Summary of Key Points
American closed-source large-model companies (such as OpenAI, Anthropic, and Google) once enjoyed high commercial premiums due to their leading model capabilities. However, this period of easy success is coming to an end. The reasons include: a slowdown in the growth of closed-source models, a rapid narrowing of the performance gap between models, the commercialization of cutting-edge intelligence technologies, and the continuous advancement of open-source models (especially Chinese models like Kimi K3), which are approaching the level of closed-source models. Additionally, the AI industry is shifting from a focus on “who has the strongest model” to “who can scale AI solutions effectively,” but this transition is facing obstacles.
1. Growth of Closed-Source Large Models Slows Down: Insufficient Computing Power, Intensified Competition
The growth rate of closed-source models has slowed down. For example, while the number of calls to Google’s Gemini API continues to increase, the growth rate has dropped from 60% to 37.5%. Anthropic’s annual recurring revenue (ARR) is still higher than OpenAI’s, but its growth rate has also slowed.
Why? On the surface, it seems that there is a shortage of computing power—everyone is in need of GPUs, and even Chinese models like Kimi K3 have had to pause accepting new users due to computational constraints. However, the deeper issue lies in the accelerated pace of competition: what used to be a several-month lead can now be overcome within weeks. For instance, Google suddenly discontinued Gemini 3.5 Pro in favor of the Flash model, and there are rumors that GPT6 and Fable 5.1 are about to be released. The duration of a model’s leadership is getting shorter, making it increasingly difficult to maintain growth.
2. Model Leadership Is No Longer a Barrier to Success: Alternating Leads Are the New Norm
In the past, it was believed that the company with the strongest model would dominate the market. However, this is no longer the case. The founder of DeepSeek suggests that Anthropic’s lead over OpenAI will not be sustained in the long term, and the three companies (OpenAI, Anthropic, and Google) will likely take turns leading.
For example, Anthropic gained momentum with Claude Code, but OpenAI quickly released GPT5.6 to overtake it. Anthropic then launched Opus 5 at a lower price than Fable 5, although its performance is only comparable to the earlier GPT5.6 Sol. Model leadership has become a temporary advantage rather than a permanent barrier; once you invest heavily in training a powerful model, competitors can quickly catch up, making it harder to profit from a short-term lead.
3. Large Models Become Standardized Components: The Era of Commercialization Has Begun
Companies no longer focus on which model is the strongest but on which one is most cost-effective to use. This marks the beginning of commercialization, where models have shifted from being a core selling point to being production-ready components.
For instance, companies use tools like Harness to decouple models from their business processes, allowing them to switch between different models easily without significant costs. Although the API model approach initially offered great potential, it is now becoming the most replaceable component in many systems. Companies are also more focused on optimizing resource usage (e.g., reducing token consumption) while achieving the same number of tasks for the same cost. The strength of a single model is no longer decisive; combinations of models and efficient scheduling capabilities are becoming more important.
4. Open-Source Models Are Competing in Low-Price Markets
Chinese open-source models (such as Kimi K3) are challenging the market share of closed-source models. Anthropic’s Haiku series, which targets low prices, has seen two of its models discontinued due to competition from open-source alternatives. Open-source models not only offer similar capabilities but also at lower costs, effectively eroding the market space of closed-source solutions.
More concerning is that open-source models are making inroads into high-value markets. Kimi K3’s parameter size has increased from 1T to 2.78T, and its performance is approaching that of closed-source models, yet the price remains much lower. This compresses the premium space for closed-source models—what used to be a significant advantage is now achievable at a lower cost with open-source alternatives.
Anthropic is trying to extend its lead by restricting chip usage and attacking distillation techniques, but history shows that it is difficult to lock in a competitive advantage with a single technology.
5. The AI Industry Is Stuck in the “Valley of Death”: Scaling Challenges
The AI industry has moved from a focus on model capabilities to scaling solutions, but facing difficulties in practical implementation. The incremental market for AI has not expanded as expected, and the growth rate of AI adoption per employee has slowed down. Companies are more concerned with the cost of using AI to complete tasks rather than the strength of the models themselves. For example, some companies that previously tracked token consumption are now cutting costs. This is like being stuck halfway up a ladder: while model capabilities are sufficient, there is no clear way to motivate companies to continue investing in AI solutions, leading to a “valley of death” in terms of practical adoption.
Conclusion
The “golden age” of closed-source large models is coming to an end, and the days of reaping profits from a single technological lead are over. In the future, success will depend on faster iteration, lower costs, better integration with business operations, and the ability to withstand the impact of open-source models. The AI industry is transitioning from a focus on model performance to practical application of AI solutions.