Global development
In recent years, several different GenAI models have been developed (see previous chapters). All identified GenAI patent families were assigned to five different model types based on information from patent abstract, claims or title. However, it must be noted that a large proportion of all GenAI patent families do not fit into any specific model type. Many GenAI patents do not include keywords about the specific model used in the patent abstract, claims or title, but instead focus on describing the use case for the patent and only give a generic description of the GenAI processes used. This makes it difficult to find and map the patents to the five core GenAI models, which also have some overlaps in content. As a result, only about 25% of all GenAI patent family publications since 2014 can be mapped to one of the five models.
Among the core models of GenAI, many patent families belong to the category of generative adversarial networks (GANs). Between 2014 and 2023, there were almost 9,700 publications of patent families in this model type with almost 2,400 patent families published in 2023 alone. Variational autoencoders (VAE) and large language models (LLMs) are the second and third largest models in terms of patent families with around 1,800 and 1,300 new patent families between 2014 and 2023 (Figure 21).
In terms of patent growth, patent families assigned to GANs show the strongest increase over the past decade. However, patent growth of GAN patent families has slowed down and GAN patent families have increased only moderately over the last three years. A similar slowdown in patent growth can be seen for patent families for VAE models and autoregressive models. In contrast, diffusion models and LLMs show much higher growth rates over the last three years, with the number of patent families for diffusion models increasing from 18 in 2020 to 441 in 2023 and for LLMs increasing from 53 in 2020 to 881 in 2023. The GenAI boom caused by modern chatbots such as ChatGPT has clearly increased research interest in LLMs.
Over the past decade, patent growth was weakest in autoregressive models (Figure 22). However, it should be noted that there is some overlap in the content of the different GenAI models and some patent families belong to more than one model. For example, there is overlap between autoregressive models and large language models, as LLMs are by definition autoregressive models, but not all autoregressive models are LLMs. As patent terminology changes over time, it is likely that due to the strong increase in popularity of large language models, newer patent families are more likely to use the term large language model instead of autoregressive model in patent titles, claims and abstracts. This would explain the weaker growth dynamics in patents for autoregressive models.
Top patent owners
When analyzing the five identified key GenAI models, Tencent tops the list for decoder-based LLM patent families, followed by Baidu. Baidu and Tencent are also the leaders in diffusion models among companies worldwide. Ping An Insurance Group has a comprehensive GenAI patent portfolio, with many patent families in all five GenAI models. State Grid is the leader in terms of GAN patent families (Table 4).
Alphabet/Google has recently significantly increased its number of LLM-related patent families and is the world’s number two for VAE models. IBM is the leader in VAE patent families and number two in GAN models, behind Baidu.
On the level of the five GenAI models, the Chinese Academy of Sciences is leading among research organizations in GAN models, VAE models and autoregressive models (Table 5). Zhejiang University is the frontrunner in patent activity in diffusion models. Tsinghua University has published the most patent families for LLMs.
Key locations of inventors
On a country level, China dominates in all five GenAI models in terms of patent families. The Chinese lead is particularly pronounced in diffusion models, where China has published more than 14 times as many patent families since 2014 as the second largest inventor location, the US (500 patent families compared to 35 patent families). China also has a very high global share of patents in the field of autoregressive models (Table 6).
The US is particularly strong in VAEs and LLMs. A large part of the GenAI patents from the Republic of Korea fall into the category of GAN models. For Japan, it is noteworthy that the vast majority of GenAI patent families cannot be linked to any of the five core GenAI models.
In India, GAN patent families accounted for a relatively high proportion of total GenAI patents. In the UK, many patent families belong to the GAN and VAE categories. Germany has a relatively strong research presence in GAN and VAE models.
