The AI Airbus: the instrument already exists, what is missing is the buyer

In early September Fritzi Köhler-Geib, a member of the Executive Board of the Deutsche Bundesbank, addressed the International Bankers Forum in Hamburg on Europe’s capacity to act. According to the published manuscript she mentioned “the idea of an ‘AI Airbus’, a joint European large-scale project in artificial intelligence”. The headlines read that as a fresh demand. The sentence, however, quotes a paper that Fritzi Köhler-Geib published on 2 July 2026 together with Claudia Plattner, President of the Federal Office for Information Security (BSI), and Kristina Sinemus, Hesse’s Minister for Digitalisation and Innovation: 37 measures across five technology fields, public since the summer.

Anyone who reduces the paper to the Airbus has read the headline and skipped the table. Its central proposal is the pooling of public demand, and that instrument is already before the EU legislator: the Cloud and AI Development Act (CADA), which the Commission tabled on 3 June 2026, provides in Article 37 that the Commission may procure cloud and AI services for the contracting authorities of the Member States as well. What is new in the paper is the buyer it has in mind, the Eurosystem, rather than the instrument. The AI Airbus itself is one of two “procedural alternatives”, and the authors have scored the feasibility of each of its steps themselves: 5, 4, 3, 3, 2 and 2 on a scale of 1 to 5. Whether the venture succeeds is not something this reading can settle, and it does not predict failure either. It reads what the authors wrote into their own paper and sets it beside what the EU has already set in motion.

The speech repeats what has been on the table since July

The paper is dated 30 June 2026 and has been public since 2 July, on the Bundesbank’s website and in a LinkedIn post by Fritzi Köhler-Geib that same morning, which also carries the actual demand: “Heute braucht es erneut ein politisch unterstütztes Konsortium von 2-4 Unternehmen aus verschiedenen europäischen Ländern im KI-Bereich”, a politically backed consortium of two to four companies from different European countries. The speech of 3 September adds nothing to that; it names the pooling of public demand and the idea in one sentence, alongside the Pontes initial launch due later this year, which an earlier article covers. The paper itself keeps its ambitions modest: according to the summary issued by the Hessian digital ministry it is “not designed as a government programme or a list of demands addressed to the federal government and the European Union”, and the Airbus idea, Fritzi Köhler-Geib says, came out of a conversation with Jean-Claude Trichet. The analogy, admittedly, is older than the paper. Peter Altmaier, then head of the Chancellery, said in January 2018, as reported by Reuters and quoted by ECIPE, “Wir brauchen eine Art Airbus für KI”, a kind of Airbus for AI, and Gaia-X launched in 2019 with the Airbus consortium as its organisational model. Yet the three authors take the analogy further than their predecessors did, with six steps and a score for each. That table is precisely what the public reading skips.

The authors grade their own roadmap

In the field of artificial intelligence the paper lists the measure “build a second frontier LLM for the EU and Germany”, time horizon “medium term”, and offers two procedural alternatives. The first is called “KI-Airbus”, described as building a company like Airbus for AI, and receives an overall feasibility of 3 and an impact of 5, each on a scale of 1 to 5 where, by the paper’s own legend, 1 means “hard” and 5 means “easy”. Beneath it sit six steps, each with its own score: identify two to four countries with LLM experience, the examples given being France, Germany, Italy and Finland (5); each “founding country” contributes an experienced company (4); the founding countries contribute the founding capital (3); the contributed companies found the AI Airbus (3); the EU awards it a paid contract to develop a European language model (2); the AI Airbus receives a privileged position in public procurement (2).

The second alternative, “privileged companies in a special zone”, means supranational companies equipped with privileges; its overall score is feasibility 1, impact 5, and the step of creating a dedicated EU legal form for them carries the 1. What both variants lack becomes obvious on reading: no amount, no companies, no legal form. A self-assessment is admittedly not a third-party review, and a five-point scale is a blunt instrument. Yet the authors are a central bank board member, the head of a federal agency and a state minister; anyone who knows how a paid EU contract comes about has documented the bottleneck here more honestly than any critic. Founding the company earns a 3. Awarding it the contract earns a 2.

