🤖 EU AI & Sovereignty Top 10
4 November 2026
Theme of the day: Brussels stopped writing new rules and started executing the ones it has. The leaked 2027 work programme, formally tabled today, puts industrial AI at the centre and quietly drops three digital files from 2018 — including the digital services tax. In the same 24 hours the EU’s top court adviser drew the copyright boundary for AI: training on protected work is broadly fine, but a chatbot may not retell the newspaper. Three of the ten items below come from paywalled Rapporteur reporting; where that is the case, only the headline, lead and metadata are reflected, and the scope is marked as unknown.
industrial policy
🔥 HIGH
Commission puts industrial AI at the centre of its 2027 work programme
The leaked draft of the Commission’s work programme — formally presented today, Wednesday 4 November — wants to “unlock” AI use across the whole economy, with a blueprint for action for industrial AI in five sectors: health, transport, agrifood, advanced manufacturing and defence/space. It follows the June report by AI envoy Jim Hagemann Snabe. Alongside it the Commission wants to strengthen Europe’s position in the AI stack — explicitly on chips, data, compute and “innovative financing instruments” — through a digital transformation package on simplification and copyright, plus a revision of the Digital Decade targets pushed from 2030 to 2027. To be scrapped: the digital services tax directive, the significant digital presence directive and the Financial Data Access regulation. Analysis: this is the first work programme in which AI sits under industrial policy instead of digital policy. “Blueprint for action” is the operative phrase — the ambition moved from legislating models to deploying them in factories.
digital taxation
🔥 HIGH
Brussels scraps the digital services tax and lowers its digital ambition
POLITICO obtained the work programme plus its annexes and describes the wider picture: the Commission is “scaling back its legislative ambitions” and prioritising competitiveness, deregulation and economic security, under the line “the EU must retain the ability to forge its own path”. For tech that means withdrawing the 2018 proposals for a digital services tax and a tax on companies with a significant digital presence — two files that existed precisely to make Big Tech contribute in the EU. Also on the list: a new digital transformation package in mid-2027 (simplification plus copyright), a delay of an EU-wide data retention proposal to “next year”, and a revision of the political advertising rules that earlier pushed Google and Meta to stop selling political ads in the EU. Analysis: a fiscal instrument is the cheapest sovereignty tool there is, because it needs no treaty change and no new agency. Dropping it tells you how much political capital the current Commission is willing to spend on making foreign platforms pay.
copyright
🔥 HIGH
Top court adviser: AI may train on copyrighted work, but chatbots may not retell the newspaper
Advocate General Maciej Szpunar delivered his non-binding opinion on Tuesday in the first AI case before the Court of Justice: using protected works for training is “in principle permitted” under the text-and-data-mining exception. But he keeps a second lock on the door — the press publishers’ right can still be infringed when a chatbot uses press articles to generate an answer, for instance by circumventing paywalls or anti-scraping measures, or when the answer recognisably reproduces the “form of expression” of the article. The operator of the chatbot is liable when the infringement is inherent to the business model or when effective countermeasures are missing. The Court usually follows such an opinion; the judgment is not there yet. Analysis: training is safe, output is not. That single distinction reshapes every RAG pipeline that summarises news — and it lands exactly where the Commission is about to legislate again on copyright.
public infrastructure
📌 MEDIUM
Commission opens a call for a European “AI Media Platform”
The Commission opened a call for pilot projects around “AI-enabled media infrastructures”, intended as modular, multilingual public infrastructure for European news and audiovisual media — AI access without every newsroom having to buy its own American API. The call sits under the “Apply AI” initiative (Horizon Bridging 2027) and runs from 20 October 2026 to the closing date of 16 February 2027. Note: the detail page on digital-strategy redirects to the Funding & Tenders portal, so budget and conditions here come from the RSS metadata and not from the full call text. Analysis: this is the most concrete answer yet to the sovereignty argument in the media sector — not a European ChatGPT, but shared plumbing that newsrooms can plug into without exporting their archives to a hyperscaler.
platform rules
📌 MEDIUM
Digital Fairness Act: 35+ app builders demand the right to take your social network with you
More than 35 builders of alternative social apps — including Eurosky co-founder Sherif Elsayed-Ali and IPFS’s Robin Berjon — are asking the Commission to put a genuine switching right into the Digital Fairness Act, due to be presented on 11 November. Every platform should hold data, identity, connections and interactions in a “personal data server” under one open standard, so users can take their network with them. They also want the UCPD to explicitly ban platforms from suppressing links to alternatives. The sore point: the DMA interoperability rules do not cover all large platforms — X falls outside, which puts blocks on “account matching” beyond the reach of EU enforcement. Analysis: portability of a follower graph is the one lever that would actually dent network effects, and the Commission is being asked to write it into a consumer-protection file rather than a competition file — a faster route, with weaker remedies.
