DeepSeek will not chip away at Silicon Valley’s sway

The global shock of the Chinese start-up DeepSeek has fueled debates on the implications for artificial intelligence development and geopolitics. Its rise was an inevitable result of a globalized yet increasingly polarized AI ecosystem. 

And although it has challenged United States dominance in artificial intelligence, Silicon Valley will not stand still.

There are three myths that must be dispelled. Immediately after the reasoning model DeepSeek-R1 launched, commentators claimed that American semiconductor export controls, which are critical for restricting China’s access to the most advanced AI chips, had failed. 

But they were not meant to stop China cold. They were designed to create a ceiling which is now visible. 

While the US is probing whether DeepSeek had access to smuggled chips, most of its training semiconductors were Nvidia H800s, which were designed specifically for the Chinese market after 2022’s first round of export controls. They are now restricted. 

Chip capacities

DeepSeek’s success with limited computing power may necessitate the adjustment of export controls. But existing measures will soon start to bite, hindering future progress as China works to mature its domestic chip capacities.

Another common misconception is that model export controls are the solution to curbing Beijing’s plan of AI ascendency. 

In response to DeepSeek’s advances, US Senator Josh Hawley introduced the “Decoupling America’s Artificial Intelligence Capabilities from China Act” that would ban the export and import of all open models to and from China. 

Ensuring that these are not being downloaded by Chinese users would likely require burdensome, universal ‘know your customer’ verification processes that would stifle domestic innovation and hamper American leadership in open-source AI.

It would also fail to mitigate misuse risks. Unlike physical goods, data like model weights can be easily transferred or hacked. It does not need to be smuggled across borders. 

With the Chinese start-up and even more efficient new models from the Alibaba-backed Qwen, China will rely less on American open-source versions. Imposing burdensome US restrictions would only drive developers towards universally accessible Chinese AI templates. 

While DeepSeek has faced criticism for being censored, developers worldwide do not pledge allegiance to flags – they will trend towards the cheapest and best model.

Open models

If the US wants to maintain global leadership in AI, it should instead encourage responsible release policies that keep open models open while ensuring they are responsibly deployed.

Tiered release strategies coupled with external audits could help prevent misuse while ensuring safe distribution.

Claims that Silicon Valley’s “moat” of money invested into “chips, talent and energy infrastructure” is gone are yet another misconception. Before DeepSeek, those investments were enough to secure US dominance.

Still, its relatively cheap success raised concerns. Reports initially stated that DeepSeek had accomplished performance gains at approximately US$6 million, or roughly 1% of the cost of training Meta AI’s Llama model.

But this number only relates to the GPU, or graphics processing unit, cost of a single pre-training run, not the entire model. And DeepSeek’s advances would not have been possible without Silicon Valley. 

Google initially advanced the “mixture of experts” architecture that DeepSeek innovated on, while OpenAI’s o1 model pioneered the use of test-time compute to increase efficiency.

Chinese researchers also reportedly “distilled” ChatGP, using it as a shortcut to generate training data.

DeepSeek may have laid a plank over Silicon Valley’s moat, but US Big Tech will use that plank to dig its moat wider. American companies will be quick to assimilate DeepSeek’s open-source advancements, once again reaping the benefits of their superior computing resources. 

Domestic rewards

And if European companies do the same, they may shake up the landscape as well.

DeepSeek’s success was not an anomaly, but the inevitable product of a global, competitive information system. This success indicates changes to development, as China has demonstrably achieved its goal of “world-leading levels” of AI by 2025, at least in open source.

Most notably, DeepSeek shows that Beijing’s investment in its talent is paying off.

China has created more than 2,300 new undergraduate AI majors since 2018. As the United States becomes a less attractive place for Chinese expertise, this investment will reap further domestic rewards. 

The country’s equity market is “heavily driven by public sentiment,” causing many companies to “ride the attention wave” and implement new technologies in products regardless of actual effectiveness. 

Local governments also face warped incentive structures for new technology rollouts. Some of DeepSeek’s deployments have been described as “publicity stunts.” 

How China closes the implementation gap will be telling for the future of the AI race.

Despite Beijing’s efforts to build a parallel open-source ecosystem, American and Chinese open-source ecospheres are heavily intertwined. 

Silicon Valley

Even if each country only uses its own open-source tools and models, developers will continue to draw inspiration from each other. Although DeepSeek laid a plank over Silicon Valley’s moat, Silicon Valley is already digging that moat wider. 

Computing power is not everything, but Silicon Valley’s bigger shovel will continue to be an advantage. Even so, the United States cannot assume it will maintain superiority. 

For now, though, the question is not if the US can compete. It is whether it will remember to dig and develop both rapidly and responsibly.

Emmie Hine is a research associate at Yale Digital Ethics Center and a PhD candidate in Law, Science and Technology at University of Bologna.

This edited article is republished from East Asia Forum under a Creative Commons license. Read the original article here.

The views and opinions expressed in this article are those of the author and do not necessarily reflect the official policy of China Factor.