Open weights: the debate shaping sovereign AI

A major debate is shaking Silicon Valley. More than a hundred companies, including Microsoft, NVIDIA, Amazon, Meta and Hugging Face, have signed an open letter supporting open weights models. Anthropic, for its part, stays cautious and only wants to publish models "without dangerous capabilities". For a European SME aiming for sovereign AI, this debate determines the architectures available to you.

Direct answer: an open weights model is an LLM whose weights are published freely, downloadable and runnable on your own infrastructure. For a company subject to the GDPR, it is the only option that guarantees data control, on the condition that you take on the infrastructure, the maintenance and the monitoring.

What is an open weights model?

An open weights model is the AI equivalent of an open source software project. The publisher releases the trained model weights, the billions of learned parameters, for download. Any team can then run the model on its own servers, fine-tune it on its data or embed it in its pipeline.

This model differs radically from closed APIs (OpenAI, Anthropic, Google) where the model stays behind a centralized infrastructure and usage terms dictated by the publisher. The open letter says it plainly: an open weights model can truly belong to the user, while a closed API is only revocable rented access.

Four consequences for a company:

• Data control: your users' data never leaves your perimeter.
• Free hosting: you choose the country of the servers, the hardware and the security policy.
• Reversibility: the model becomes a company asset, not a rented service.
• Cost: you pay for compute, not for a per-token bill.

This control is exactly what small businesses look for when they put data sovereignty at the heart of their AI strategy.

Who signed the open letter in favor of open weights?

The letter gathered more than one hundred Silicon Valley companies. The best-known names are Microsoft, NVIDIA, OpenAI, Intel and Amazon. Its thesis is simple: the transparency and reusability of weights are progress, and the ecosystem should help an AI that can be verified and reused.

The arguments of the signatories converge:

• Auditing models is a security gain: you can check what the system does.
• Research and industry should not depend on the goodwill of three or four publishers.
• Sovereignty is a function of owning the model weight, not renting it.
• Overly restrictive usage policies hurt legitimate innovation.

On the other side, Anthropic is the most visible refusal. The company supports "open-weights models that do not have dangerous capabilities": a nuanced position. An open weights model can indeed be fine-tuned again and misused by a malicious actor. The editorial control of a closed API is simply gone.

Why a French company should follow this debate

This debate is not theoretical for a French SME. It crosses two issues: technological sovereignty and GDPR compliance.

On sovereignty, your options boil down to two families:

1. Process through an API: your data is sent to a third-party provider. Actual location, contract terms and hosting are the provider's.
2. Host an open model: your data stays under your control, on hardware you define, in the Union or in the country of your choice.

In my work as a freelance AI expert, the second option has become the default as soon as personal or sensitive data is involved. It is not a luxury, it is a direct answer to GDPR.

Open weights and GDPR: which articles apply?

GDPR does not talk about model formats. It talks about data processing. The key articles work like this:

• Article 28: any processor that handles data on your behalf must be covered by a contract. With an open model hosted by you, the processing stays under your responsibility.
• Article 44: governs transfers of data outside the European Union. A model hosted in the EU removes the transfer question, unlike a US API exposed to the CLOUD Act (2018 law).
• Article 30: you must keep a register of processing activities. You can trace a local model: you know where it runs, which version, which logs.

The main constraint is therefore the location of the infrastructure. The CNIL reminds you of its key point: in your architecture, data protection is built at the processing level, not only at the application level.

Hosting an open weights model in Europe: sizing

Let us get down to architecture. A model is deployed according to its size:

Range	Use case	Base infrastructure
1 to 8B	classification, summarization, extraction	8-16 GB VM
8 to 32B	RAG, assistance, document analysis	24-48 GB GPU or 1 server
70B and above	cutting-edge generation, complex agents	GPU cluster

The first step is sizing. Before buying a GPU, the VRAM simulator estimates the memory footprint of a chosen model. A calculation done early avoids expensive surprises.

My advice: start small and with a specific use case. Evaluate an 8B model on a business test set, then scale up only if results justify it. For the implementation, see the guide on GDPR-compliant local LLM agentic coding or the method to integrate an LLM without violating GDPR.

What risks must you anticipate when self-hosting?

An open weights model changes the distribution of responsibilities. With local deployment, several risks fall on your team:

• Updates: you must follow released versions and rerun your quality tests.
• Security: a model reachable on a network can be used for exfiltration, via prompt injection.
• Traceability: requests must be logged and processing explanations must be maintained.
• Licenses: open weights come with multiple licenses, some commercial, others restrictive. Check early.

None of these risks are blocking. They must be treated like any other critical software component.

Open weights or proprietary API: choosing the right scope

Criterion	Open weights	Proprietary API
Cost	depends on infrastructure	subscription or per-use
Deployment	your responsibility	simple
Data out of scope	no if hosted in Europe	depends on the provider
GDPR contract	internal	provider
Production release	under your control	provider
Updates	you	provider

The decision is therefore not binary. For sensitive data, I favour open weights: the cost of infrastructure compares to a licence. For a one-off, low-sensitivity function, an API remains reasonable. These trade-offs usually appear during an LLM integration project, in our conversations about AI for European businesses.

Open weights and French data protection policy

The debate echoes a French concern: digital sovereignty. Since the GDPR, French regulators have repeatedly emphasised that cloud and AI choices must keep data readable and under European legal protection. The AI Act also contributes: it adds transparency duties for general-purpose models, whether open or closed.

For a company, this translates into process obligations:

• Document every deployment: version, location, purpose.
• Validate licenses before a production rollout.
• Define a policy for prompt injection and logs.
• Plan a fallback if a model becomes unmaintained.

These rules are the same as for any external software, but they deserve attention because a model is a living component.

Conclusion

The Silicon Valley letter opens a new chapter: owning the weights becomes the support of sovereign AI, and it intersects the GDPR obligations of the European framework. For a French small business, the architecture choice is a sovereignty choice: the model must be under control and the location certain.

In practice, the path is progressive: an evaluation workshop, a prototype, a controlled deployment. That is exactly the type of flow I run when a company needs agents and sovereign AI. This choice of architecture and its compliance benefits are best discussed with a GDPR-aware partner.

Let's talk.