When G7 leaders meet to discuss artificial intelligence, the sticking point is no longer just ethical or regulatory. A more concrete question is on the table: who has the right to decide which organizations can use frontier models? This marks a step change: we are past general rules on transparency or safety, and into the material control of access to the most capable LLMs—the ones reshaping entire industries.

The Battleground: Closed Models vs. Open Ecosystems

Frontier models—generative AI systems with advanced reasoning—are now a strategic resource. On one side, big cloud providers and private labs offer regulated access via APIs, with usage limits, audits, and safety filters decided centrally. On the other, pressure grows for open-weight models or partial releases that allow public bodies and companies to run inference in controlled environments without relying on third parties. The G7 sits at the center of this tension, with some nations pushing to guarantee national autonomy and others fearing security risks or uncontrolled proliferation.

Why Access Control Is a Sovereignty Issue

When a government or enterprise cannot run an LLM locally because the vendor denies access or imposes restrictive conditions, digital sovereignty is compromised. Sensitive data remains tied to external servers, usage policies can change unilaterally, and fine-tuning customizations become impractical. In defense, healthcare, or finance, that scenario is untenable. That is why the G7 discussion is not abstract: deciding who controls access means defining the real technological autonomy of member states and their businesses.

The On-Premise Deployment Angle

For those evaluating on-premise architectures, the stakes are high. Self-hosting LLMs on proprietary infrastructure requires hardware with adequate VRAM (often enterprise-grade GPUs) and robust serving frameworks, but it offers total control over data, latency, and regulatory compliance. If G7 decisions push toward tighter vendor control, organizations will need to accelerate investment in local stacks to avoid being locked into closed ecosystems. Conversely, a regulated opening could favor a hybrid model where frontier models are distributed on-premise under shared guarantees. At AI-RADAR, we regularly explore the trade-offs of self-hosting: from quantization to reduce memory footprint to fine-tuning pipelines in air-gapped environments—tools that become essential when access is not guaranteed.

A Long-Term Struggle

The G7 tug-of-war is only the beginning. The ability to train and run frontier LLMs is concentrating in few hands, and access control risks becoming a geopolitical weapon. For enterprises, the answer cannot be passive waiting: investing in internal skills, evaluating open-source hardware and frameworks, and designing architectures that avoid single-vendor dependency are tangible steps to preserve freedom of choice. The debate over who controls model access is the debate over who will control innovation in the next decade.