Can Europe build its own foundation models for medicine, or is it permanently reliant on US Big Tech?
- Nelson Advisors
- 3 hours ago
- 13 min read
Evaluating Europe’s Capacity for Medical AI Autonomy
The global paradigm shift toward generative artificial intelligence and foundation models has transformed healthcare research, diagnostic radiology, digital pathology and drug discovery. However, this technological frontier has intensified a critical strategic vulnerability for Europe: a profound structural reliance on United States technology conglomerates for computational hardware, cloud infrastructure, and frontier foundation models. While American technology leaders deploy multi-billion-dollar clusters to train trillion-parameter multi-modal systems, European policymakers and healthcare institutions are navigating a complex operational trilemma. This trilemma pits the mandate for technological data sovereignty and strict regulatory oversight against the operational necessity of accessing state-of-the-art diagnostic and reasoning capabilities.
Whether Europe can construct and maintain its own sovereign foundation models for medicine, or if it is structurally locked into permanent reliance on US Big Tech, depends on a nuanced dynamic. Europe cannot win a brute force competition in general purpose large language model (LLM) compute against US hyper scalers.
However, an alternative technological trajectory is emerging. By leveraging dense, structured clinical registries, unified health data frameworks, specialised biology-native architectures, and open weight model strategies, European research ecosystems are carving out a defensible niche in clinical and biological AI.
The Infrastructure Trilemma: Compute Capacity and Hyperscaler Dependency
A fundamental obstacle to European AI sovereignty is the structural disparity in computational infrastructure. Training frontier medical foundation models, which integrate clinical natural language, high-resolution diagnostic imaging, spatial transcriptomics and whole genome sequencing, requires massive graphics processing unit (GPU) clusters operating with high-interconnect bandwidth.
The European Union has sought to bridge this gap through public investments led by the European High Performance Computing Joint Undertaking (EuroHPC JU) and its "AI Factories" initiative. Flagship installations such as JUPITER in Germany, LUMI in Finland, Leonardo in Italy, and MareNostrum 5 in Spain provide world-class supercomputing power. JUPITER, for example, incorporates approximately 24,000 NVIDIA GH200 Grace Hopper Superchips, delivering up to 90 FP8 ExaFLOPS of AI compute performance. Aggregate EuroHPC public compute capacity spans roughly 57,000 high-end accelerators.
Despite these public investments, Europe accounts for approximately 4.8% of global GPU cluster performance, compared to 74.5% hosted within the United States.
A single American technology corporation, Meta, deployed an infrastructure footprint equivalent to nearly 600,000 NVIDIA H100 GPUs—more than ten times the entire public supercomputing fleet of the European Union. Furthermore, private capital expenditure by individual US hyper scalers ranges from $60 billion to $75 billion annually, dwarfing the EU’s multi-year public HPC budget of €10 billion.
Infrastructure & Resource Metric | EuroHPC JU Network (Combined Flagships) | US Hyperscaler Ecosystem (Aggregate Top Tier) | Structural Implication for Medical AI |
Total High-End Accelerator Count | ~57,000 GPUs across primary systems | >2,000,000 H100-equivalent GPUs deployed | Public EU compute is an order of magnitude smaller than private US clusters. |
Share of Global AI Supercomputing Capacity | ~4.8% global share | ~74.5% global share | US hosts the dominant share of hardware capable of training frontier models. |
Primary System Architectural Focus | Scientific HPC simulations, FP64 precision, emerging FP8 partitions | High-throughput deep learning, distributed FP8/FP4 training clusters | EuroHPC systems were historically built for physics/climate modeling, requiring retrofit for LLMs. |
Annual Capital Expenditure (CapEx) | ~€10 Billion multi-year public investment budget | $60 Billion – $75 Billion annually per major firm | US private capital scale accelerates hardware iteration cycles far past public budgets. |
Commercial Access Mechanisms | Fast Lane (4-day approval), Playground (2-day SME access) | On-demand API, dedicated cloud instance reservations | EU public access friction is lowering, but scaling remains bound by capacity limits. |
This compute asymmetry creates an operational dilemma for European healthcare providers and AI developers. While EuroHPC facilities offer streamlined access programs like "Playground" and "Fast Lane" to provide small to medium enterprises (SMEs) with GPU hours within days, the total available capacity remains bottlenecked. Consequently, European medical AI developers frequently resort to US cloud providers, Amazon Web Services (AWS), Microsoft Azure and Google Cloud Platform (GCP), which collectively control approximately 70% of the European cloud market, compared to just 15% held by domestic cloud providers.
