The AI Act just got teeth, and self-hosting open weights may have made you a provider

The GPAI obligations in Chapter V have technically applied since August 2025. What changed on 2 August 2026 is that the Commission can now actually enforce them: request information, demand access to a model for evaluation, require risk mitigation, and fine up to 3% of global annual turnover or order a model withdrawn from the market. The Article 50 transparency duties came into force on the same date.

Everyone I work with read the headline as being about the frontier labs. Mostly it is. But there is a second-order question that lands directly on the architecture I spend my life advocating, and it deserves a straight answer rather than a reassuring one.

Deployer or provider?

The Act's obligations split by role. A deployer using a model has a comparatively light set of duties. A provider who places a general-purpose model on the market carries the Chapter V load: technical documentation, information for downstream providers, a copyright policy, a training-content summary, and for systemic-risk models, evaluation, mitigation, incident reporting and cybersecurity.

If you pull Llama, Mistral, DeepSeek or Qwen weights and serve them internally on vLLM behind your own gateway, which are you?

The honest answer is: it depends on what you did to the weights and who you serve.

• Serving unmodified open weights to your own employees. You are a deployer. This is the common case and it is fine.
• Fine-tuning, then serving internally. Still internal use, but you now hold a modified model and the documentation trail matters, because you cannot point downstream at someone else's model card for behaviour you introduced.
• Fine-tuning, then making it available to customers or partners - even bundled inside your product, even free. This is where the provider question becomes real, and where I have seen European enterprises assume they are safe because "we didn't train it from scratch." Substantial modification plus placing on the market is the trigger, not the pre-training bill.

The compute thresholds for systemic risk are far above anything an enterprise fine-tune touches, so the heavy systemic-risk tier is not the issue. The base provider duties are.

What this means for the sovereign-AI argument

I have argued for two years that jurisdiction beats residency, and that open weights on EU-owned metal is how you get there. That argument survives 2 August intact. It arguably gets stronger: the more of the stack sits under your own control, the easier every documentation and evidence obligation becomes, because you are not asking a US provider to attest to something on your behalf.

But self-hosting is not a regulatory shortcut. It relocates the paperwork rather than deleting it. If you took the weights, you took the documentation duty for what you did next.

The practical list

For clients running open weights in production, these are the four things worth doing this quarter:

1. Write down your role, per model, in one page. Deployer or provider, and why. If you cannot answer in a paragraph, you have found the gap.
2. Keep the provenance chain. Which base weights, which licence, which checkpoint, what data went into the fine-tune, who approved it. This is a git-and-registry problem, not a legal problem, and it is easy while you are doing it and near-impossible to reconstruct a year later.
3. Ship Article 50 disclosure now. Users interacting with an AI system need to know. Synthetic output needs marking. This is UI copy and a metadata field, and it is the cheapest item on the list.
4. Note the 2027 line. Models placed on the market before August 2025 generally have until 2 August 2027. That is a grace period on a specific cohort, not on your programme.

None of this is dramatic. It is the same lesson as GDPR: the organisations that treated documentation as a build-time artefact spent a fraction of what the ones who treated it as an audit-time scramble did.