There was something about the artificial intelligence (AI) hype that aroused my skepticism; that it absolutely had to have, not just gigabytes or terabytes, but petabytes and exabytes of data storage; that it needed millions of the highest-end computer chips selling for $40,000 each; that it required hundreds of billions of dollars of investment. Was it not possible to do with less? It’s just software, after all. Couldn’t diligent application of innovative brainpower make it much more efficient over time, perhaps even quickly, the way vigorous efforts in fields powered by math and computer programming have before — indeed, in the field of AI itself? The statements about “generative AI”, and large language models such as OpenAI’s ChatGPT, have seemed hyperbolic.

More than 10 months ago, in a March 19 article, I used an example to show that ChatGPT and others like it were not that big a step beyond Googling. I then said:

“But if it could have been done as little more than a small step beyond Googling, as in my example, does it need all that advanced equipment and massive quantities of data to work? Isn’t it just possible that it could work nearly as well, and just as well for most applications — even including some of the most advanced applications — with equipment that is not on the very leading edge of technology but somewhere closer to the trailing edge?”

We have just learned the answer to that question: Yes, it can. A Chinese generative AI chatbot called DeepSeek has been all over the news because it does the job at a much-lower cost.

Was it inevitable that it would be a Chinese company that would accomplish this task first? No, it wasn’t. But it was not unlikely. Why?

To understand why, consider a quote about the buildup of the polysilicon industry in China in the 2010s. The quote is from David Fishman, China-based specialist for Hong Kong energy consulting firm the Lantau Group. Fishman was interviewed on Jan 23, before the DeepSeek news came to light and went viral, on Chris Keefer’s excellent podcast Decouple. Fishman said:

“China was just starting to build panels, but a lot of the upstream components, especially the polysilicon, they were still dependent on imports. They were still importing polysilicon and then, I want to say it was like 2012, 2013, as Chinese panels began to flood out into the world, there was the first of the anti-dumping concerns and one of the outcomes of that was a move to limit or ban silicon exports to China. It’s very reminiscent of where we are now, with chips actually is that if we cut them off at the knees, they won’t be able to sell panels anymore and instead they turned around and built the silicon industry in like three years and then scaled it up and went on to dominate it, and now nobody produces silicon anymore except for China. … And then very quickly, if you give China a reason to pursue something, to back it with state capital and consider it a key strategic priority to diversify or eliminate that exposure to an unfriendly political adversary, they’ll do it really fast.”

China’s share of the global polysilicon production market in 2005 was essentially zero. The United States’ was 55 percent. In 2012, China’s share was still only 30 percent. But in 2023, China’s share was over 80 percent, and the US’ and Europe’s were each 5 percent.

Nothing spurs innovation like constrained resources and pressure to perform. The US’ and Europe’s curtailing of the availability of high-end chips to China and other sanctions served to constrain those resources and increase that pressure. Hence, it was not unlikely that some Chinese entrepreneur would find a way to do more with less.

In the US, by contrast, efforts are focused primarily on securing massive amounts of money, massive amounts of data storage, and massive amounts of energy. A supersufficiency of resources reduces the pressure to innovate efficiencies. It makes efficiency less necessary. This difference between the US and China may be the main reason it was a Chinese company that came up with this breakthrough in AI efficiency.

You would think the US and its allies would have learned from the polysilicon experience, and its failure to contain China’s rise through bans and sanctions. But no. Why not? Because much of the US’ actions against China are performative, to make it seem to voters that the US administration is being “tough on China”. Almost all of the actions in this sphere are focused on appearances rather than results, even if the government officials dictating the actions don’t realize it. They are caught in a straitjacket, in which actions and words that don’t have the right performative effect are immediately jumped on by opposing politicians and opposing voter constituencies, even if those voting constituencies and those politicians are poorly informed and are more preoccupied with presenting the right impression than with getting the best results.

Nevertheless, the US is likely to catch up. DeepSeek has lit a fire under it. DeepSeek’s success is being characterized in the US as a “Sputnik moment”, drawing a parallel with the scare the US experienced in 1957 when the Soviet Union launched Sputnik, the first artificial satellite to circle the globe. It made the US fear that the Soviets were ahead of them in technology. But then they greatly ramped up their efforts in space, succeeding in landing the first man on the moon in 1969.

China is a much more formidable competitor than the Soviet Union was, of course. But DeepSeek’s AI model is open source. That means others can take advantage of it in their modeling too. In the AI sphere, the US and China may ultimately run neck and neck, borrowing from each other and thus, in effect, collaborating implicitly. Let us hope, optimistically, that this is the result.

The author is a mathematician and economist with expertise in finance, energy, and sustainable development. He is an adjunct associate professor at the Hong Kong University of Science and Technology.

The views do not necessarily reflect those of China Daily.