To mitigate data residency concerns, US hyperscalers have launched "Sovereign Cloud" initiatives within Europe. While these isolated regional infrastructures guarantee that protected health information (PHI) resides within EU borders to satisfy General Data Protection Regulation (GDPR) mandates, they do not eliminate systemic dependency.
The underlying hardware, hypervisors, and foundational software stack remain controlled by American entities, leaving European healthcare systems vulnerable to extraterritorial regulatory actions, such as the US CLOUD Act, and enterprise lock-in.
Data Strategy and Governance: The EHDS Framework and Implementation Friction
While the US maintains an advantage in raw compute and capital, Europe possesses a structural asset: centralised, universal public healthcare systems that generate comprehensive, longitudinal patient records. Recognising that data quality and diversity dictate model efficacy, the European Union enacted Regulation (EU) 2025/327, establishing the European Health Data Space (EHDS).
Published in the Official Journal in March 2025, the EHDS framework mandates both primary data sharing for clinical care and secondary data reuse for scientific research, health innovation, and algorithm development. The secondary use provisions create a legally enforced, federated data-sharing architecture across all 27 Member States. Under this governance structure, national Health Data Holders, such as hospitals, public registries, and biobanks, are obligated to make electronic health data available to statutory Health Data Access Bodies (HDABs) established in each Member State.
Entities seeking to train medical foundation models must submit a formal application to an HDAB detailing the research purpose, requested datasets, and technical security parameters. Upon issuance of a data permit, data processing occurs exclusively within a air-gapped Secure Processing Environment (SPE). External extraction of raw data is prohibited; researchers may only export aggregated, non-reidentifiable analytical results. These national access bodies are interconnected globally through the federated HealthData@EU network, creating a cross-border research framework.
EHDS Phase / Component | Regulatory Target Date | Statutory Mechanism | Operational Implications for AI Training |
Regulation Entry into Force | March 26, 2025 | Regulation (EU) 2025/327 enacted | Initiates the transition window for Member States to establish legal and technical infrastructure. |
Primary Use Application | March 26, 2027 | Mandatory exchange of Patient Summaries & ePrescriptions via MyHealth@EU | Standardizes clinical record formats, establishing baseline data harmonization across borders. |
EHR System Conformity | March 26, 2028 | Harmonized EU certification for Electronic Health Record manufacturers | Eliminates proprietary vendor formats, enforcing open API accessibility across clinical software. |
Secondary Use Enforcement | March 2029 | Operational HealthData@EU network & mandatory HDAB data permits | Opens standardized access to cross-border clinical, genomic, and biobank datasets for model training. |
Extended Secondary Datasets | March 2031 | Inclusion of social determinants, population health, and environmental data | Enables multi-modal contextual pre-training for broad population-level health risk modeling. |
Despite its potential, the EHDS introduces secondary operational frictions that could hamper fast-moving AI developers. First, the timeline for secondary use enforcement extends to March 2029, leaving a multi-year bridge period during which European data access remains fragmented across national borders. Second, the technical shift requires connecting previously isolated hospital servers to standardized external Application Programming Interfaces (APIs). Legacy electronic health record (EHR) infrastructure across European hospitals often lacks modern cybersecurity protections, significantly increasing the cyber-attack surface at the API layer.
Finally, the tension between the EHDS secondary use rules and existing GDPR mandates introduces regulatory complexity. While the EHDS expands access to pseudonymised datasets via SPEs, it maintains an unconditional patient opt-out right for secondary data use. If opt-out rates spike in specific demographics or Member States, training datasets could suffer from selection bias, compromising the generalisation performance of resulting medical foundation models.
European Champions vs. US Dominance: Strategic Model Architecture
The competitive landscape of medical foundation models highlights two contrasting strategic approaches: the American paradigm of massive, general-purpose LLMs fine-tuned for healthcare, versus a European pivot toward targeted open-weight architectures and biology-native reasoning systems.
US Big Tech dominance is exemplified by systems such as Google’s Med-PaLM 2 and its commercial successor MedLM, alongside Microsoft’s deep integrations of OpenAI architectures into Epic Systems' hospital workflows. Med-PaLM 2 achieved expert-level performance on the US Medical Licensing Examination (USMLE) with scores exceeding 86.5%, leveraging self-supervised training across multi-billion-parameter text, vision, and genomic pipelines (Med-PaLM M). These commercial models are tightly integrated into proprietary cloud ecosystems, providing end-to-end clinical documentation, automated charting, and diagnostic support.
