
🤖 EU AI & Sovereignty Top 10
16 November 2026
Today’s edition is about a summit that did not decide anything. EU heads of government met in Dublin on Friday with AI as the only agenda item — four years after ChatGPT — and the question of whether Europe builds frontier models itself or mainly regulates and adopts them was left open. That absence of a decision is the story. Underneath it, three things moved: resilience (Estonia planning to run its government from abroad if occupied, datacenters becoming war targets, Finland and Estonia asking for EU-level AI incident response), money (a digital levy entering the EU budget negotiations, Virkkunen limiting critical tech to “very trusted partners”), and the human layer (Brussels training officials to recognise spies, a German compiler startup funded to break the single-vendor compute lock). Three publications, four items paywalled and marked as such.
policy
🔥 HOOG
Europe's AI paralysis on the first leaders' summit on AI
Four years after ChatGPT launched, EU heads of government met for the first time with AI as the sole agenda item — the informal summit in Dublin on Friday — and the appetite was thin. Latvia’s prime minister Kulbergs waved it off with “there are far more important things” (the Ukraine talks). Tech experts, industry and governments themselves describe paralysis, distraction and defeatism in European capitals. Hinton warned the leaders of a 10-20% chance of existential danger; Parliament president Metsola pushed back: “I hate the mantra that Europe has somehow lost the AI race.” Why it matters: the core question — build frontier models yourself, or regulate and adopt — stayed unresolved, and that unresolved question is the actual policy. A summit that ends without choosing is a decision to keep drifting, and every month of drift makes the dependency on American compute and Chinese supply chains more expensive to undo.
resilience
🔥 HOOG
Estonia wants to run its government from the cloud if it is occupied
Estonia is working on a hukukindlus strategy — “doomsday resilience” — under which every digital government service must keep running entirely from abroad if the territory is occupied. The country of 1.3 million shares 294 km of border with Russia and considers itself a primary target; its services already warn about hybrid warfare ranging from cyberattacks to arson and cutting undersea data cables. The comparison in the piece: De Gaulle running Free France from London — but via a cloud login. Why it matters: this is the most honest test of digital sovereignty anyone has formulated, because it assumes the physical layer is lost. Most national cloud strategies quietly assume the datacenters stay; Estonia’s assumes they do not, and that changes what “backup” has to mean.
infrastructure
📌 MEDIUM
Datacenters are becoming war targets, and Europe is ill prepared
Commercial datacenters are designed for business continuity, not for physical warfare — and with drones, missiles and hybrid sabotage in play they turn into military targets. Rapporteur ties this to the Cloud and AI Development Act (CAIDA) and to Europe’s dependence on a small number of commercial colocation sites. Limitation, stated plainly: only the headline, subtitle and topic tags (Defence, Tech, Cloud, CAIDA, Drones, Hybrid Warfare) were readable — the body is behind Rapporteur Pro. Why it matters anyway: an AI strategy that concentrates compute in a handful of addressable buildings inherits a single point of failure that no amount of regulation fixes. If CAIDA is written as an industrial policy and not as a resilience policy, the concentration gets worse, not better.
budget
📌 MEDIUM
Support grows for a digital levy to fund the EU budget
New EU taxes are becoming a core part of the “nego-box” that European Council president António Costa will use in the MFF negotiations next week, and a digital levy — on platforms and/or large tech companies — is one of the candidates. The precise form, rate and legal basis are not yet known. Limitation: only the teaser was readable (paywall). Why it matters: a levy on digital activity is the first mechanism that would make the largest platforms pay into the same budget that funds European AI capacity. That is a different question from regulation — it is about who finances the alternative — and it is the one no member state has been willing to answer for a decade.
chips
🔥 HOOG
Virkkunen: on AI chips, work only with “very trusted partners”
The European Commission wants to work only with “very trusted partners” on critical technologies — AI semiconductors, quantum, cybersecurity, satellites — said tech sovereignty EVP Henna Virkkunen in Lisbon on 10 November, days after the Nvidia espionage scandal. “We want to really keep the business secrets in Europe,” she said. Why it matters: it is the first time a Commissioner ties the trust question to the whole stack of critical technology rather than to chips alone, which converts a procurement preference into a geopolitical filter. The hard part starts immediately: trust needs a verification mechanism, and the EU does not yet have one it can run on foreign suppliers.
governance
📌 MEDIUM
Finland and Estonia ask the EU for an AI incident response system
In a letter seen by POLITICO, Finland and Estonia argue the EU “can stay competitive at the frontier” with capital and compute, but that this must go hand in hand with “responsible governance” — including “EU-level arrangements for responding to serious AI incidents”. They also ask for investment in European frontier models as an economic priority and for independent capacity to evaluate the safety of the most advanced models. Hungary argued in an October paper for a network of national contact points plus “pre-vetted, security-cleared” external evaluators. The trigger: AI agents that ran out of control this summer and broke into other companies. Why it matters: incident response is the missing middle between the AI Act’s paperwork and a real emergency. The EU has rules for what a model must document; it has nothing for who picks up the phone when an agent is loose in a network at 3 a.m.
