The Roscongress Foundation’s review shows how large AI language models are transforming from custom-made technology products into standardized tools

Large AI language models are moving from the category of unique technological products to the category of standardized tools under the influence of a number of factors. This is stated in the analytical review ‘Large language models of generative AI: from custom-made products to standardized goods’ prepared by the experts of the Roscongress Foundation.

The US–China rivalry in the field of artificial intelligence has largely contributed to this. Sanctions on the supply of the most powerful chips to China have stimulated innovation, forcing AI developers to optimize the architecture of models. China is implementing a long-term AI development strategy based on a phased approach with clear targets. The outcome of the Chinese development model is the formation of the AI sector, which includes about 4,500 companies. Unlike the US model, where private venture capital funds play a key role, Chinese technology companies receive their main funds through a system of state development institutions and large banks with government participation. The Chinese model is also characterized by the active participation of regional authorities, which create special technology development zones and provide significant tax incentives. 

The effectiveness of this approach is confirmed by statistics: while there are fewer companies in the AI sector – 15% of the global total versus 34% in the US – Chinese developers have created a comparable number of language models – 36% versus 44% in the US. Together, companies from these two countries account for 80% of these models. This demonstrates the higher impact of the state-centred funding model in the context of critical technology development.

The Chinese have unveiled DeepSeek-V3 and DeepSeek-R1 models that rival OpenAI products. The appearance of DeepSeek-V3 on the MoE architecture has set a new price level for companies that provide access to LLM – a system of billions of computational blocks or ‘neurons’ trained on huge arrays of text data. Using the most advanced models from OpenAI costs an order of magnitude or two more than similar quality models from DeepSeek. In addition, the Chinese company has published its models in free access under the MIT licence, which allows their free use and modification for any purpose, including commercial ones. Derivative models based on the Chinese company’s developments can be licensed under any conditions.

Thus, the business model of OpenAI and other AI providers is being jeopardized. An additional factor setting the trend towards cheaper LLM usage is the release of so-called ‘distilled’ models, which use fewer parameters while retaining most of their useful properties.

The development of large language model architectures and knowledge distillation techniques are significantly lowering the barriers to entry in the generative AI field, facilitating the transition from a market dominated by a very small number of companies to a more competitive environment.