The instrument already exists, and it is called the anchor customer

The procurement part of the paper is more concrete than the Airbus. In the cloud field it calls for a “framework contract of the EU Commission from which all public institutions can draw services from the top 3 to 5 EU-native cloud providers”, feasibility 3, impact 5, with the step “pool the demand of the Eurosystem in EU-native cloud”. Right beside it stands a second framework contract for the “top 2 to 3 US cloud providers”, feasibility 3, impact 4, with the same pooling step. The paper runs on two tracks; a buy-European paper it is not. On procurement law it names as one step Germany’s procurement acceleration act, which the paper wanted to see take effect, as it did, on 1 July 2026; according to the procurement office of North Rhine-Westphalia the act has applied since exactly that day. That step was already done when the paper appeared.

The European framework has been on the table since 3 June 2026. On that day the Commission presented proposal COM(2026) 502 for the Cloud and AI Development Act; its policy page names as one aim “establishing a common EU-level procurement framework for public administrations to leverage their joint purchasing power”, and Article 37 of the draft reads: “The Commission may carry out procurement activities to procure data centre services, cloud computing services, software and AI systems for itself and for Union entities and for contracting authorities of Member States”. The draft also defines four sovereignty levels, up to “full transparency and control over their software supply chain and no interference from a third country”. In Parliament, according to the Legislative Observatory, Reinier Van Lanschot and Diego Solier have been rapporteurs since July and August; no Council position is documented in the primary sources. The anchor customer mechanism is already in a live call elsewhere: the EuroHPC Joint Undertaking opened its call for up to seven AI gigafactories on 30 July 2026 and writes that the public funding “will act as an anchor customer”, expected to unlock more than €20bn of private investment; the deadline is 12 November 2026, with selection due in early 2027. Regarding the gigafactories, the paper notes: “Measure is already being implemented by the EU.” The EU is, admittedly, procuring compute rather than models, and the CADA is a proposal rather than a law. Yet what the paper demands for the AI Airbus in steps 5 and 6 already exists as a mechanism in both procedures. The open question is who signs the framework contract, on what criteria, and whether the Eurosystem joins as a buyer, rather than whether Europe builds an Airbus.

Two papers, one instrument, two endpoints

Seven days after the paper, on 9 July 2026, the Monopolies Commission delivered its 26th biennial report, and its foreword reads like a rebuttal: “Keine Subvention, kein nationaler Champion ersetzt, was ein funktionierender Markt leistet”, no subsidy and no national champion can replace what a functioning market delivers. Its chairman Tomaso Duso told the Tagesspiegel: “Strukturreform schlägt Subvention”, structural reform beats subsidy. On the instrument, however, the commission stands beside the paper. In paragraphs 904 to 906 the report distinguishes three procurement variants: pooling demand without geographic preference; the state as anchor customer with a European preference, tied to “objective, verifiable and substantively justified criteria”; and a general European preference, which it rejects. Its recommendation chooses the middle one, “limited to clearly diagnosed transformation failure and existing dependencies”. The dividing line is surgically precise: both papers want the state as a buyer; the sovereignty paper moves on to founding a company; the report stops at procurement. The cut falls at steps 4 to 6 of the Airbus roadmap, and that is exactly where the authors entered 3, 2 and 2. Their own scale drops at the boundary where the Monopolies Commission gets off.

The history behind the analogy helps more than the paper gives it credit for. Airbus began, according to a case study by the US institute American Compass, with an intergovernmental agreement in July 1967 between France, Britain and West Germany; the British government withdrew, while the British manufacturer Hawker Siddeley stayed on board with the wings. Airbus Industrie was founded, according to Airbus, on 18 December 1970 as a groupement d’intérêt économique under French law, with SNIAS, Deutsche Airbus, Hawker Siddeley and VFW-Fokker as partners; the “Franco-German consortium” of the paper was three-nation from the outset. Its first operating profit came, by the same case study, in 1990, “20 years after its launch”; by 2018 the state launch aid had, on an estimate by the industry-aligned US institute ITIF, exceeded $22bn, and the dispute over it before the World Trade Organization lasted 17 years. Matthew Kilcoyne of the Center for Data Innovation, an ITIF body that argues for a “Western, allied digital industrial system”, holds those two decades against the model’s advocates. Since the 1990s Airbus has been, as the paper rightly says, an established force: for 2025 Airbus SE reported, in its release dated 19 February 2026, revenue of €73.4bn and 793 deliveries against 600 at Boeing, which, according to CNBC, nevertheless booked 1,173 net orders against Airbus’s 889, outselling it for the first time since 2018. The model is a duopoly that still fights for every order after five decades.

