AI Open vs. Closed Weights Battle Lands on Price

The fight over open-weight AI models is no longer just a philosophical debate. It is turning into a cost-of-capital question for every company that uses large language models without building its own.

Nearly every major Western AI company, including OpenAI, Google, and Cohere, signed an open letter last week supporting open-weight models. Jensen Huang of Nvidia posted it online. The letter argues that open-weight models strengthen competition and give customers more control. It came after reports that the US administration wanted to ban open-weight Chinese AI models under national security grounds.

The only top North American lab that did not sign was Anthropic. CEO Dario Amodei later wrote a blog post saying Anthropic is not against open weights. He is concerned about Chinese "distillation" operations that train AI off US models, arguing they could bring China's frontier capabilities to within months of the US. In a separate letter, Anthropic and over 1,200 AI company employees called for an international effort to pace frontier automated AI development. Yoshua Bengio backed that call.

For the companies that do not build their own LLMs, the debate may come down to price. Anthropic's Claude Code is the choice model for many developers, but tokenmaxxing has become prohibitively pricey. Chinese open-weight models now perform well enough at a fraction of the cost. The share of tokens spent on cheaper Chinese models by US companies has risen sharply in 2026, according to data cited in the open letter's supporting materials.

The difference is material. A developer running a batch of 10 million tokens on Anthropic's top-tier model pays about $300 at current enterprise rates. The same batch on an open-weight Chinese model costs roughly $12, according to pricing data from three cloud providers. For a startup burning through $50,000 a month in inference costs, the math is simple.

That price gap is why the open letter matters beyond policy circles. If the US administration imposes restrictions on Chinese open-weight models, American companies lose access to the cheapest inference on the market. They would either absorb higher costs or switch to closed US models at even steeper prices. Either way, the savings vanish.

The counterargument, raised by Anthropic and others, is that cheap Chinese models carry security risks. Distillation operations can extract the reasoning architecture of a frontier model, then replicate it in a cheaper open-weight version. That speeds up China's catch-up timeline, Amodei argued. The trade-off is between short-term cost savings and long-term strategic risk.

For now, the market has voted. Open-weight Chinese models have gained adoption not because developers prefer them, but because they work well enough at one-tenth the cost. The open letter's signatories are betting that regulation will not kill that price advantage. Anthropic is betting that security concerns will eventually justify the premium.

Both sides may be wrong about the policy outcome. But for the companies that just need to run inference, the decision is already made.

NVDA (Nvidia) carries an Alpha Score of 70, labeled Moderate, at $190.06, down 3.53% today. Nvidia's Jensen Huang posted the open letter, and the company invested $5 billion into Safe Superintelligence, an AI startup founded by former OpenAI chief scientist Ilya Sutskever.