When US Politicians Call to Ban Chinese AI, They Should Fear Their Own Lagging Progress

If the US is really afraid of China's open-source AI, it should first ask itself: why are CEOs of top global AI chip companies actually speaking up for Chinese models?

Jensen Huang's remarks are worth chewing on repeatedly. He wasn't being polite; he was basing his comments on NVIDIA's balance sheet. NVIDIA's FY2024 data center revenue exceeded $47.5 billion, with the China region accounting for about 20%. Any calls for a "ban" mean NVIDIA loses this market. But more critically, Jensen Huang sees a paradox: the more the US bans, the stronger China's open-source models become.

Look at DeepSeek-V2. This model scored 90.2 on the MMLU benchmark, 1.5 percentage points higher than Meta's Llama 3.1 70B, while training costs were only 1/3 of the latter. The iteration speed of the Chinese open-source community is astonishing—from Qwen to Baichuan, then the Yi series, every new model is sprinting towards SOTA. This isn't "catching up," it's "parallel running."

Compare the two routes:

US Route: Closed source + compute monopoly. OpenAI, Google, Anthropic lock models behind APIs, maintaining advantages through compute barriers. But costs are too high—GPT-4o inference cost per call is ~$0.03, while an equivalent performance open-source model costs only $0.005.

China Route: Open source + ecosystem win-win. DeepSeek, Zhipu, Alibaba Tongyi all choose open source, even releasing weights. Developers can deploy, fine-tune, and commercialize for free. This directly lowers the barrier to AI applications, attracting global developers into the Chinese ecosystem.

The competitive landscape is reversing. According to IDC data, in Q1 2024, open-source models accounted for 47% of global AI model deployments, a 12 percentage point increase year-over-year. And the GitHub Star growth rate of Chinese open-source models is 2.3 times that of US open-source models. This isn't a "threat," it's a "replacement."

What Jensen Huang is truly worried about isn't the Chinese models themselves, but that US politicians' calls for "bans" are creating an information cocoon. If US enterprises cannot use Chinese open-source models, they are abandoning the world's most active open-source community. US AI companies will lose iteration speed because the feedback loop of closed-source models is far slower than that of open-source communities.

More fatally, bans will accelerate the maturity of China's independent compute ecosystem. Huawei Ascend 910B's compute power has reached 60% of H100, and models like DeepSeek have improved training efficiency to 85% of H100 through optimization. Once Chinese models reach scale in the open-source community, the "moat" of US chips becomes a "besieged city."

My prediction: By the end of 2025, Chinese open-source models will comprehensively surpass US closed-source models of equivalent size on key benchmarks. At that point, US politicians will realize that what they should fear most isn't Chinese AI, but the high wall they built with their own hands.

Original Link: https://www.ithome.com/0/980/294.htm