China’s cost competitiveness in artificial intelligence (AI) large models has once again become a global focal point. A latest analysis from UBS Securities indicates that the training costs for Chinese AI models are only one-tenth of those of their overseas counterparts, and this is not driven by cash-burning subsidies, but rather by the ability to achieve healthy gross margins of 20% to 40% on API businesses. Meanwhile, Nvidia CEO Jensen Huang has publicly endorsed China’s open-source models, arguing that US companies should widely adopt them, and warning that Washington’s restrictive mindset could slow down America’s own pace of innovation.
UBS Securities China Internet sector analyst Xiong Wei stated that China’s AI large models are demonstrating significant systemic cost advantages in the global market. She pointed out that the training costs for Chinese models are currently only one-tenth of those of leading overseas firms, while inference API pricing is also just 10% to 20% of comparable overseas models. Xiong Wei emphasized that this low-price advantage does not stem from aggressive price subsidies or loss-making concessions; Chinese model developers can still achieve healthy gross margins of 20% to 40% on their API businesses, possessing sustainable global competitiveness.
Xiong Wei further analyzed that the root of the cost advantage lies in the dual improvement of training and inference efficiency. During the training phase, domestic models achieve this through architectural design innovations and algorithmic improvements in GPU utilization efficiency, combined with an active open-source ecosystem that facilitates rapid technology dissemination. During the inference phase, domestic models adopt architectures such as Mixture of Experts (MoE), where the parameter ratio activated per inference is only in the single digits to 10%, far lower than the 15% to 30% seen among US peers, significantly reducing computational load. Simultaneously, optimized computing cluster scheduling algorithms have enabled leading manufacturers to achieve GPU utilization rates exceeding 70%. Coupled with China’s cost advantages in electricity rates and data center infrastructure, these factors collectively form a structural moat.
Regarding global market trends, Xiong Wei observed a notable shift: the market has moved from an early “Token Max” model of blindly consuming tokens toward refined budget management that emphasizes return on investment (ROI). As enterprise applications deepen, some overseas companies have already begun adopting Chinese open-source models to handle specific workflows. She believes that the core bottleneck for Chinese manufacturers’ commercial expansion and inference revenue currently lies primarily in insufficient computing power supply, rather than market demand.
Kimi K3 Seen as a “Second DeepSeek Moment”
The recent release of the Kimi K3 model by Chinese AI startup Moonshot AI has become a flashpoint for market discussion. The model boasts 2.8 trillion parameters, making it the largest open-source model globally by parameter count. It focuses on reasoning, long-context programming, and agentic AI tasks, with performance in some tests approaching the flagship products of OpenAI and Anthropic. Demand surged immediately after the model’s launch, at one point forcing Moonshot AI to suspend new subscriptions.
Xiong Wei described the release of Kimi K3 as a “second DeepSeek moment,” arguing that it further validates the prosperity and commercial pricing power of China’s open-source ecosystem by narrowing the performance gap with top-tier models while maintaining a price advantage. Furthermore, compared to the globally hyper-focused text models, Chinese manufacturers are also demonstrating unique strengths in the multimodal domain, with some showing impressive financial performance, making this a monetization breakthrough worth watching in the second half of the year.
Jensen Huang: The US Should Use China’s Open-Source Models
In an interview with Axios, Nvidia CEO Jensen Huang took a stance on China’s open-source models that starkly contrasts with Washington’s position. He stated bluntly: “These Chinese models are excellent. Excellent open-source models should be used.” Huang believes the US has no reason to fear the development of China’s open-source AI. Cheaper AI models will drive more enterprises and individuals to use AI, ultimately expanding demand for Nvidia chips, data centers, and computing resources.
Huang criticized the market for previously misjudging the impact of DeepSeek and now misreading the significance of Kimi K3 again. He asserted that open-source models will not cause US companies to be “driven out of the market by China,” stating, “That scenario is impossible, the probability is zero.” Regarding US efforts to restrict Chinese AI models on security grounds, Huang raised questions, pointing out that open models do not imply the existence of so-called “backdoors,” and that enterprises can control risks through methods such as secure isolation environments. Conversely, openness allows more researchers to discover vulnerabilities and improve defenses. He warned: “If the world is left with only one model, one attack point, one source of failure, I think the world will be more vulnerable.”
However, the US government has raised questions about the technical origins of Kimi K3. White House tech policy chief Michael Kratsios accused Moonshot AI of using distillation to obtain capabilities from Anthropic’s model Fable and using restricted Nvidia GB300 chips through servers in Thailand. US Treasury Secretary Scott Bessent stated that the government supports open-source AI but will not tolerate intellectual property theft, and if Chinese models are found to involve infringement, sanctions against related companies are not ruled out. These allegations have not yet been confirmed by Moonshot AI.
Wall Street: Model Competition is Bullish for Chip Hardware
Although the debut of Kimi K3 once triggered a sharp sell-off in chip stocks, replaying the rout seen when DeepSeek R1 launched in early 2025, Wall Street analysts remain firmly bullish on the long-term prospects of AI hardware manufacturers. Bank of America stated that the release of China’s open-source models further reinforces its investment thesis favoring the memory industry. BofA analysts noted that Chinese open-source models offer highly competitive API pricing, reflecting a business model choice rather than necessarily reflecting hardware costs. When enterprises choose to download open-source models like Qwen or GLM instead of paying for Anthropic’s Claude Opus, it drives up memory capacity demand on the enterprise side.
UBS Americas Chief Investment Officer Ulrike Hoffmann-Burchardi believes that even if competition from overseas AI developers could weaken demand for top-tier US AI labs, there is no reason to expect an oversupply in the high-end GPU market. She pointed out that tech giants view AI as an existential “winner-takes-all” battle; even if short-term financial returns are low, companies will continue to purchase hardware to prevent competitors from gaining a permanent technological advantage.
China’s AI Capex Remains Prudent, Avoiding a Bubble
Regarding capital expenditure and market competition, Xiong Wei pointed out that compared to their US counterparts, Chinese cloud vendors and major internet companies are more prudent in their AI investments, placing extreme importance on ROI. This steady, methodical strategy has effectively prevented an AI bubble from forming in the Chinese market. In the long term, as inference demand increases, AI capital expenditure will still rise steadily. Addressing concerns that certain tasks with more inference steps increase token consumption, she acknowledged that this narrows the cost advantage for complex tasks, but overall, domestic models still possess significant cost-effectiveness.
In summary, China’s AI industry is reshaping the global market landscape with a highly competitive cost structure. From UBS’s data analysis to Jensen Huang’s public endorsement, and Wall Street’s sustained optimism on hardware demand, signals from all sides point to one trend: low-cost, high-performance open-source models are breaking down existing barriers in the AI field, and demand for computing infrastructure will not shrink but will instead expand further due to the proliferation of applications.