The Nvidia Doctolib deal announced in October 2026 is a minority investment project, not an acquisition. According to public information, Nvidia is gaining neither control of Doctolib nor access to medical records. However, the transaction raises a broader question for European artificial intelligence: can Europe keep its data in Europe while depending on American components, software, and expertise?
What does the deal between Nvidia and Doctolib actually provide for?
On October 13, 2026, Nvidia and Doctolib announced a planned minority investment in the French digital health company. The amount, equity percentage, and valuation were not disclosed. No acquisition, majority control, or Nvidia access to medical records appears in the official announcements.
The distinction matters. A minority stake means that an investor owns less than half of the capital and does not automatically control the company’s decisions. It may nevertheless come with contractual rights: an observer seat, enhanced information rights, vetoes over certain transactions, or commercial commitments. As of October 21, 2026, these potential rights are not public.
Doctolib says it has been using Nvidia technologies for about a year and plans to benefit from the group’s specialists’ support to develop healthcare-focused artificial intelligence products and models in Europe. According to Doctolib in 2026, the platform serves approximately 580,000 healthcare professionals and 100 million patients in France, Germany, Italy, the Netherlands, and the United Kingdom.
This transaction is part of a broader strategy. On October 12 and 13, 2026, Nvidia announced planned investments in four companies based in France: Doctolib, Entalpic, Genesis AI, and Pigment. Nvidia also says it committed nearly $10 billion to the European technology ecosystem in 2026, following investments in Alice&Bob, Gradium, and Scintil Photonics.
Nvidia Doctolib: who controls health data?
In October 2026, Doctolib stated that it alone retains control of health data and hosts it in the European Union or the United Kingdom, depending on national requirements. The Nvidia Doctolib announcement grants no access to medical records. The financial investment and data hosting are therefore two legally and technically separate relationships.
La documentation de confidentialité publiée par Doctolib en juillet 2026 mentionne AWS EMEA et S3NS avec Google Cloud Platform parmi ses fournisseurs d’infrastructures certifiées HDS, la certification française d’hébergement des données de santé. Les données de service sont indiquées comme stockées dans l’Union européenne. Atos/Eviden héberge en France une clé de chiffrement de Doctolib. Nvidia n’y figure ni comme hébergeur de santé ni comme sous-traitant des données.
In July 2026, Doctolib also documented AES-256 encryption for stored data and TLS 1.2 for data transfers. An additional AES-256 layer on the server side is applied to health data. Encryption makes information unreadable without the appropriate key, but its value depends on key separation, access rights, and the traceability of operations.
Confusing an investor, a processor supplier, and a hosting provider therefore leads to a mistaken reading of the risk. Conversely, a European location alone is not enough to guarantee complete autonomy: the law applicable to the provider, key management, and the ability to change platforms also matter. Managing a sensitive provider deserves the same precautions as the securing data entrusted to a service provider.
Is Europe becoming dependent on Nvidia technologies?
European dependence on Nvidia cannot be measured by the nationality of the data center. It becomes apparent when a project can no longer train, deploy, or maintain its models without the accelerators, libraries, or tools of a single provider. European hosting can therefore remain technically dependent on a non-European technology stack.
An accelerator is a component designed to execute the calculations required for artificial intelligence quickly. Nvidia’s GPUs, or graphics processing units, dominate many AI projects because of their power and software ecosystem. The risk does not come solely from the chip: code, team expertise, and operating procedures can ultimately become dependent on the same ecosystem.
This dependence is not necessarily a mistake. Choosing mature technology can reduce time to market and execution risk. Honestly, rebuilding an entire European supply chain solely to claim theoretical autonomy makes little sense if the product has neither users nor a validated business model.
The problem begins when switching becomes economically unrealistic. Code rewriting, model retraining, performance degradation, a shortage of skills, and data migration can all accumulate. In the projects we undertake, we often see lock-in emerge in details overlooked during initial planning: an incomplete export format, a managed service that cannot be reproduced, or a key controlled by the provider.
The Nvidia Doctolib case thus shows that supposedly proprietary AI can remain dependent on external components. Developing its own models allows a company to be less dependent on APIs such as those from OpenAI or Anthropic, but it does not eliminate dependence on computing capacity. Choosing a European model such as Mistral AI and its Large 4 model does not, by itself, resolve the infrastructure issue either.
