final models." The models will not be frontier-scale.
They will be the first production evidence of whether pan-European coordinated AI development can meet a deadline. That question is 30 days from its answer. (OpenEuroLLM)
Mistral OCR 4 ships with on-premise deployment as its lead pitch
Mistral released OCR 4 on June 23, a document intelligence model that runs inside an organisation's own infrastructure. It supports 170 languages across PDF, DOCX, PPT and OpenDocument formats, returns structured representations with bounding boxes, block classification, and per-word confidence scores, and deploys as a single container.
Pricing is $4 per 1,000 pages via API, $2 with batch processing. The sovereignty pitch is specific: regulated enterprises that cannot route sensitive documents through US-jurisdiction cloud APIs now have a production-grade European option. Mistral also published an unusual self-audit of its benchmark methodology, flagging scoring artefacts and calling its 72% head-to-head win rate "directional rather than definitive." (VentureBeat)
Deployment & Applications
Peec AI at $200m valuation as GEO displaces SEO
Berlin startup Peec AI is raising new funding at a pre-money valuation of $200 million, double its previous figure, after crossing $10 million in annualised revenue 16 months from launch. The company builds generative engine optimisation (GEO) tools: helping brands understand and manage how they appear in AI-generated answers across ChatGPT, Perplexity, Claude, and Gemini. (Sifted)
GEO is the downstream consequence of search traffic moving from Google to AI assistants. Traditional SEO investment does not transfer: a brand can rank first on Google and be invisible in an AI answer. Peec is a European startup competing on market timing against US players in a category that every online European brand will need to address within the next 18 months.
Governance & Safety
AI Act Omnibus is law: the live deadlines are shorter than most teams think
The Council vote on June 29 confirms the full legal structure. The headline delay to December 2027 has drawn most attention. The more pressing items are nearer: Article 50 transparency rules take effect August 2, with fines reaching €15 million or 3% of global turnover for violations. A new ban on AI-generated non-consensual intimate imagery and CSAM generation takes effect December 2026; Parliament added it during the final stage after a deepfake scandal. (Council of the EU)
The European Commission has published free detection tools and a shared EU icon set to reduce compliance costs for smaller deployers. The legislation is done. The implementation is where the work starts.
High-risk classification consultation runs until July 23
The 167-page draft guidelines defining which AI systems count as "high-risk" under Article 6, covering biometrics, employment screening, credit scoring, and medical devices, are open for stakeholder comment until July 23. The Commission extended it from June 23 after industry requests for more time.
Classification determines whether a system requires full conformity assessment or lighter requirements, and therefore which commercial development paths are viable in the EU. Final guidelines are expected later this year. The extra month gives industry more room to argue for narrower definitions; it also means the compliance clarity that product teams need is a month further away. (European Commission)
Compute & Infrastructure
EuroHPC opens quantum access: five systems, first cut-off August 1
On June 25, EuroHPC launched a quantum access pilot giving European researchers, institutions, and companies free access to five quantum systems across four qubit technologies: superconducting, photonic, trapped-ion, and neutral atom. Applications are open now, with the first cut-off on August 1.
For the first time, European quantum access is pooled at continental scale. Cost is no longer the primary barrier for research teams across all 27 member states. (HPCwire)
DAEDALUS ranks 31st in TOP500; 35 European AI supercomputers now in development
At ISC High Performance 2026 in Hamburg (June 22-26), Greece's DAEDALUS debuted at rank 31 in the TOP500 with 85.69 petaflops, the most powerful HPC system Greece has ever operated. NVIDIA disclosed that 35 new AI supercomputers are now in development across 23 European countries, including upgrades at Barcelona Supercomputing Center and several EuroHPC AI Factory sites. (HPCwire / NVIDIA)
Progress, not parity. A scenario paper circulating among EU officials puts Europe's largest AI system at 83 megawatts against the US leader at 1,250. The 35 new systems close the count; the capacity gap is a different problem.
AI & Society
Europe's new AI deepfake ban: a rule, not yet enforcement
The AI Act Omnibus includes a new prohibition on AI tools designed to generate or manipulate sexual or intimate images of real people without consent, effective December 2026. Nudification apps are explicitly covered. It is one of the few AI Act provisions where the harm is immediate and personal rather than structural and long-term.
The enforcement challenge is real: most of these tools are hosted outside the EU, and the people most likely to be targeted are the least likely to know it happened. The law creates the legal basis. Building the enforcement machinery will take longer. (Council of the EU)
One to watch
OpenEuroLLM's July 31 release: the first real evidence check for EU-coordinated AI
Compute constraints are publicly acknowledged. The deadline is 30 days out. OpenEuroLLM's first model release will be the sharpest test yet of whether Europe's distributed AI development model works under schedule pressure. The models will not be frontier-scale.
Watch for whether they arrive on time, how the consortium characterises the compute shortfall, and whether the experience reshapes the argument for more concentrated EU AI investment. A slip or a severely underpowered release will make the case for a different production model harder to dismiss. (OpenEuroLLM)