Frontier LLMs are becoming infrastructure.

People increasingly use them as the interface to the world's knowledge. They write software, learn, conduct research, communicate, and solve problems through them. As they become general-purpose infrastructure, they should increasingly be treated like other infrastructure.

The closest historical analogy is the telephone network.

Telephone companies made enormous upfront investments to build nationwide networks. Society allowed them to recover those investments and earn profits by charging for access. It did not allow them to decide who could make lawful phone calls or what services could be built on top of the network. They became common carriers: providers of a neutral service.

Commercial LLMs should evolve in the same direction

AI providers should be free to charge whatever the market will bear for inference. They should compete on model quality, latency, reliability, and price. But once a customer pays for inference, the resulting outputs should be theirs to use for any lawful purpose.

That includes model distillation.

Blocking distillation is analogous to an incumbent telephone company claiming the right to prevent customers from using its network to build competing communications services. Common carrier law evolved in the opposite direction: the operator could charge for access, but not dictate what lawful services customers built on top of the network.

The same principle should apply to LLMs. The provider is selling computation, not control over what customers lawfully do with its results.

Making distillation a protected right would likely increase token prices in the short term. AI companies would need to price inference closer to its true cost, because selling below cost would no longer subsidize only customer acquisition—it would also subsidize future competitors.

That's how competitive markets should work. Anyone building a competing model would also pay those same market prices to distill existing models.

Competition would shift from contractual restrictions to engineering: better models, lower costs, and more efficient infrastructure.

The natural way to give back

This is also consistent with how frontier models came to exist. They were trained on humanity's accumulated knowledge: scientific research, literature, open-source software, and countless public contributions spanning centuries. The engineering required to build these systems deserves to be rewarded. The reward should be payment for providing inference, not the power to control how customers use the outputs they have purchased.

As LLMs become essential infrastructure, neutrality should become part of the bargain. Society provides the knowledge these models are built on; society should retain the right to build on them.