Barcelona AI Factory Startups Put MareNostrum 5 to Work

Barcelona’s AI Factory is turning supercomputing access into a practical resource for startups, pairing compute capacity with technical support and training. The programme at the Barcelona Supercomputing Center (BSC) has moved beyond its pilot phase and is now operating at full scale. Its early cohort shows how companies are applying that support across fields where specialized data and operational constraints shape what an AI system can do.

BSC AI Factory is among Europe’s first AI factories and the first in Spain. Since launching in April 2025, it has delivered more than 800 services to private and public-sector clients. Those services sit within a wider effort to make advanced computing available for scientific and technological work, while helping companies develop and test their own applications.

Compute access paired with hands-on support

The accelerator programme provides startups with access to sovereign compute, technical assistance and training. The mix matters: access to computing infrastructure can help a team run demanding workloads, while expert support can help it adapt tools and workflows to a specific application. The source does not report how much compute each startup receives or how the programme measures results, so the 800 services figure describes activity rather than a quantified impact on products or customers.

An upcoming extension of MareNostrum 5, BSC’s supercomputing infrastructure powered by NVIDIA accelerated computing, is expected to provide access through the BSC AI Factory and other AI factories across Europe. That planned expansion could give more teams a route to supercomputing resources. The article does not specify the extension’s capacity, timing or allocation process, leaving those details open.

Some participating companies also belong to NVIDIA Inception. Services from BSC AI Factory are now available through the Inception portal. The programmes offer different forms of support: BSC supplies access to its accelerator and computing resources, while Inception provides member training through NVIDIA Deep Learning Institute open courses, business mentoring, venture capital connections and other resources.

Three projects recognized at the pilot demo day

Three Inception members were named AI Champions at the accelerator’s pilot initiative demo day: Qatium, OneCareAI and Avatia. Their work illustrates a range of development paths, from building operational models for utility networks to using AI in clinical analysis and pharmaceutical information services.

Qatium develops a platform for water networks that is already used by more than 250 utilities in 16 countries. Its Neural Water Intelligence technology is in early production and uses proprietary AI models trained on MareNostrum 5. The system is built with NVIDIA accelerated computing and draws on NVIDIA cuGraph, the NVIDIA PhysicsNeMo framework and NVIDIA Nemotron open models. Qatium says the technology helps utilities make operational decisions. The available description does not report accuracy, savings or other measured results, so the number of utilities using its platform provides the clearest indication of current reach.

OneCareAI, a research spin-off from BSC, used NVIDIA NIM microservices to develop technology intended to identify signs of hidden atrial fibrillation in ECGs that appear normal. Atrial fibrillation is an irregular heart rhythm associated with stroke risk. The stated aim is to help clinicians identify potentially at-risk patients before a stroke occurs. That describes a clinical use case, not evidence of clinical effectiveness: no study results, detection rates or deployment details are included in the source.

Avatia built a voice-first system with multiple AI agents for pharmaceutical companies. The agents are designed to support sales representatives and provide doctors with transparent, curated medicine information, with the goal of helping appropriate treatments reach patients sooner. Avatia’s system uses NVIDIA NIM microservices, the NVIDIA Nemotron 3 open model family and the open source NVIDIA NeMo Curator library. The source does not describe how the system checks its information or measures whether it changes treatment decisions; those will be important details for assessing its use in clinical settings.

A varied cohort, with results still to establish

The wider group includes Agrikola AI, Elna Health, Lokutor, Montevive AI, Oppy, Pack2Earth and Up2Smart. Their areas include precision agriculture, medical imaging and knowledge access, speech services, enterprise security, sales intelligence and sustainable packaging. The reported applications range from detecting breast cancer subtypes before biopsy and helping doctors find medical knowledge more quickly to using robots to protect crops and developing biobased compostable materials as alternatives to petroleum-based plastics.

That range makes the accelerator’s role easier to see: it gives teams working on distinct problems a shared route to compute, tools and technical help. But a list of applications does not establish performance or readiness. For developers evaluating the programme, the next useful evidence will be how these projects perform in real workflows, what compute and data they require, and how teams handle accuracy and reliability in high-stakes settings. BSC’s planned MareNostrum 5 extension could widen access; the startups’ deployment results will show what that access enables.

Learn more about NVIDIA Inception and the BSC AI Factory Accelerator Programme.