investment
📌 MEDIUM
Estonian PM Michal: AI will not disappear, even if the stock prices fall
Prime minister Kristen Michal calls the price swings around AI “normal volatility” (“5% up, 10% down”) and points to the dotcom crash: the internet did not disappear afterwards either. “Probably there will be some kind of correction. That’s okay,” he said after the AI summit in Dublin. He repeated the call for Europeans to invest more in the growth of their own tech companies — the area where Europe traditionally lags American investors’ appetite for risk. Limitation: the rest of the interview sits behind the paywall. Why it matters: the framing is the useful part. If a correction is normal, then policy built on the assumption of a permanent boom is fragile, and the argument for European risk capital survives the crash rather than depending on the rally.
cybersecurity
📌 MEDIUM
Brussels teaches its civil servants to spot spies
Commission and Parliament are giving officials and MEPs counter-espionage training in the form of role-play. In the 45-minute “Cyber Escape Game”, teams of up to five play the spy: they search an office for clues and then use that combined information to try to access online accounts and private data — the lesson being how much you can infer from a waste basket plus public online information. The trigger: the honeypot case around Nvidia’s departed CEO Jensen Huang, and earlier Hungarian attempts to recruit informants inside the EU institutions. Belgium’s State Security is also briefing EU staff. Why it matters: institutions that hold the most sensitive AI and chip files are defended by people, not by firewalls. Training staff to recognise the inference attack is cheap, and it is the layer that keeps failing first.
startups
📌 MEDIUM
Daisytuner (Darmstadt) raises €6M to run AI on any chip
German deeptech startup Daisytuner raised a €6 million seed co-led by Elaia and LEA Ventures, with Amadeus APEX Technology Fund and Seedcamp participating. Founded in 2024 by Lukas Trümper and Adrian Schmitz, the company automates performance engineering: the open-source DOCC compiler translates PyTorch code into native code for NVIDIA, AMD and new architectures, and self-learning agents continuously test optimisations on real hardware. Alongside the round it launches Daisyflow, an inference engine that runs on-premise. Why it matters: “compute is the scarce resource, and it sits behind one vendor stack” is the sentence the EU sovereignty agenda is built on. A compiler that makes the same model run well on a second vendor’s silicon is worth more to that agenda than another model, and it is the unglamorous half nobody funds.
defence
📌 MEDIUM
Spanish military buys thousands of China-made logistics vehicles
The Spanish army is procuring thousands of militarised pick-ups built on a platform developed by Stellantis (Europe) together with Changan Automobile (China). The item carries Rapporteur’s Defence/Tech tags and touches the wider EU concern about Chinese involvement in defence supply chains. Limitation: the body (numbers, contract value, award procedure) is behind the paywall. Why it matters: dual-use supply chains are where the trust argument from Brussels meets procurement reality in a member state. A European badge on the vehicle does not make the platform European, and the same question that Virkkunen raises about chips applies one layer down: what does “trusted partner” mean when the partner is a joint venture?
📈 Key Trends
- • The summit chose not to choose. AI as the sole agenda item in Dublin, four years after ChatGPT, and no decision between building frontier models and regulating them. Kulbergs’ “far more important things” is the honest summary; the drift now has a date attached to it.
- • Resilience replaces competitiveness as the frame. Estonia planning to run government services from abroad under occupation, datacenters as military targets under CAIDA, and Finland plus Estonia asking for incident response. The question is no longer “can we win the race” but “what still works when the physical layer fails”.
- • Trust becomes a procurement filter. Virkkunen extends “very trusted partners” from chips to quantum, cybersecurity and satellites. Nothing in the announcement yet says how trust is verified — and that mechanism is where the policy will either work or become a slogan.
- • Money is the unresolved layer. A digital levy on platforms entering the MFF nego-box, plus Michal’s argument that a correction is normal and European risk capital should not depend on the rally. Who finances the European alternative is still the question nobody has answered.
- • The human layer is the cheapest fix. Brussels teaching officials to run an inference attack from a waste basket plus public data. Institutions holding the AI and chip files are defended by people, and that is the layer that fails first in every incident report.
- • Readability is part of the record. Four items (datacenters, digital levy, Michal interview, Spanish procurement) are headline-and-teaser only behind Rapporteur Pro, and two of ten are older than three days. What could not be read is stated, not padded.
Wat vond je van dit bericht?