Mr Trump’s hosting the day after the launch of “the largest AI infrastructure project in history” shows he appreciates the potential. But so does the rest of the world and, most of all, China. Even Mr. Trump was attending his inauguration. A Chinese firm suddenly released the latest impressive large language model (LLM).
China’s catch-up is surprising because it was so far behind—and because the US went out of its way to slow it down. Joe Biden’s administration fears that advanced AI could secure Chinese Communist Party (CCP) military dominance. So the US has curbed exports to China of the best chips for training AI and cut off China’s access to many of the machines needed to create alternatives. Behind its protective wall, Silicon Valley has swaggered. Chinese researchers devour US papers on AI; Americans rarely returned the compliment.
Yet China’s most recent advances are upending the industry and embarrassing U.S. policymakers. The success of the Chinese model, combined with industry-wide changes, could turn the economics of AI on its head. America must prepare for a world in which Chinese AI is breathing down its neck.
China’s LLMs are not very good. But they are cheap to make. QWQ, owned by e-commerce giant Alibaba, was launched in November and is less than three months from being the top US model. DeepSeek, whose creator was spun out from an investment firm, is ranked seventh by one benchmark. It was apparently trained using 2,000 second-rate chips—versus the 16,000 first-rate chips for Meta’s model, which DeepSeq beats on some rankings. The cost of training an American LLM runs into the millions of dollars and rising. DeepSec’s owner says it spent under $6M.
American firms can copy DeepSec’s technologies if they wish, as its model is open-source. But cheap training will change the industry at the same time as model design is evolving. China’s opening-day release was DeepSec’s “Reasoning” model, designed to compete with the state-of-the-art offering by OpenAI. These models talk to themselves before responding to a query. This “thinking” produces an even better answer, but also an even better answer, but an even better answer, but an even better answer, but an even better answer, but it Also gives a better answer, but also gives a better answer, but also gives a better answer, but also gives a better answer, but also gives a better answer, but also gives a better answer but also gives a better answer, but uses more power Does. As the quality of output increases, costs mount.
The result is that, just as China has driven down the fixed costs of building models, so the marginal costs of querying them have been rising. If those two trends continue, the economics of the tech industry will be turned upside down. In web search and social networking, copying a giant incumbent like Google involves huge fixed costs of investment and the potential to incur huge losses. But the cost per search was infinite. This and the network effect inherent to many web technologies have created winner-take-all markets.
If good-enough AI models can be trained relatively cheaply, the models will proliferate, especially as many countries are desperate to have their own. And a higher cost-per-query might similarly encourage more built-for-purpose models that deliver efficient, specialized answers with minimal queries.
Another consequence of China’s success is that the US faces asymmetric competition. It is now clear that China will innovate around constraints such as the lack of the best chips, whether by efficiency gains or by compensating for the absence of high-quality hardware with greater volume. China’s homegrown chips are getting better, including those designed by Huawei, a technology firm that achieved widespread adoption of its telecommunications devices a generation ago with a cheap-and-clever approach. Had done.
If China remains close to the border, it may be the first to leapfrog superintendence. Should that happen, it could yield more than just a military advantage. In a superintelligence scenario, winner-take-all dynamics may suddenly reassert themselves. Even if the industry stays on track today, widespread adoption of Chinese AI around the world could give the CCP enormous political influence, at least as worrisome as the concern that a Chinese-owned The video-sharing app that was futuristic in the US. remains unclear.
What should Mr. Trump do? His infrastructure announcement was a good start. The US should clear legal barriers to building data centers. It should also ensure that it is easier to hire foreign engineers, and reform defense procurement to encourage faster adoption of AI.
Some argue that they should also repeal chip-industry export restrictions. The Biden administration acknowledged that the ban failed to include Chinese AI. Yet that doesn’t mean it accomplishes nothing. In a worst-case scenario, AI could be as deadly as nuclear weapons. The US would never send components to its adversaries for nukes, even if they had other ways to obtain them. Chinese AI would certainly be stronger still if it got easy access to the very best chips now.
agencies or agency
More important is to bring back Mr. Biden’s draft “AI proliferation rules,” which would control which countries have access to American technology. It’s designed to force other countries into the US AI ecosystem, but the tech industry has argued that, by lying on the sidelines, the red has done the trick. Tape, it’s the opposite. Once again America will be put in front.
FOR SUBSCRIBERS ONLY: Sign up to our Weekly to see how we design each week’s cover cover story newsletter,
© 2025, The Economist Newspaper Limited. All rights reserved. From The Economist, published under license. Original content can be found at www.economist.com
) CCP (T) Donald Trump (T) CHAT (T) AI Model
#Chinese #catching #creating #dilemma #Donald #Trump
