Summary of Key Points
Over the past two years, our discussions about the Sino-US AI competition have mainly focused on the press releases of star companies like OpenAI and Zhipu, as well as the industrial planning documents of the two governments to discern trends. In essence, we have been watching the “internet celebrity movements” and “policy intentions,” without truly understanding the underlying foundations of the industry. This time, Rand Corporation put in a lot of effort to identify 26,000 AI-related companies, using large models as a strict filter to eliminate those that merely repackaged and integrated third-party models. In the end, they selected 1,181 companies that actually developed AI independently: 743 in the United States and 438 in China. The conclusion drawn from this analysis completely overturns the prevailing narrative: Sino-US AI is not competing on the same track. Instead, both started from the same technological foundation but have grown in vastly different industrial environments, evolving into two almost unrelated development paths. There is no clear leader or loser; each has adapted to its own environment.
---
Detailed Breakdown and Interpretation
1. The AI landscape we’ve observed over the past two years was actually just a 10% warm-up
Everyone’s attention has been drawn to the star companies selling model APIs—competing in terms of model size, performance scores, and the speed of launching new products. However, the real data reveals something quite different:
- Regardless of country, only about 10% of companies actually make money by selling model APIs, totaling around 100 companies in total.
- The remaining 90% of US companies and 75% of Chinese companies focus on practical AI applications. Their customers don’t care whether the underlying technology uses GPT or open-source models; they only care whether the AI can get the job done and reduce costs.
Interestingly, the proportion of companies that start from scratch training large models is the same in both the US and China, at 19%. The common misconceptions, such as China rapidly catching up with large models while the US is stagnating, or the US focusing on Transformers while China is secretly developing new architectures, are unfounded. In other words, we have been following the performances of the 10% of companies that garner 90% of the traffic and funding, while missing out on the 90% of the application markets that determine the profitability of the AI industry.
2. The most striking contrast between Chinese and US AI: similar brains, but completely different approaches
Rand assigned nine technical labels to the companies and found a surprising similarity: the distribution of technologies used (Transformers, reinforcement learning, diffusion models) was almost identical on both sides. However, when it comes to practical applications, the differences are clear:
- In the US, reinforcement learning is mainly used to make models speak more naturally and answer questions in a human-like manner (i.e., training AI to act like “gold medal customer service”).
- In China, reinforcement learning is used to train robots to move, autonomous vehicles to navigate, and factory production lines to adjust parameters (i.e., training AI to act like “skilled workers”).
- 61% of US AI companies are pure software companies with no hardware involvement, focusing on healthcare, data analysis, and cybersecurity. More than 60% of Chinese AI companies work with hardware; the proportion of companies developing humanoid robots is six times that of the US, and those working on industrial robots and autonomous vehicles is twice as high. This is not determined by policies but by market demands: in China, where labor costs are lower, AI is used to improve the efficiency of blue-collar workers; in the US, where labor is more expensive, AI is used to support white-collar jobs.
- Another interesting point is the proportion of open-source models. Although the percentages (15% in the US and 17% in China) seem similar, the quality of the open-source models differs significantly. The top US companies (OpenAI, Anthropic, Google) only open-source smaller, edge models, while Chinese companies (Tencent, Alibaba, DeepSeek, Zhipu) open-source their most powerful models for free.
- Currently, Chinese models are as widely used as US models globally, and the pricing power in the mid-range market no longer lies with the US. This is similar to the early home appliance market, where imported brands sold expensive products, while Chinese companies offered cheaper alternatives, eventually taking over the mid-range market.
- The most counterintuitive aspect is that the same proportions of companies use open-source models do not reflect similar development paths. This suggests that the direction of the industry is influenced by market demands rather than policy.
3. A particularly misleading statistic: the same proportions, but opposite trajectories
The report highlights a “statistical trick”: although the percentages of companies using open-source models are similar (15% in the US and 17% in China), the reality is quite different. The US’s open-source efforts come from smaller companies, while the top three companies (OpenAI, Anthropic, Google) keep their flagship models proprietary. In China, the open-source efforts come from leading companies like Tencent, Alibaba, and DeepSeek.
- The mid-range market pricing power for AI models is no longer in the US’s hands. This reflects the fact that AI is adapting to local market conditions: in China, where labor costs are lower, AI is used to improve efficiency in manufacturing; in the US, where labor is more expensive, AI is used in software services.
- The final outcome shows that US AI has thrived in coffee shops near universities, while Chinese AI has developed in factories along the Pearl River and Yangtze River Delta regions.
4. The most surprising finding: the same proportions, yet completely different paths
The report reveals that the actual technology choices of Chinese and US companies are similar, but their applications are vastly different. This suggests that the direction of the industry is driven by market demands, not by policy.
- Rand also notes that only about two-thirds of the core technical data is publicly available, with many companies (68% in China and 60% in the US) not disclosing their model architectures or learning paradigms.
- This is because US AI companies are often independent startups that actively share their technology, while Chinese AI teams are often part of state-owned enterprises or companies that do not need to publicize their technology for funding. Rand’s tools can only process English information, making it difficult to gather comprehensive data.
5. There is no AI “race”; rather, each is evolving independently
Rand’s conclusion challenges the notion of a Sino-US AI competition. The two sides are not competing on the same track but have developed differently due to their unique environments. If future AI must interact with the physical world, China’s strengths in hardware-based applications will be a significant advantage. If AI will primarily operate in the virtual world, the US’s software-based approach will be more viable. Since the future environment is uncertain, the company that best adapts to its environment will survive.
---
In summary, this report provides a nuanced understanding of the Sino-US AI landscape, highlighting the importance of market demands and technological adaptation rather than a competitive narrative.