In response, European pioneers are eschewing direct competition on brute-force parameter scale, focusing instead on structural domain expertise, open-weight transparency, and multi-modal biological reasoning. Rather than attempting to train multi-trillion parameter general models from scratch, the European paradigm relies on taking specialised base architectures, ingesting domain corpora, such as PubMed Central, clinical trial repositories and spatial transcriptomics and performing continual pre-training. The resulting open-weight engines are deployed either directly on-premises within hospital perimeters to meet strict data sovereignty requirements or integrated into iterative, agentic wet-lab feedback loops for automated drug discovery.
Mistral AI: The Open-Weight and Sovereign Deployment Strategy
Paris-based Mistral AI has positioned itself as a core provider of open-weight foundation architectures. Supported by a €1.7 Billion financing round valuing the company at €11.7 Billion (with anchor investments from semiconductor leader ASML), Mistral has pursued a dual deployment strategy. It delivers high-reasoning models, such as Magistral Small and Magistral Medium, which utilise explicit reasoning chains for multi-step analytical problem solving, while allowing enterprise customers to host models locally or within sovereign perimeters.
In healthcare, community driven adaptations such as BioMistral have demonstrated the power of continual pre-training. By executing targeted pre-training of open-weight base architectures on specialised corpora like PubMed Central, BioMistral integrates clinical domain memory without requiring the compute resources of a ground-up foundation model. Furthermore, Mistral’s dedicated "AI for Science" division builds agentic frameworks designed to automate molecular target discovery, simulate numerical physics, and predict complex bio-molecular interactions.
Owkin: The Agentic Biological Superintelligence Approach
French AI biotech champion Owkin has embraced a biology-native strategic vision, operating on the premise that while American tech giants have captured general-purpose textual LLMs, the domain of biology-native reasoning remains uncolonized. Rather than building conversational text bots, Owkin constructs multimodal, agentic architectures designed to act as "Autonomous AI Scientists" for biopharmaceutical R&D.
Owkin’s flagship platform, K Pro, accesses structured multi-modal datasets, incorporating spatial transcriptomics, digital pathology, and clinical histories, to form and test biological hypotheses autonomously. Owkin grounds its foundation models through closed-loop validation in physical wet-lab infrastructures and continuous feedback loops from major oncology networks.
Through the €33 Million Bpifrance-funded PortrAIt consortium, developed alongside Europe's leading cancer center, Gustave Roussy, Owkin is deploying digital pathology AI tools across French hospitals to extract predictive biomarker signatures directly from routine tissue slides. At the Franco-German Digital Sovereignty Summit, Owkin, Gustave Roussy, and Charité (Germany) announced a joint pan-European agentic infrastructure. This initiative aims to structure and harmonize biomedical data across borders, creating biology-native reasoning models to automate therapeutic discovery.
However, European start-ups face a persistent structural paradox: distribution capture. Despite their sovereign positioning, European foundation model developers remain dependent on US cloud infrastructure for global commercialization. Mistral’s flagship models are distributed via Amazon Bedrock, Google Cloud’s Vertex AI, and Microsoft Azure. While this grants global distribution, it means European innovations frequently generate revenue and API traffic that reinforce the dominance of US cloud ecosystems.
Regulatory Frameworks, Market Access and Capital Asymmetries
Beyond compute hardware and data access, Europe’s path toward medical AI autonomy is shaped by a stringent regulatory environment and fragmented market access pathways.
The EU AI Act and Medical Device Harmonisation
The EU AI Act (Regulation (EU) 2024/1689) imposes strict compliance obligations on artificial intelligence technologies, establishing explicit risk tiers. Under Annex III of the AI Act, AI systems integrated into standalone Software as a Medical Device (SaMD) or those serving as safety components of medical devices—are classified as High-Risk.
This classification requires medical AI systems to undergo formal conformity assessments, enforce strict data governance to eliminate training bias, maintain continuous risk management logs, and ensure human oversight mechanisms. Crucially, the compliance timeline sets a hard deadline of August 2026 for high-risk AI applications, extending to August 2027 for SaMD that requires third-party Notified Body review under the Medical Device Regulation (MDR) or In Vitro Diagnostic Medical Devices Regulation (IVDR).
Because European Notified Bodies already face severe administrative backlogs in clearing traditional medical hardware under MDR, adding complex foundation model software assessments creates a potential regulatory bottleneck. US developers, backed by capital reserves, can absorb these compliance costs more easily than resource-constrained European startups.
European Reimbursement Fragmentation: DiGA vs. PECAN
Even when a European medical AI model secures a CE mark under the MDR and satisfies the EU AI Act, it confronts a fragmented reimbursement landscape. Unlike the single commercial market of the United States, where developers negotiate directly with major private insurers or Medicare/Medicaid, Europe requires country-by-country Health Technology Assessment (HTA) negotiations.
