The Influence of China’s AI Models on Silicon Valley’s Tech Surge
Jensen Huang recently stated that: “China is going to win the AI race.” He later moderated his position, indicating that the US’s competitor was merely “nanoseconds behind,” and emphasized the importance of America emerging victorious “by racing ahead and winning developers worldwide.” Huang’s initial assessment lacks depth, and the reasons behind his negative stance on the US are evident. He has dedicated the past year to asserting that America gains from his company’s sale of chips to the Chinese market, even in light of Washington’s increasing export controls. However, he is justified in his concern that the competition for developers might already be fading. In recent weeks, a nuanced change has become more evident. For months, speculation has been brewing that low-cost, open-source Chinese AI models might attract global users away from US offerings. However, it seems they are also subtly gaining favor in Silicon Valley.
Chamath Palihapitiya recently stated that a company he collaborates with has transferred significant workloads to Kimi K2, developed by Beijing-based Moonshot AI. The open-source model, he remarked, is “frankly just a ton cheaper than OpenAI and Anthropic.” Shortly after Brian Chesky acknowledged that he did not incorporate his travel app with OpenAI’s ChatGPT due to the connective tools not being “quite ready.” According to Chesky, Airbnb’s new service agent relies on over a dozen different AI models. They are “relying a lot” on Qwen lineup: “It’s very good.” “It’s also fast and cheap.” His remarks carry significant weight, particularly in light of Chesky’s close personal connection with Sam Altman. The list of overt acknowledgments continues to expand. Thinking Machines Lab, the startup established by OpenAI’s former Chief Technology Officer Mira Murati, announced in a blog post that its recent research was inspired by and developed from the efforts of Alibaba’s Qwen3 team.
However, what could be even more intriguing are the nuanced ways in which this shift seems to be establishing itself. Cursor, an AI coding leader with significant buzz and a valuation of approximately $10 billion, unveiled a new version of its assistant last month. Speculation on the internet has increased, suggesting that it may have been developed using a Chinese AI tool such as DeepSeek. This theory gained traction after a tech investor noted that the system changed its inner monologue to Mandarin during his usage. Another prominent US-based company, Cognition AI Inc., valued at approximately $10 billion, seems to have developed its new coding agent using a foundational model from Zhipu AI, recognized globally as Z.ai. Following the observations made by social media investigators, the Beijing-based firm appeared to validate their conclusions in a tweet, stating it “highlights the positive impact and value of open-source contributions to the ecosystem.” The two US companies did not respond promptly to my emailed requests for comment. However, the permissive licensing nature of Chinese AI models allows companies to freely build products on top of them. Data confirmed that Chinese models have surpassed their US counterparts in cumulative downloads by developers. The transition unfolded gradually, then culminated abruptly: In early 2024, Meta Platforms Inc.’s Llama achieved 10.6 million downloads, while Alibaba Qwen lagged with a mere half a million. As of last month, Qwen reached a remarkable 385.3 million cumulative downloads, surpassing Llama, which recorded 346.2 million.
Derivative systems built on Qwen now represent over 40 percent of new language models featured on Hugging Face, whereas Meta’s share has decreased to 15 percent. Geopolitical concerns have arisen regarding international users gravitating towards Chinese AI tools, particularly the risk of Communist Party ideologies being embedded in the outputs. However, for developers eager to launch products, particularly in the coding and software development fields, these risks tend to diminish in importance — especially when considering price and performance. Open models are available for download, allowing for fine-tuning and local execution, which addresses concerns regarding content and data privacy. It is too early for Huang to announce a victor. The US maintains distinct advantages regarding access to cutting-edge chips and computing power, which are essential components in the development of advanced systems. However, Beijing’s emphasis on low-cost and open-source initiatives is undeniably drawing in developers, who are essential to AI innovation. If Washington genuinely seeks to prevail in the long term, it must begin by questioning why Silicon Valley is already shifting allegiances.
