Demis Hassabis, CEO of Google DeepMind, broke his silence with a piece that blends technological vision and geopolitical strategy. In an essay published on Substack, echoed on X and relayed by Axios, he calls for a "global AI watchdog" and specifies that leadership should remain firmly in the hands of the United States. This is no ordinary appeal: it comes from the world’s most powerful private lab for LLM development, and it reads like an attempt to shape regulation before less predictable actors do.

The move must be read against two dynamics that AI-RADAR tracks closely: the growing demand for self-hosted stacks, and the tension between regulatory compliance and dependence on non-EU cloud platforms. The idea of a Washington-led watchdog, however neutral it may sound technically, redefines the balance of power. If standards and audit mechanisms are calibrated to US regulatory and commercial interests, non-American companies – European ones first, but also Asian firms – will face a stark choice: align with rules defined elsewhere, with direct consequences on data residency and software supply chains, or accelerate on-premise infrastructure plans to retain control and sovereignty.

Here the issue becomes structural. An international watchdog with American traction could, in practice, tilt regulation in favor of cloud providers legally based in the US, which can offer integrated certifications and privileged channels to the regulator. Open models distributed under permissive licenses, today the backbone of self-hosted environments, might end up under scrutiny for "systemic risks" that are hard to assess outside US jurisdiction. For organizations already running inference pipelines on bare metal or air-gapped Kubernetes clusters, the message is clear: the path toward technological sovereignty will demand even bolder investment, because global rules – if they ever arrive – will be written with cloud operators in mind, not self-hosters.

Hassabis’ silence on hardware aspects and enforcement modalities is itself loaded with implications. There is no mention of chips, of export-restricted integrated circuits, of VRAM or firmware attestation mechanisms. Yet any watchdog with global ambitions will sooner or later have to address on-device inference and the distribution of quantized models, which escape central control by definition. It is plausible that the next chapters of this debate will see proposals for use licenses conditioned on hardware checks, directly affecting decisions by those who are currently weighing whether to buy GPUs for on-premise training or rely on cloud services.

From a European and Italian vantage point, what is at stake is the viability of GDPR in an AI ecosystem that tends to locate governance outside the continent. Hassabis’ appeal never mentions Europe except as a passive recipient, and that is in itself a signal: America’s regulatory acceleration risks hollowing out EU efforts, unless businesses leverage local architectures to demonstrate effective, not merely declared, compliance. In this landscape, on-premise deployment ceases to be a technical choice and becomes a political-industrial act – the only way to negotiate audit terms that respect both European regulation and the intellectual property of data.