How can the sovereignty of an AI project be measured?
In 2026, the European Commission’s Cloud Sovereignty Framework distinguishes several dimensions: the origin of components, the portability of models and interfaces, control of data and keys, the location of processing, and continuity without provider assistance. A serious analysis evaluates each layer separately rather than assigning a global “sovereign” or “non-sovereign” label.
Portability refers to the ability to move an application, its data, and its models to another environment without disproportionate reconstruction. Since September 12, 2025, the European Data Act has included provisions intended, among other things, to facilitate switching between cloud services and data-processing services. However, the law does not make an architecture portable by magic.
Here is a framework that can be used directly during a tender or architecture review:
- Components: Which GPUs, system software, and proprietary services are essential, and is there a tested replacement?
- Models and interfaces: Can the model weights, parameters, and logs be exported in documented formats?
- Data and keys: Who can read the data, administer access, and revoke encryption keys?
- Location: In which countries do storage, training, inference, and backup take place?
- Continuity: Can the company operate if the provider raises prices, interrupts a service, or withdraws its support?
Inference is the stage during which an already-trained model produces a response. It can process sensitive data even when training took place elsewhere. The actual flows must therefore be mapped, rather than merely asking for the address of the main server. A resilient architecture also provides for a cloud disaster recovery plan in the event of a data center outage.
| Layer being assessed | Decision question | Proof to be requested | Main risk |
|---|---|---|---|
| Components | Does the service work on another accelerator? | Benchmark test or fallback architecture | Scarcity or rising computing costs |
| Models and APIs | Can the model be changed without rewriting the product? | Export formats and interface documentation | Software lock-in |
| Data and keys | Does the client retain control over access? | Rights matrix and revocation procedure | Uncontrolled access |
| Processing location | Where do storage, backup, and inference take place? | Mapping data flows and listing subprocessors | Contractual non-compliance |
| Continuity | Can the service survive the provider’s withdrawal? | Tested exit plan and contractual deadlines | Business interruption |
What clauses should be included before choosing an AI provider?
In 2026, an artificial intelligence contract must specify the locations where processing takes place, subprocessors, data usage, ownership of the results, export formats, and the exit plan. It must also set out return deadlines, migration costs, and the applicable conditions if an API, model, or component disappears.
The least visible pitfall concerns exit costs. A provider may autormize the export while charging for the transfer, delivering a format that is difficult to reuse, or imposing a deadline incompatible with the business. A useful reversibility clause therefore describes a testable procedure, a responsible party, a deadline, and the exact scope of the items to be returned.
On the agency side, the best practice is to build a portability prototype before going into production: export a dataset, replace an API, and restore the service in a separate environment. This test takes time initially, but reveals dependencies before they become structural.
For Nvidia Doctolib, the public information available as of October 21, 2026, does not make it possible to assess governance rights, any potential board representation, GPU purchase commitments, or exit clauses. Two shortcuts must be avoided: presenting Nvidia as the owner of medical data, or assuming that a minority stake has no strategic impact.
The right trade-off is not absolute independence. It is a known, limited, and reversible dependency. For an SMB, it is often better to secure two decisive points—the export of data and a replaceable architecture—than to fund a complex multicloud architecture that will never be tested.
Cadrer les dépendances techniques, contractuelles et réglementaires avant le development évite la plupart des migrations d’urgence. Un regard extérieur peut alors aider à distinguer les garanties démontrables des simples promesses de souveraineté.
FAQ about Nvidia’s investment in Doctolib
Did Nvidia acquire Doctolib?
Nvidia did not acquire Doctolib. The two companies announced on October 13, 2026, a plan for a minority stake, the amount and percentage of which were not disclosed.
Can Nvidia access Doctolib's medical records?
According to public announcements from October 2026, Nvidia does not receive any access to medical records. Doctolib states that it retains sole control over health data.
Where does Doctolib host health data?
In July 2026, Doctolib stated that it stored its service data in the European Union on HDS-certified infrastructure, including AWS EMEA and S3NS with Google Cloud Platform. Certain national requirements may also lead to hosting in the United Kingdom.
Is AI hosted in Europe sovereign?
AI hosted in Europe is not automatically sovereign. In 2026, the European Commission distinguishes between location, key control, portability, the origin of components, and the ability to operate without the provider.