There is an old joke in pharmacology: the first pill costs $100 million, but the second pill costs 99 cents. A significant technological advance carries a huge, upfront cost and risk. China’s DeepSeek’s new Large Language Model R1 represents a case study of how second-mover companies can circumvent upfront research costs.
DeepSeek’s success hit US tech companies hard because investors recognize this second-mover practice and know it is a real threat to American technological dominance. It is dangerous not just because DeepSeek’s R1 is a cost and energy-efficient alternative to OpenAI and other major AI players.
The concern is that the Chinese model of IP theft creates new and severe economic risks. For years, the Chinese government has been carrying out a systematic campaign of such theft, including the cyber-campaign Salt Typhoon, positioning Chinese companies as cheaper alternatives to American competitors.
DeepSeek depended on data from Open AI’s GPT-4 to train its own R1 chatbot, a process called distillation where student AI models interrogate parent models, mimic their logic, and “suck” their knowledge from them. As a matter of law, distillation might count as IP theft, though the question has not been tested in court. OpenAI alleges that DeepSeek used OpenAI’s APIs to perform distillation, which would violate the API’s terms of use, and is investigating filing a court case.
Although analysts have focused on the Chinese theft of military applications, DeepSeek shows how the problem also impacts civilian tech. From quantum technology and robotics to biotechnology and AI, China steals secrets in various sectors, the US and other Western intelligence agencies assert. According to a 2017 US Trade Representative report, Chinese theft of American IP costs between $225 billion and $600 billion annually.
Of course, the Chinese IP threat may be overstated. Some analysts believe Republicans exaggerate the numbers for political purposes. DeepSeek did innovate. R1 produces results similar to those of the major AI players at a fraction of the cost, with much lower energy use, and with few costly Nvidia chips.
OpenAI, Anthropic, Google, and other major firms poured hundreds of millions of dollars (in fact, tens of billions, after all the bills are settled) in upfront costs, building cloud-computing infrastructure and hiring legions of AI engineers. By contrast, R1 is a project that looks at OpenAI and bootstraps it.
Companies such as Deep Seek are a part of the life cycle of any technology. They reduce the cost of training the system and create distilled models that meet consumer needs at a much lower price. We want companies to develop efficient low-cost alternatives. That is how a market is supposed to work; it is part of the reason we have personal computers in our pockets.
But the central purpose of IP law is to protect companies who fund the upfront research. A company that pours billions into a research project deserves the opportunity to profit from that investment. We want to make sure that companies get compensated for their research to motivate them. Spending 11 figure sums on research and development, only to have a Chinese-backed startup swoop in and steal it, stamping out an alternative for cents on the dollar (or, in the case of DeepSeek, giving the code away outright), looks like a worst-case scenario.
DeepSeek’s goal is not profitability: it is undercutting the market and hurting US AI companies. That coincides with Beijing’s economic strategy. If DeepSeek ripped IP away from US firms, leaving them stuck with the bill for the research costs while winning a decent share of the profits, the next generation of US technology investors could be spooked and refuse to fund future research. For China, this outcome is a win-win. For the US, it is a lose-lose.
Joshua Stein is a Technology Policy Analyst at the Software and Information Industry Association. His work focuses on technology policy and economics.
Bandwidth is CEPA’s online journal dedicated to advancing transatlantic cooperation on tech policy. All opinions expressed on Bandwidth are those of the author alone and may not represent those of the institutions they represent or the Center for European Policy Analysis. CEPA maintains a strict intellectual independence policy across all its projects and publications.
2025 CEPA Forum Tech & Security Conference
Explore the latest from the conference.
