Germany's innovation agency SPRIND closed applications yesterday for its Next Frontier AI Challenge, a €125 million competition to fund up to three European frontier AI labs targeting the next architectural paradigm rather than the current transformer generation. Jury pitches on June 24-25 will be the first public test of whether European teams have built credible next-paradigm hypotheses. The challenge closes the same week Mistral's first sovereign compute facility is due to come online outside Paris — a rare convergence of research ambition and infrastructure reality in European AI.
Research & Labs
SPRIND closes €125M frontier lab challenge; pitches June 24-25
Germany's federal agency SPRIND accepted applications until June 1 for its Next Frontier AI Challenge, a €125 million, 24-month competition to fund up to three European frontier AI labs from scratch. Jury pitches are scheduled for June 24-25; the first ten funded teams begin work in July.
The challenge explicitly disqualifies transformer optimization, model reproductions, conventional agent architectures, and brute-force scaling. SPRIND is betting on the next architectural paradigm: state-space models, energy-based systems, neuro-symbolic approaches, scientific foundation models, novel training frameworks. The argument is that the current S-curve is already owned and European teams cannot compete on it at scale with European budgets.
SPRIND is designing toward a €1 billion follow-on raise per winning lab, though that capital would have to come from external investors. Whether it materializes depends entirely on whether the technical bets pay off; the answer will not be visible before 2028.
CuspAI and Kemira compress years of PFAS chemistry into six months
Cambridge-based CuspAI and Finnish water treatment company Kemira announced on May 21 that they had used generative AI to design new materials for removing PFAS from drinking water. CuspAI's platform searched a space of roughly 300 trillion possible material structures and returned 5,000 novel metal-organic frameworks with full property data for three priority PFAS molecules: GenX, PFBS, and PFOS. Around 20 candidates are now moving into further development.
The discovery phase took six months. Comparable work through traditional experimental chemistry would have taken years. This is the first commercial partnership to apply generative AI end-to-end to PFAS material design rather than screening existing compounds.
PFAS contamination is an unresolved public health problem across European water systems, with enforcement pressure building under the EU's revised Drinking Water Directive. CuspAI has offices in Cambridge, Amsterdam, and Berlin.
(Kemira / Intelligent CIO Europe)
Deployment & Applications
Spain presses ahead with AI rules despite $20M US lobbying campaign
Spain's digital transformation minister Óscar López confirmed the country would proceed with draft rules requiring social media platforms to disclose how their recommendation algorithms work and imposing tighter controls on high-risk AI systems. The confirmation came after US tech companies spent roughly $20 million on European lobbying in Q1 2026, with federal filings showing the pressure targeted these exact Spanish proposals.
Spain has also announced plans to ban social media use by teenagers and to hold executives personally liable for hate speech on their platforms. Whether its position shifts EU-level negotiations or remains a national outlier will depend in part on how CAIDA — the Cloud and AI Development Act expected this week — is finalized.
Malta becomes first country to give all citizens free ChatGPT Plus
OpenAI signed a deal with Malta's government to provide every Maltese citizen and resident with 12 months of ChatGPT Plus at no cost, the first partnership of its kind between OpenAI and a national government. Participants must first complete a short AI literacy course developed by the University of Malta; upon completion, access is granted through Malta's national digital identity system.
The programme, called AI for All, launched its first phase in May and is managed by the Malta Digital Innovation Authority. The literacy-gate requirement is notable: Malta is testing whether pairing access with basic training changes how citizens actually use the tool, a model with obvious relevance to broader European AI competency debates.
(The Next Web / Euronews)
Governance & Safety
AI Omnibus: what the May 7 deal actually changes
The EU's Digital Omnibus on AI political agreement extends the high-risk compliance deadline for standalone Annex III systems (biometrics, critical infrastructure, employment, law enforcement, border management) from August 2026 to December 2027. AI embedded in regulated products under Annex I gets even more time, until August 2028. Co-legislators framed this as a sequencing correction: the technical standards companies need to implement the rules are not yet published.
The deal bans AI systems whose primary purpose is generating non-consensual intimate imagery, with a December 2026 compliance deadline. Simplified documentation, proportionate quality management requirements, and reduced fine caps previously available only to SMEs are now extended to small mid-cap companies. The AI Office gains stronger enforcement powers over general-purpose AI models. Formal adoption by Parliament and Council is expected before August.
AI transparency consultation closes June 3; Code of Practice due this month
The European Commission's consultation on draft transparency guidelines under Article 50 of the AI Act closes tomorrow, June 3. The guidelines cover the obligations kicking in August 2: providers must inform users when interacting with AI, and deployers must disclose when people are exposed to deepfakes, emotion recognition, or AI-generated content on public-interest matters.
The final Code of Practice on marking and labelling AI-generated content is expected to be published in parallel this month, two months before those transparency rules apply. The code is voluntary but designed to serve as a compliance demonstration path. Companies that have not yet mapped their Article 50 obligations are running short on time.
Compute & Infrastructure
Mistral's first data center goes live this month outside Paris
Mistral AI's first owned compute facility, in Bruyères-le-Châtel south of Paris, is due to become operational by end of June — the first time a European frontier AI company has brought its own GPU cluster online rather than relying entirely on US cloud providers. The facility holds 13,800 Nvidia GB300 GPUs across 44 megawatts of capacity, financed through an €830 million debt facility backed by Bpifrance, BNP Paribas, and five other French and international banks.
The Bruyères-le-Châtel facility is the first step toward Mistral's target of 200 megawatts of European compute by 2027, with a second site under development in Sweden. Whether European-owned compute at this scale changes the actual cost structure for European AI labs — or whether it mainly changes the sovereignty optics — will become clearer once the facility is running and utilization data is public.
AI & Society
Pope Leo XIV's encyclical: AI must not make lethal decisions or concentrate power
Pope Leo XIV published his first encyclical, "Magnifica Humanitas," on May 25. The 42,000-word document condemns delegating lethal or irreversible decisions to AI systems and argues that AI must be freed from the logic of military, economic, and cognitive competition. The pope describes this process as "disarming" AI: not rejecting the technology, but refusing the assumption that technical power automatically confers the right to govern.
The framing draws on Pope Leo XIII's Rerum Novarum response to the Industrial Revolution. That document shaped Western European labor law over decades, not through direct legal force but through sustained institutional pressure and political coalition-building. Whether a papal encyclical can do comparable work in AI governance is an open question. What is not open is that institutional voices arguing for limits on AI power concentration are rare right now.
(Vatican News / National Catholic Reporter)
One to Watch
SPRIND jury pitches, June 24-25
Ten teams will present to SPRIND's jury in Berlin on June 24-25. The pitches are the first public test of whether credible European teams have developed genuine next-architecture AI hypotheses, or whether the challenge attracted incremental work reframed as paradigm shifts. SPRIND's disqualification criteria are specific: no transformer optimization, no model reproductions, no conventional agent architectures. Teams that advance will have had to defend a concrete technical claim about a new scaling dimension against expert reviewers. The jury choices will reveal what Europe's best technical minds believe the next paradigm might be, and who they trust to build it.