On 21 October 2026, Italian humanoid developer Generative Bionics announced at CES Unveiled in Milan that it is working with AMD to build what it calls Europe's largest "physical AI data factory". Silicon Republic reports that the first 5,000 sq m hub in Italy is due to be operational by the end of 2027. A second site is planned "elsewhere" in Europe and a third will serve global markets. AMD's Helios rack-scale AI infrastructure will power the data network.
The stated aim is "humanoid sovereignty": ensuring that the data, models and know-how behind Europe's industrial humanoids are generated, trained and validated on the continent. The backdrop is lopsided. The same report cites an analysis putting China at more than 97% of the humanoid market, with Agibot and Unitree together accounting for about 75% of sales.
The case for treating this as a policy matter
The strongest argument for state involvement is that physical AI has a data problem language models did not have. Text can be scraped. Robot training data has to be generated through teleoperation, simulation and real-world trials, and that takes capital and floor space. If a handful of foreign firms hold the datasets that define industrial robot behaviour, European manufacturers may end up as customers of models they cannot inspect or adapt. A government that wants an alternative has a reasonable case for public money and a coherent plan.
Italy has tried to supply both. Law 132/2025 (Legge 23 settembre 2025, n. 132, "Disposizioni e deleghe al Governo in materia di intelligenza artificiale") was published in the Gazzetta Ufficiale on 25 September 2025 and entered into force on 10 October 2025. According to A&O Shearman's analysis, it introduces a national AI strategy that the Interministerial Committee for Digital Transition updates every two years. AgID and the national cybersecurity agency ACN serve as the national authorities. The OECD.AI policy navigator says the law authorises up to EUR 1 billion in state-backed venture capital for AI, cybersecurity and telecommunications companies.
What the law does, and what it does not
The EUR 1 billion is venture capital, not a robotics subsidy, and it is not a bill for any single project. Generative Bionics' own funding shows how the model works in practice. It launched in late 2025 with a EUR 70 million seed round backed by CDP Venture Capital, with AMD Ventures, Eni Next, Duferco, RoboIT and Tether also taking part. That is the pattern policy should favour: state-linked capital that sits alongside private investors and does not replace them.
The regulatory side of Law 132/2025 is less obviously helpful to a robotics start-up. A&O Shearman describes a law that adds organisational and procedural requirements, including transparency, controls, training and documentation, that reach beyond high-risk cases into sensitive sectors such as work, healthcare and justice. It also takes a more prescriptive line on generative AI and deepfakes, with criminal offences and traceability duties, than the EU AI Act, which relies mainly on information duties and codes of conduct. Healthcare is one of the sectors Generative Bionics names as a target market.
This is the gap between a national AI strategy and an industrial strategy. A strategy document refreshed every two years can set priorities. It cannot supply the thing a data factory runs on, which is a large volume of high-quality embodied data, collected under rules that let firms use it.
Where the EU framework matters more than the Italian one
For industrial humanoids, the binding constraints are more likely to come from Brussels than Rome. The European Commission's AI Act page says the Act became generally applicable on 2 August 2026. Annex III high-risk use cases, such as employment, have been pushed back to 2 December 2027. High-risk systems embedded in regulated products, which would include many machines, have an extended transition to 2 August 2028. The AI Omnibus that made those changes entered into force on 27 July 2026.
That timetable lines up with the factory's own. The first hub opens at about the time the high-risk obligations start to bite. Those obligations cover dataset quality, activity logging, technical documentation and human oversight. A data factory that builds these into its pipelines from the start could turn compliance into an advantage over rivals that must retrofit it. That is the best version of the sovereignty argument, and it rests on engineering, not protectionism.
The risk is that member states layer their own requirements on top. Law 132/2025 states that it is to be read in conformity with the EU regulation. Even so, every national add-on raises the cost for a company that wants to operate in a second European site and then a global one, which is exactly the three-site plan Generative Bionics has announced.
A proportionate agenda
Three things would help more than another strategy revision.
- Data access with clear rules. Make it simple to run real-world trials in factories, hospitals and warehouses, with predictable safety and privacy requirements, so the data factory fills with real data and not only simulation.
- Capital that follows private lead investors. Use the venture-capital authorisation for co-investment alongside private money, and publish how it is being deployed so outcomes can be judged.
- One rulebook. Resist national additions to the AI Act where the EU text already covers the risk, and keep the transition dates the Omnibus set.
The 97% figure is a reminder of how far behind Europe is, but not a reason to wall off the market. Europe's strongest position is to be the place where industrial robots are trained on trustworthy, well-documented data and can be sold anywhere. A 5,000 sq m hub is a modest start. Whether it becomes more depends on capital, compute and permission to collect data, which no strategy document supplies on its own.