The U.S.-China competition in artificial intelligence has entered a second phase, according to a Chosun Ilbo report: the fight is no longer only about semiconductors, but about control of AI models themselves.

This shift matters because the first phase had a clear instrument. Washington restricted access to advanced chips. Beijing poured subsidies into domestic fabrication. The battleground was silicon, and the weapons were export licenses and fab budgets.

The second phase is harder to police. A model is not a crate of GPUs crossing a border. It is weights, code and training data that move as files. That structural difference forces both governments to build new tools of control, and neither has finished building them.

What does "model control" actually mean?

The report frames the change as an extension of the same strategic logic. If advanced chips determine who can train frontier models, then the models become the next chokepoint. Governments that once tracked wafer shipments now face a target that can be copied, leaked or open-sourced in minutes.

Readers should note what the source does not say. The Chosun Ilbo piece does not provide:

• Specific new regulations or executive orders with dates
• Named companies, model families or parameter counts
• Analyst estimates of market size or revenue impact
• Statements from U.S. or Chinese officials

The story is a framing argument, not a disclosure of a concrete policy event. That gap deserves attention before anyone treats it as a policy announcement.

Why the semiconductor phase is no longer sufficient

The first phase produced results, but its limits are now visible. Chip controls slow an adversary's training runs. They do nothing once a capable model exists and circulates in weights and weights alone.

This is the structural problem the report gestures at. A fab takes years and billions of dollars to build, so export controls bite hard and slow. A model checkpoint takes seconds to transfer, so equivalent controls on models would need a completely different enforcement apparatus — one that does not yet exist in public form.

The skeptic's question follows naturally. If Washington could not fully stop chip diversion through third countries, can it stop model weights from crossing borders at all? The source offers no answer. It identifies the shift, not the mechanism.

What is not said matters most

Every promise of "model control" carries a caveat here. The article does not cite a single regulator, nor a draft rule, nor a company reaction. Without those anchors, the claim rests on the observable trend: AI competition has broadened beyond hardware, and both states now treat models as strategic assets.

That trend is real, but the policy response remains speculative. Journalistic caution applies. A shift in the object of competition is not the same as a shift in the ability to control it.

Where the rivalry goes next

The commercial picture changes if model-level controls arrive in force. Companies that sell access to models, rather than chips, would inherit the compliance burden that Nvidia and its peers carry today. Chinese developers would face a new layer of restrictions layered on top of the existing chip ceilings.

For now, none of that is confirmed. The Chosun Ilbo account signals a direction of travel, and the direction suggests the next round of U.S.-China measures will target the models built on restricted silicon, not merely the silicon itself.