EuroHPC signed a contract on 26 January 2026 to equip MareNostrum 5 in the Barcelona AI Factory with AI-optimised GPUs and storage. Startups, SMEs and sector teams can therefore run pilots on shared supercomputing instead of building their own infrastructure, with training and technical support.
According to the announcement by the European High-Performance Computing Joint Undertaking, this is not only about heavier hardware, but also about supporting smaller organisations with data access, HPC resources, training and technical expertise in sectors such as finance, law, energy, media and the public sector. That makes the upgrade more than a story about computing power.
The programme of AI Summit Barcelona places this in context. The event on 22 and 23 September 2026, according to the official tracks, includes separate tracks for AI For Builders, AI Foundations, AI For Today and AI For Growth, alongside other tracks. In our assessment, this supports the interpretation that the event, besides model development, places emphasis on implementation: how do organisations turn shared compute and expertise into controllable, auditable workflows?
What exactly does the upgrade of MareNostrum 5 within the Barcelona AI Factory involve?
The EuroHPC Joint Undertaking describes the upgrade as the addition of AI-optimised GPUs, accelerated hardware, fast interconnects and large-scale storage to MareNostrum 5, which is hosted and managed by the Barcelona Supercomputing Center. These components appear verbatim in the source.
More important for the reader is the second layer the source mentions: the AI Factory supports startups and SMEs with data access, HPC resources, training and technical expertise. The computing power is the infrastructure; the support determines whether a smaller organisation can actually make use of it.
How does Barcelona fit into the wider European AI Factory network?
Barcelona does not stand alone. The European Commission reported in October 2025 that the network of AI Factories had been expanded to 19 facilities across 16 member states, supported by more than 2.6 billion euros in combined commitments from the EU and participating countries.
In our assessment, this is the point that often drops out of summaries: it is not about one Spanish investment, but about a repeatable model of distributed access. The network character makes shared compute and expertise potentially available as a European option, but the concrete access conditions must be established per AI Factory and per workflow.
What does shared AI supercomputing mean in concrete terms for startups and SMEs?
The practical consequence that the sources imply but do not spell out: smaller organisations can start concrete pilots without building a data centre themselves. That changes the up-front trade-off. Instead of an investment decision about hardware, it becomes a choice about which workflow is suitable for shared compute.
- When shared compute fits: heavy, occasional training or evaluation tasks that do not justify owning hardware.
- When it grates: workflows with sensitive or confidential data that require a bounded environment.
- What the AI Factory adds: training and technical expertise, so smaller teams without their own HPC specialists can get started.
Anyone taking this trade-off seriously must choose between local and cloud AI per workflow and be able to justify that choice. Shared supercomputing is not a default answer, but an option that suits certain tasks.
Which controls must an organisation set up before sensitive data goes through shared compute?
The sources explicitly mention sectors such as law, finance, energy and the public sector. Those are precisely the domains in which data location, traceability and accountability weigh heavily. The EuroHPC announcement describes the support; on setting up controls the source says nothing. That responsibility lies with the using organisation.
In our assessment, the usable test is operational. Before sensitive information enters a shared environment, an organisation should be able to answer the following questions concretely:
- Which data does the workflow use, and is it traceable to its source?
- Does sensitive information stay within an appropriate environment, or does it leave the controlled context?
- Are evaluation results reproducible, so an outcome can be accounted for later?
- Can production behaviour be monitored after the pilot goes into use?
- Who carries the human responsibility for the final decision?
The fourth and fifth questions deserve extra attention. Anyone who later has to account for an AI-supported decision does well to record human oversight of AI decisions demonstrably. And in workflows involving personal data it is unavoidable to keep personal data in embeddings controllable, even when it persists in derived form.
What can you do now to prepare a sector pilot?
The concrete step is modest and testable. Select one bounded workflow, determine whether it is suitable for shared compute, and record the five control questions above in writing beforehand. Only then does it make sense to get in touch with an AI Factory about data access and support.
Anyone who approaches this systematically places the pilot within a broader practice of applying AI in professional practice, in which infrastructure, data minimisation and accountability come together. The contract for MareNostrum 5 lowers the threshold to start; it does not shift the responsibility for controllable workflows. That, even with shared compute, remains with the using organisation.
Sources and references
Sources: The article draws on the EuroHPC Joint Undertaking, the European Commission and the official programme of AI Summit Barcelona.