Germany and France have established fast-track reimbursement frameworks for digital health applications (DiGA and PECAN, respectively), but structural differences persist.
Operational Metric | German DiGA Framework (BfArM) | French PECAN Framework (HAS / CNEDIMT) | Strategic Impact on Scaling AI |
Enacting Legislation | Digital Healthcare Act (DVG 2019) / DigiG (2024) | Social Security Financing Act / PECAN (2023) | Pioneers formal statutory reimbursement for digital therapeutics. |
Covered Patient Population | ~73 Million statutory health insurance beneficiaries | ~68 Million national health insurance beneficiaries | Broad coverage, but isolated to domestic populations. |
Device Scope & Risk Tiers | Class I and Class IIa (expanding to IIb in 2026) | Digital Medical Devices (DMN) & Telemonitoring systems | France provides explicit early pathways for remote monitoring workflows. |
Provisional Reimbursement Duration | 12-month trial window to prove positive care effects | 12-month non-renewable provisional coverage window | Allows early revenue generation while formal RCT data is collected. |
Reimbursement Rates & Cap | Negotiated post-trial with GKV-Spitzenverband | Capped initial package (~€435) + follow-up (max €780/yr) | French framework enforces tighter statutory price caps on digital solutions. |
Cross-Border Reciprocity | None; clinical trials must meet BfArM standards | None; requires local HTA & French benefit proof | High friction: Evidence generated in Germany is not automatically accepted in France. |
This lack of cross-border reciprocity means a medical AI foundation model approved for reimbursement under Germany’s DiGA directory must essentially re-do HTA evaluations, clinical trials and localised software integrations to enter France under PECAN or market to NHS trusts in the UK. Other European nations remain categorised as "fast followers" (Belgium, Italy, Netherlands) or lack formal reimbursement pathways entirely, forcing startups to navigate a patchwork of regional health authorities.
The Capital Markets and Venture Gap
Underpinning both infrastructure and market entry challenges is Europe’s structural venture capital shortfall. While early-stage seed capital for European AI research is relatively robust, late-stage growth capital is severely limited. The European economy faces an estimated annual investment shortfall of €270 Billion compared to the United States across technological sectors.
Europe's capital ecosystem is hindered by fragmented financial markets, shallow deep-tech venture funds, and conservative institutional asset allocation. As a result, when European medical AI startups reach the scale-up stage, requiring tens of millions of Euros for multi-centre clinical trials and GPU cluster reservations, they often rely on American venture funds or agree to strategic buyouts by US corporate entities.
Comparative Assessment: Sovereignty Ambitions vs. US Dominance
A holistic evaluation of Europe's posture across the medical AI value chain illustrates a stark contrast between regulatory ambitions and operational market realities. In computational infrastructure, the European ambition centers on establishing public EuroHPC AI Factories and enforcing sovereign regional clouds. However, the market reality is defined by American dominance, with US entities controlling nearly three-quarters of global GPU supercomputing capacity and over 70% of the European cloud hosting market. This leaves European developers dependent on foreign hyperscalers for large-scale training runs.
In data assets and governance, European policy aims to leverage centralized healthcare systems by federating cross-border patient records through the European Health Data Space. The operational reality, however, is slowed by legacy hospital IT environments, high cybersecurity vulnerabilities at the API layer, and complex patient opt-out mechanisms under GDPR that threaten dataset completeness. While Europe possesses superior longitudinal clinical records, turning these assets into AI-ready training sets remains a multi-year administrative and technical undertaking.
Regarding model architecture, Europe has ceded ground in general-purpose text LLMs, where US Big Tech holds a firm lead. Instead, European pioneers are succeeding by focusing on open-weight architectures, targeted clinical fine-tuning, and biology-native reasoning systems. Companies like Mistral AI and Owkin demonstrate that European research excels when applied to spatial biology, digital pathology, and closed-loop biopharmaceutical discovery rather than broad conversational chat interfaces.
Finally, in regulation and market access, Europe leads globally in safety standards through the EU AI Act and structured digital health reimbursement schemes like DiGA and PECAN. Yet, the market reality reveals severe friction: backlogged Notified Bodies under MDR delay software clearances, while a lack of cross-border HTA reciprocity fragments the European internal market. Combined with a massive venture funding deficit, European medical AI innovators face high barriers to achieving commercial scale within their home markets.
Strategic Trajectory and Recommendations
The analysis indicates that Europe cannot achieve technological self-sufficiency by attempting to replicate the US hyperscaler model of general-purpose, compute-heavy LLMs. The capital intensity, hardware concentration, and cloud footprint of US Big Tech remain unmatchable by public EU budgets alone.
