Companies from Silicon Valley and Europe, including DoorDash, Siemens, and Airbnb, are increasingly adopting Chinese AI models to reduce costs and dependence on US frontier labs. This shift is largely driven by the significantly lower cost of Chinese open-weight models, which are between 10 and 60 times cheaper than their proprietary US counterparts, according to Vipul Ved Prakash of Together AI. Flo Crivello, founder of a company that switched, hailed it as "transformative," saving millions of dollars and improving performance.
While cost is a primary driver, geopolitical considerations have also played a role, particularly in Europe. The Trump administration's export controls on Anthropic's Mythos and Fable models last month highlighted the risks of over-reliance on US technology. This spurred companies like German HR startup Timebutler to offload tasks to models like Alibaba's Qwen to reduce dependence on US labs, even after the ban was overturned. Per Roman of Bullhound Capital noted that while China was once the main worry, the US is now a bigger concern in Europe.
Another significant draw of Chinese models, such as those from DeepSeek and Z.ai, is their open-weight nature, allowing companies to host them on their own servers and fine-tune them for specific data and uses, offering more control. Sam Bresnick of Georgetown University highlighted that many workloads don't require the most premium models, making the "cheaper, faster models" from China a viable and attractive option. This explains why Chinese AI models have rapidly surpassed US rivals in token consumption this year, according to OpenRouter data. The performance gap is also narrowing, with Z.ai's GLM-5.2 in June being praised by Silicon Valley technologists, including Marc Andreessen, co-founder of Andreessen Horowitz, for matching or beating American big lab public AI models.
Despite the increasing adoption of some Chinese models, the US Commerce Department's preliminary cyber-focused evaluation found that Moonshot's Kimi K3 performed significantly below leading US frontier AI models in that domain. However, other assessments, such as data cited in Jefferies' report, estimated Moonshot AI’s open-source Kimi K3 to offer about 95% of the performance of Anthropic’s Claude Fable 5. Chris Wood, Jefferies’ Global Head of Equity Strategy, warned of potential "massive capital destruction" in the US AI sector due to the competitive threat from cheaper Chinese open-source models, especially as top Chinese models processed nearly five times more tokens (36.39 trillion vs. 7.39 trillion) on OpenRouter during the week ended July 19. The Silicon Data LLM Token Expenditure Index also shows a 25% drop in average token prices since late May, raising concerns about the profitability of large language models.