Two differences from the model matter more than the number of years. First, Airbus sold to solvent airlines that had no free alternative; a European frontier model competes with free base versions and openly published weights, which is an assessment rather than a measured quantity. Second, the two European candidates answered the consortium question before the paper appeared: Mistral was not founded by states but recapitalised privately on 9 September 2025 with €1.7bn, with ASML as its largest shareholder; Aleph Alpha, the German candidate, went to Canada’s Cohere on 24 April 2026, with the support of the federal government according to Reuters. The notion of sovereignty behind the Mistral case is set out in “Sovereignty is not an address”. And the model already sits in a coalition with the candidate: Airbus has, by its own account, been a member of the “European Tech Creators” since May 2026, together with ASML, Ericsson, Mistral, Nokia, SAP and Siemens. No company had to be founded.

The diagnosis holds, the number has four reference bases

The paper’s dependency diagnosis appears on page 2: “Stand 2025 importieren wir in Europa 80 % der IT-Infrastruktur und Technologien von nicht-europäischen Anbietern”, 80% of IT infrastructure and technologies imported from non-European providers, with footnote 1 pointing to the EuroStack report of the Bertelsmann Stiftung of 13 February 2025. That report states: “More than 80% of Europe’s digital technologies and infrastructures are imported.” The Draghi report states on page 59 of its Part A: “The EU relies on foreign countries for over 80% of digital products, services, infrastructure and intellectual property”, and refers to the Commission’s Digital Decade report of 27 September 2023, where the sentence first appears without a methodology footnote. In the speech of 3 September it becomes “over 80% of digital infrastructure and technologies”, without a footnote. Four stations, three reference bases, one digit, and the “as of 2025” is the publication year of the cited source, not a measurement year. That is no reproach to the authors, who did set the footnote, and nobody disputes the order of magnitude. The speech’s second number rests on firmer ground, namely that in 2022 “almost two thirds of card payments in the euro area” ran through non-European providers: the ECB report of 28 February 2025 measures “approximately 61%” and records that 13 of the then 20 euro area countries rely entirely on international card schemes. The card question belongs to the articles on the digital euro; here it explains why the Bundesbank argues from its own exposure as a buyer. A number that narrows with every retelling serves as a direction. As a reference base for a procurement decision it does not.

For banks the critical infrastructure line matters, not the Airbus

The most concrete line of the paper for institutions sits in the cloud field: tighten regulation, for instance for critical infrastructure (KRITIS), “so that the use of sovereign providers can be made partly mandatory”, feasibility 4, impact 3. What remains open is whether “partly mandatory” means infrastructure, models or both, and who decides which provider counts as sovereign; the vocabulary for that comes from the CADA draft and the BSI’s C3A criteria. For an institution that has aligned its cloud contracts with the Digital Operational Resilience Act (DORA) this is a second axis: so far supervisors ask about concentration and exit capability, while the paper’s measure asks about origin. The instrument against concentration does, admittedly, exist already and it is called oversight: the European supervisory authorities published the list of ICT third-party providers under direct DORA oversight on 18 November 2025, 19 providers including AWS EMEA, Google Cloud EMEA, Microsoft Ireland and Oracle Nederland, but also SAP, Deutsche Telekom, Capgemini, Colt and Orange; five of them have a European parent.

Two German houses show how far apart the answers lie. Deutsche Bank presented itself, in its release dated 25 August 2026, as “key design partner” for the Financial Research Agent in Google’s Gemini Enterprise for Financial Services; the supervisory side of that partnership is covered in “When the agent supervises the trader”. Finanz Informatik, by its own account, runs “a 100% on-premise AI platform in its own certified and highly available data centres”, with open-source models on NVIDIA processors; the platform logic behind that is described in “The platform lever”. Both paths end at the same boundary, and the paper concedes it on chips: it is “unlikely that any state will be able to build a fully independent supply chain in the medium or long term”. That leaves the dimension the paper lacks: who pays, and how much. For the AI Airbus it names no amount; the reference figures lie right beside it, €1.7bn for one Mistral round, more than €20bn of expected private money for the gigafactories, more than $22bn of launch aid. A discussion paper is certainly not a budget. Yet whoever draws the Airbus comparison draws its financing with it, and the score of 2 for the paid EU contract is the only place where that calculation appears in the paper.

Recommendations

Glossary