DeepSeek and its R1 model have been the talk of the town when it comes to generative AI. According to – admittedly sparse – public information, the R1 model is as good or even better than rival generative AI models, while consuming less resources, being cheaper to train, and providing the software on an open source basis. At the same time, at least according to an article in the Financial Times, OpenAI “says it has evidence China’s DeepSeek used [OpenAI’s] model to train” its own application. OpenAI executives have made some noises about possibly suing DeepSeek over… uh… something? Given the legal swamp that is IPR rights in machine generated content, that might be easier said than done.
That said, if DeepSeek was going to get into trouble over something, it might be infringement of patents, which is to say the core technology and technical solutions behind on the one hand OpenAI’s and on the other hand DeepSeek’s AI models. A patent gives its owner the right to prohibit others from using the protected invention in those jurisdictions in which the patent is registered. We decided to have a look at what kind of patents DeepSeek and OpenAI have been filing for.
As an important disclaimer, patent applications can be, and often are kept secret for up to 18 months from when they are filed, and sometimes applicants file patents not in the name of the parent company but its subsidiaries or individual inventors. In other words, our findings are more of a brief surface level exploration rather than an exhaustive deep dive into the IPR protected by these companies. Any conclusions we present are more in the nature of educated guesses than firm legal opinions.
DeepSeek has a Small and Defensive Patent Portfolio
DeepSeek made a big splash when it entered the world stage. In contrast, their public patent portfolio is very modest, with four granted patents and ten pending applications in China. They cover things like RDMA (Remote Direct Memory Access) communication methods between multiple processing units; data compression storage processing system; distributed deep learning training and resource management; and planning paths in a network topology.
If you look at DeepSeek’s patents, you are first struck by the territorial scope. All of these patents are only granted in China, although the latest applications could still be extended to other countries (we will know more in late February, when their priority date ends).
Furthermore, most of them do not even relate to the core machine learning technology that you would expect to see in AI related patents. Overall, this looks like a purely defensive portfolio, covering the very core markets and domicile of the company in China, with little ambition to defend the company’s competitive advantage by asserting patents against competitors.
Other AI Companies Aren’t Actually That Different
DeepSeek shouldn’t be examined in isolation. The picture becomes a little clearer when you look at the patent portfolios of other generative AI companies like OpenAI, Mistral AI and Anthropic.
OpenAI’s portfolio of public patents and applications is shockingly small for such an immensely valuable company – 12 applications and two granted patents, most of them in force only in the USA. More will surely come as already filed applications become public, but this is still a very modest and very defensive portfolio for a technology company. Indeed, OpenAI states on its own website, under “Our Approach to Patents”, that they pledge to “only use our patents defensively”, i.e. they won’t assert patents against others so long as they are not attacked. Mistral AI and Anthropic, on the other hand, seem to have no public applications or granted patents at all.
What we can conclude from this is that patents do not seem to be a relevant moat (barrier for others to seize a company’s competitive advantage) for these generative AI companies. In conclusion, we are none the wiser about OpenAI’s threats of lawsuits against DeepSeek – neither company’s patents are really geared for attack either in terms of their number or territorial scope. It seems more likely that OpenAI will have to answer the competition posed by DeepSeek by providing better, cheaper services and products than their Chinese competitors. That might not be such a bad thing.
