The US administration is in talks with leading AI companies on a set of voluntary standards for releasing new models. According to the Financial Times, an announcement could come as early as next week. The guidelines – non-binding – would establish benchmarks and timelines for the most advanced systems, and clarify who gets access, within the United States and abroad.

At first glance, the initiative looks like an attempt to channel innovation without braking it with rigid rules. But the voluntary nature reveals its complexity: in a sector where competition is driven by model scale and technology availability, shared benchmarks and access restrictions – even on paper – can tip the scales for deployment decisions, especially for organizations operating outside US jurisdiction.

For those already considering on-premise or hybrid setups, there are two intertwined challenges. On one side, possible export controls on models trained on sensitive data or deemed to have critical capabilities could limit the ability to download and run certain LLMs on self-hosted infrastructure outside the US, echoing familiar dynamics from dual-use technology controls. On the other, the existence of shared standards – albeit voluntary – can help those who must meet internal compliance or GDPR requirements: a model released according to recognised parameters offers more assurances when documenting the supply chain of an application handling protected data.

It is no coincidence that the discussion comes as the EU’s AI Act takes shape and the US issues successive executive orders on AI safety. Major model vendors – often headquartered in America – are caught between the pressure to release quickly (and at scale) and the need to manage international compliance. For end users, particularly those unwilling to outsource the entire inference pipeline to the cloud, the evolution of these voluntary standards is a signal to watch: they could become the precursor of stricter requirements, or instead create a self-regulation perimeter that forestalls invasive regulation.

The question of checks and balances remains open: voluntary standards risk being ignored by those who release unfiltered open-weight models, creating a two-tier system between those who adhere and those who do not. For an organization evaluating an on-premise LLM, the provenance and release regime of the model are far from details: they affect audits, accountability, and even the feasibility of customisation through fine-tuning. As the formal announcement is awaited, the European and Italian tech community would do well to read these moves as a possible watershed between AI tamed by governance and a gold rush where effective control stays in very few hands.