AI Under Sanctions: The Clash Between Closed and Open Routes

I noticed an interesting detail—US senior officials publicly claiming to investigate Chinese AI companies' "distillation" behavior. The wording itself exposes their cognitive bias regarding technological evolution.

"Distillation" in the AI field is one of the most common technical paths in the open-source model community. Simply put, it involves using large models to generate synthetic data to train smaller models. This isn't theft; it's a standard move in technical iteration. After Meta's Llama series went open-source, it has been distilled countless times by developers globally, and Zuckerberg's attitude towards this was "welcome."

Now, the US is characterizing this common technical operation as "intellectual property theft." The logical chain is: Chinese AI companies call US large models via API, use the knowledge from these models to distill competitive models with comparable performance, and then provide inference services at lower prices.

This is essentially a dispute over technical routes, packaged as a security narrative.

Comparison of Two Models

Dimension	US-style Closed	China-style Open
Technical Path	Closed API + Tech Blockade	Open Source Model + Distillation Iteration
Business Model	Charge per token, high threshold	Extremely low inference cost, universal access
Risk Attitude	Security first, development later	Development first, governance later
Ecosystem Strategy	Control core tech chain	Build application layer ecosystem

What is the core moat of US AI companies? It's the model foundation built through computing power accumulation, and profits obtained through API charging. Chinese models like DeepSeek and Alibaba's Qwen series have achieved performance close to GPT-4 at much lower costs, directly shaking the pricing power of US AI companies.

[!info] The essence of sanctions is not cracking down on "theft," but protecting the pricing power of US AI companies. When Chinese models reduce inference costs to one-tenth of the US, the API charging model faces collapse.

From a technical perspective, "distillation" itself is not illegal. GPT-4's API service terms do prohibit using outputs to train competing models, but during the "distillation" process, what is used is not raw output but processed synthetic data, which is a gray area. The US now wants to use the basket of "national security" to contain this.

What is truly worth noting is the timing of this US statement—China's AI industry is at a critical turning point from "following" to "running alongside."

Key Data on China's AI Industry

• Contribution of open-source models accounts for 38% globally, second only to the US
• Large model inference costs have dropped to 1/10th of the US
• Application-layer innovation is starting to feed back into base models
• Domestic computing chip ecosystem is taking shape

The US choice is: since they cannot suppress China in technical iteration speed, they will use political tools to cut off technology flow. This precisely confirms the typical characteristics of "technological hegemonism"—when the lead in technology advantage is caught up, non-market means are employed.

From an industrial perspective, the actual effect of sanctions may backfire. Chinese AI companies have already significantly reduced the proportion of direct calls to US APIs in the past six months, switching to domestic models and open-source communities. Sanctions actually accelerated the ongoing process of domestic substitution.

[!tip] The marginal utility of sanctions is diminishing. In 2023, 60% of Chinese AI companies' model training relied on US API calls; by mid-2024, this ratio had dropped below 30%.

The true focus of the game is not "distillation" itself, but the autonomous controllability of next-generation AI infrastructure. What the US wants to strangle is China's systematic breakthroughs in AI chips, frameworks, and toolchains, not just a few model parameters.

The choice is clear: either continue dancing within someone else's framework, or build your own stage to perform.

Original Link: https://www.ithome.com/0/982/221.htm