However, permanent reliance is not inevitable if European policy and industry stakeholders pivot toward a domain-specific strategy focused on biological intelligence, open-weight architectures, and federated data assets. Europe retains an advantage in scientific research, clinical expertise, and structured patient data repositories.
To translate these systemic strengths into sustainable technological autonomy, the European AI ecosystem should prioritise four strategic vectors:
Capitalise on Biology-Native AI
Rather than allocating scarce public compute toward building localized clones of general text engines, public and private capital should target biology-native reasoning platforms. Domains such as automated biological hypothesis generation, spatial transcriptomics, structural biophysics, and digital pathology represent open frontiers where scientific context matters more than brute-force parameters. Initiatives like Owkin's agentic biological infrastructure demonstrate how European consortia can achieve global leadership in these specialised domains.
Accelerate Operationalisation of EHDS Secure Processing Environments
EU Member States must prioritise the technical roll-out of the EHDS ahead of the statutory 2029 deadline. Establishing standardised, highly secure APIs and federated SPEs across major academic medical centres will allow European AI developers to train multi-modal models on diverse, population-scale datasets. This will create a data moat that foreign technology firms cannot easily replicate due to strict cross-border data transfer limitations.
Harmonise Digital Health Reimbursement Pathways
The European Commission and Member State health authorities should establish a unified cross-border HTA framework for medical AI. Building on the foundations of Germany's DiGA and France's PECAN, a "European Mutual Recognition" pathway for software medical devices would allow an AI application validated in one Member State to rapidly gain provisional reimbursement access across the EU. This would dramatically expand the addressable market for homegrown start-ups, attracting the late-stage venture capital needed to scale.
Institutionalise Open-Weight and Sovereign Edge Deployment
To insulate clinical infrastructure from geopolitical disruptions and cloud vendor lock-in, European healthcare systems should standardise on open-weight foundation architectures deployable on-premise or within sovereign public clouds. Supporting players like Mistral AI in developing domain-specialized, open-weight base models ensures that clinical workflows remain audit-ready, GDPR-compliant, and independent of external API endpoints.
By integrating its public supercomputing investments with unified health data access, streamlined regulatory clearance, and domain-specific biological AI research, Europe can establish a resilient, competitive, and sovereign medical AI ecosystem.
Nelson Advisors > European HealthTech, MedTech, Digital Health Investment Banking
Nelson Advisors specialise in Mergers and Acquisitions, Partnerships and Investments for Digital Health, HealthTech, MedTech, Health IT, Consumer HealthTech, Healthcare Cybersecurity, Healthcare AI companies. www.nelsonadvisors.co.uk
Follow Nelson Advisors LinkedIn Page > https://www.linkedin.com/company/nelson-advisors/
Nelson Advisors regularly publish Thought Leadership articles covering market insights, industry trends, deal commentary, market analysis & predictions. https://www.healthcare.digital
Nelson Advisors publish Europe's Leading Healthcare Technology Investment Banking Newsletter every week, join 5000+ HealthTech and MedTech subscribers today! https://lnkd.in/e5hTp_xb
#VentureCapital #PrivateEquity #Founders #SeriesA #SeriesB #Founders #SellSide #TechAssets #Fundraising #BuildBuyPartner #GoToMarket #PharmaTech #BioTech #Genomics #NelsonAdvisors #HealthTech #MedTech #DigitalHealth #HealthIT #Cybersecurity #HealthcareAI #FemTech #Mergers #Acquisitions #Partnerships #Growth #Strategy #NHS #UK #Europe #USA #Canada
Nelson Advisors LLP
Hale House, 76-78 Portland Place, Marylebone, London, W1B 1NT
Meet Nelson Advisors @ 2026 Events
Digital Health Rewired > March 2026 > Birmingham, UK
NHS ConfedExpo > June 2026 > Manchester, UK
HLTH Europe > June 2026, Amsterdam, Netherlands
HIMSS AI in Healthcare > July 2026, New York, USA
Bits & Pretzels > September 2026, Munich, Germany
World Health Summit 2026 > October 2026, Berlin, Germany
HealthInvestor Healthcare Summit > October 2026, London, UK
HLTH USA 2026 > October 2026, USA
Barclays Health Elevate > October 2026, London, UK
Web Summit 2026 > November 2026, Lisbon, Portugal
MEDICA 2026 > November 2026, Düsseldorf, Germany
Venture Capital World Summit > December 2026 Toronto, Canada
Comments