The AI Act takes effect, the US splits over open models: what changes for decision-makers today

Two stories landed days apart, one from Europe and one from the United States, and together they say more about where open-model regulation is heading than any benchmark leaderboard. On August 2, 2026, the European Commission activated its enforcement powers over general-purpose AI (GPAI) models under the AI Act. In the same week, more than 230 American companies — from Nvidia to Microsoft, from Meta to Mistral — signed a public letter opposing any ban on the use of open-weight models, while Anthropic and OpenAI publicly distanced themselves from it. For companies in Switzerland that chose an open-first architecture for compliance reasons, these are two stories worth reading together.

As of yesterday, the AI Act is no longer just theory

Obligations for GPAI model providers have been in force since August 2, 2025, but for a year the Commission lacked the concrete power to enforce them. As of yesterday, August 2, 2026, that transition period is over: the European AI Office can now request models' technical documentation, run its own evaluations, impose risk-mitigation measures, restrict or withdraw a model from the EU market, and issue fines of up to 3% of global annual turnover or €15 million, whichever is higher. The same day also brought new transparency rules into force: chatbots must disclose they are AI systems, deepfakes must be labelled, and AI-generated content must carry a machine-readable marker.

For models placed on the EU market after August 2025, compliance is required immediately; for models released before that date, providers have until August 2027. But the distinction that matters for a company simply using a model, open or closed, hasn't changed: compliance responsibility doesn't fall only on the lab that trained it — it also falls on whoever puts it into service in the European Union.

In the US, the debate has flipped: ban open models, or defend them?

Kimi K3's success in mid-July ignited a discussion in Washington that had so far stayed in the background: banning the use of open-weight models built by Chinese labs, on national security grounds. Nvidia CEO Jensen Huang responded publicly on July 28, defending openness as an engine of American technological leadership. A few days earlier, a group of companies — initially 25, including Nvidia, Microsoft, Meta, IBM, Dell, Palantir, Hugging Face, Mistral, Mozilla, the Linux Foundation, Andreessen Horowitz and Y Combinator — had published the "Open Weights and American AI Leadership" letter. Within days, signatures more than tenfolded, passing 230, with OpenAI and Google among the later signatories.

The letter's position is blunt: banning legitimate American businesses from using open-weight models wouldn't stop anyone determined to misuse them — malicious actors are rarely properly incorporated companies — and would mainly shield the labs that sell access to their closed models via API from competition.

Anthropic pushes back: openness, but with more caution

Anthropic didn't sign the letter, and instead published its own position paper worth reading in full: it acknowledges that an open-weight model carries a harder-to-manage risk than a closed one, because once weights are published they can't be recalled or subjected to centralized guardrails. Its main concerns are authoritarian regimes using open models to gain a lasting military advantage or to entrench domestic repression, plus the risk of cyber or biological misuse. But Anthropic converges on the same point as the letter's signatories: banning use — not development, use — wouldn't solve these risks, and it proposes chip export controls and pre-release evaluations of the most capable models as an alternative to a blanket ban.

What actually changes for a Swiss company using open models today

• The American debate is about using specific models on US soil, not in the EU or Switzerland: no obligation changes for you today, but it's a sign that "which model, trained by whom, run where" remains politically unstable — one more reason to build an architecture that lets you swap models without rebuilding everything.
• The letter's central argument — that the risk lies in who trains and who misuses a model, not in the mere existence of the weights — is the same origin-versus-execution distinction we've written about for months: one more reason, not fewer, to keep running open models on your own infrastructure or a Swiss sovereign cloud.
• The AI Act is no longer an obligation on paper: if you use a GPAI model, open or closed, in the European Union, check now whether your provider's technical documentation satisfies the AI Office's new enforcement powers, before someone else checks it for you.
• The new transparency rules — chatbot disclosure, deepfake labelling, marking of generated content — apply as of yesterday: if your product includes a conversational interface or generates content for the public, now is the time to verify compliance, not during an audit.

This is exactly the work we do every week with our clients: read the regulatory obligations before they become a problem, and build an architecture where the open model you choose today stays compliant tomorrow — regardless of how the political debate over where it was trained evolves.