verifiable AI
🔥 HIGH
Anthropic builds cryptographic proof of which model answered
With “FlashProve” Anthropic is working on provable inference: a lab publishes a fingerprint of a model — without releasing the weights — that lets anyone verify afterwards, cryptographically, that an answer came from that model. That makes it hard to quietly substitute a cheaper model or to claim someone else’s benchmark results. The specification and code are open source on GitHub; there are no plans yet to put it into production. Analysis: this is precisely the verifiability that the AI Act’s transparency obligations need in practice — GPAI documentation only means something if a downstream provider cannot misdescribe which model it is reselling. A public fingerprint is a cheaper enforcement mechanism than a regulator with audit rights.
export controls
📌 MEDIUM
Nvidia asks cloud customers to cut back on China business — and hits a gap in the export rules
Nvidia has asked cloud customers, including the neoclouds it finances, through sales and regional teams to diversify their customer mix and take on less Chinese revenue; new contracts must also disclose the end users of the chips. The guidance is not mandatory and does not apply retroactively, but it targets a known hole: US export controls restrict the delivery of hardware, not remote access through datacenters abroad — which is how Chinese AI developers can still reach advanced Nvidia chips. Analysis: a vendor asking its own customers to police end use is what happens when a rule is written for boxes and the product is sold as a service. For Europe it is a reminder that “sovereign compute” is also about who can legally sit on the other end of an API.
supply chain
📌 MEDIUM
Huawei tells customers to source their own HBM; black market asks ~$4,000 per unit
Huawei cannot supply enough High Bandwidth Memory for its own Ascend chips (the 960DT arrives early next year) and is telling customers to arrange the memory themselves — in practice, to go to the black market. Samsung and SK Hynix, together more than 80% of the world market, are not allowed to sell directly, so smugglers bundle HBM with cheap server processors below the export threshold and ship via Hong Kong, where the modules are separated again. Smuggling prices for HBM3E went from a few hundred dollars to roughly $4,000 per unit in a year, and Korean customs had found $560m in violations by May — more than the previous annual record. Analysis: memory, not compute, is now the binding constraint on Chinese AI hardware, and the price signal says the controls are biting while the leakage continues.
venture capital
📌 MEDIUM
Robin Capital closes a €15m second fund for AI, robotics and space
Berlin-based Robin Capital (solo GP Robin Haak) closed Fund II at €15m, taking it to €30m across two early-stage funds, with 17 investments already — including SkySpy.ai, Marble Imaging and Helicity Space (intelligence, defence, earth observation, space). Tickets run from €100k to €600k with a clear weighting towards AI applications, autonomous systems and space infrastructure; a third €15m fund is planned. It is one of a series of European closes this year (Vanagon €20m, Angel Invest €40m, Auxxo €33.3m, Masia €18m) totalling €141-156m of fresh capital for deep-tech and AI — the counterweight to the American compute lead, at fund sizes that are still two orders of magnitude smaller.
inference economics
🔥 HIGH
Revolut cuts AI coding costs 90% by swapping Claude for a self-hosted open-weight model
Revolut rolled AI coding tools out to roughly 2,000 engineers and saw inference costs fall by about 90% after making self-hosted models (GLM-5.1, later GLM-5.3 from China’s Z.ai) the default instead of Claude Code or OpenAI Codex: $0.30 per million tokens, about a tenth of the commercial price, served on 56 Nvidia B200s. The biggest gain was not writing code faster but the disappearance of token budgets — debugging and Jira tickets became “free”. Abitbol nuances the sovereignty question himself: Chinese open weights are mainly risky through APIs running in China; on-prem or in a regional cloud he sees no problem. Analysis: this is the first hard number for European firms weighing the self-hosted stack as both a cost and a dependency strategy — and the answer depends on where the weights run, not where they were written.
📈 Key Trends
- • Execution replaces legislation — The 2027 programme adds almost no new AI law. The sovereignty fight has moved to procurement, compute and financing instruments — the places where a Commission can act without a co-legislator.
- • The fiscal lever is being put down — Scrapping the digital services tax removes the cheapest way to make foreign platforms pay. Expect the “level playing field” argument to be replaced by “attractiveness” in every digital speech from now on.
- • Training is fine, output is not — Szpunar’s opinion splits the copyright question in two. Every RAG or news-summarising product in Europe now has a legal boundary drawn around its answer text, not its training set.
- • Memory is the new choke point — HBM at ~$4,000 on the black market and Nvidia policing cloud end users both say the same thing: the constraint moved from fabs to supply chains and from hardware to access.
- • Cost, not sovereignty, drives the stack — Revolut’s 90% saving is the argument European CTOs will actually bring to their board. Sovereignty sells to governments; unit economics sell to everyone else.
Wat vond je van dit bericht?
