Building clinical AI on Europe’s supercomputing infrastructure

How does Elna build a trustworthy medical knowledge base at scale?

Every answer Elna gives a clinician rests on a knowledge base underneath it — the medical guidelines, drug formularies, and evidence from scientific papers it draws from and cites. Building a trustworthy medical knowledge base at European scale requires serious computing power. This summer, we got access to some of the most powerful compute infrastructure in Europe to do exactly that: the MareNostrum 5 supercomputer.

What is the BSC AI Factory Acceleration Programme, and how was Elna selected?

The BSC AI Factory Acceleration Programme is an intensive, two-month initiative by the Barcelona Supercomputing Center (BSC) designed to fast-track 15 deep-tech AI startups from research to market. We were one of 15 startups chosen from a pool of more than 100 applicants, to be part of the May to July 2026 cohort. It’s a competitive program, and being selected meant standing up to scrutiny on the same dimensions investors and hospital partners look at: technical maturity, team, and the relevance of what we’re building.

What did Elna do with MareNostrum 5?

As part of the program, we were given compute access to MareNostrum 5, one of Europe’s most powerful supercomputers, hosted at BSC under the EuroHPC initiative. We used it to annotate over 20 million evidence-based medical knowledge sources — structuring and enriching medical literature, guidelines, and formulary data at a scale that would have been cost-prohibitive for a company at our stage on commercial cloud infrastructure otherwise. That work now directly underpins Elna Core, the knowledge engine behind every answer Elna gives.

How does the annotation pipeline actually work?

Under the hood, that means running multiple open-source language models side by side — including models from the Qwen, Llama, and Mistral families, served through vLLM — with each model independently annotating the same source material. Rather than trusting any single model’s output, we reconcile the different models’ annotations through automated fuzzy consensus voting, flagging the cases where they disagree for further review. Quantization techniques let us fit a 32-billion-parameter model onto a single GPU alongside the smaller models in the ensemble, so all three could run concurrently on MareNostrum 5’s NVIDIA H100 nodes rather than in sequence — in testing, that setup sustained close to 100 documents annotated per second on a single GPU. It’s the same principle — multiple models checking each other rather than relying on one — that underpins how we think about trust in the product itself.

Why does it matter that this runs on European infrastructure?

It also matters for a second reason beyond scale: where that compute actually runs. Keeping clinical data and model training on European-sovereign infrastructure — rather than US-hosted, generic cloud — reinforces the data-residency and trust guarantees European hospitals need, and it’s a deliberate part of how we’ve built the company, not an afterthought bolted on for compliance.

What’s next for Elna Core?

This access also opens the door to deeper work ahead: benchmarking open-source language models against our own clinical evaluation sets, and fine-tuning smaller, specialist models on the annotations we’ve produced — distilling the judgment of today’s multi-model ensemble into faster, cheaper models purpose-built for clinical annotation, on that same European-sovereign infrastructure. It’s part of how we’re building Elna’s knowledge engine to be purpose-built for clinical medicine, not just powered by general-purpose AI.

The progress we made over the course of the program earned us a further mark of confidence: we were shortlisted among the pitch finalists for the Demo Day within the BSC AI Factory Acceleration Programme, based on what we’d built and shown by that point.

Barcelona Supercomputing Center’s support here — non-dilutive, and requiring no equity — is the kind of infrastructure access that lets a small team punch well above its size. We’re grateful for it, and for what it’s let us build.