Is America Poised to Overreact to DeepSeek?

The last several days have been dubbed a “Sputnik moment” for Americans, who are discovering that foreign software engineers are much further ahead than previously thought. On January 20th, the Chinese company DeepSeek announced the release of their newest large language model (LLM) “DeepSeek R1.” Reportedly trained for just $6 million, DeepSeek R1 achieves capabilities that more than rival the output of American AI giants like Anthropic, OpenAI, Meta and Google, whose models have required billions in investment. 

DeepSeek RI combines several innovative training techniques to create a model whose performance serves as proof of concept for exponential efficiency gains, even though the techniques are not especially novel in themselves. To illustrate the achievement, as one retired Microsoft engineer put it, “It’s akin to building a Ferrari in your garage out of spare Chevy parts…and it’s really just as good as a regular Ferarri.” (You can read up on mixture of experts, chain of thought, and reinforced learning versus supervised fine tuning for more information on the technicals.)

DeepSeek is a relatively new startup founded in 2023 by Liang Wenfeng, who is no newcomer to AI. He previously founded a notable Chinese hedge fund, High Flyer, which used AI to trade stocks, seemingly in an effort to replicate the success of the legendary algorithmic trading firm Renaissance Technologies. After piloting a few other LLMs with the financial backing of High Flyer, DeepSeek came out with R1. 

Why was this all over the news, and why does anyone care? There are a couple of reasons. For one, it showed that while the U.S. has traditionally excelled in software development, China can match and even surpass American tech giants on important AI benchmark tests. The fear is that AI is the new atomic bomb for 21st-century warfare, and America is in an arms race. Two, news of R1’s efficiency caused the largest single-day stock market loss in history, with Nvidia losing $589 billion in market capitalization because some fear a drop in demand for GPUs, which are used to process the data used to train the models. 

The commotion prompted a lot of online debate about America’s AI strategy, with the White House investigating nationality security issues and commentary on America’s export controls over chips. The U.S. has implemented all sorts of policies aimed at restricting foreign access to particular tech, from the aforementioned export controls, the CHIPS and Science Act, the Clean Network Program, and various other protections and subsidies to go along with proposed tariffs. Policymakers are not seeing the long game.

What is often lost in these debates is the importance of competition and the fact that restrictions can compromise long-term viability. America’s current direction shows shortsightedness, which in many ways bears unfortunate similarities to its approach to the steel industry. 

Right now, America still dominates the software market. Silicon Valley is so rich that if it were a country, it would be on the list of wealthiest nations. That should not be taken for granted. In the aftermath of WWII Americans had a virtual monopoly on steel production, in no small part due to the destruction of Europe. For a time, the “Steel City” of Pittsburgh was one of the richest cities on Earth. Considered strategically vital, the steel industry has long been heavily protected by tariffs, subsidies, and special carve outs. Throughout the 20th century, American policymakers shielded domestic steel producers from foreign competition, believing this would preserve industrial dominance. As Scott Lincicome has pointed out, “U.S. steel companies and workers since the 1970s have probably received more government assistance than any industry in the country.” 

However, this protectionism ultimately backfired: without the pressure to innovate and improve efficiency, U.S. steel companies stagnated while global competitors, particularly in Japan and Europe, developed superior production methods and inefficient American firms lagged behind. American steel lost its technological edge. A similar risk looms over AI today. If the U.S. relies too heavily on restrictive policies rather than allowing robust homegrown innovation and efficiency, it may find itself outpaced by competitors who are forced to innovate. 

China is pouring billions into its domestic chip production capabilities and is already utilizing existing hardware to effectively bypass these restrictions. With DeepSeek specifically, High Flyer bought huge bulks of Nvidia A100s before the U.S. government imposed export restrictions, so it is difficult to say if the restrictions would have prevented R1’s rollout. 

In any case, DeepSeek’s advancement shows that AI is not solely dependent on access to massive computational resources. To the extent that China distorts and impairs its own domestic AI market, that is a them problem. The double-edged sword is that the export controls limit U.S. companies like NVIDIA from accessing a lucrative Chinese market, potentially reducing their revenue and slowing domestic R&D investments. 

R1 is open source, meaning more developers can build on it without prohibitive costs. This makes the technology more accessible and lowers the barrier for entrepreneurs to innovate. Moreover, the increased efficiency has many downstream benefits, such as reducing demand on the electrical grid. Firms will have more R&D capital available since they won’t have to spend all their money on training. Since R1 is open source, it makes the know-how publicly available for researchers and developers. It is worth keeping in mind, however, that the $6 million price tag for R1 is likely a conservative estimate that discounts DeepSeek’s R&D and components cost. Still, it certainly will be cheaper to employ.

History teaches that overreaction and heavy-handed intervention often backfire. U.S. carrot-and-stick tech policies risk undermining the very goals these measures are meant to achieve. DeepSeek’s innovation should be seen as both a wake-up call, but also a tempered cause for optimism.

The post Is America Poised to Overreact to DeepSeek? appeared first on The Beacon.

READ MORE