DeepSeek-R1 Launch Triggers New Trends in the AI Industry, Guo Mingzhi

Details Future Development
New AI Industry Trends Triggered by DeepSeek-R1 Launch
On January 31, Tianfeng International Securities analyst Guo Mingzhi released a report on January 29, stating that the launch of DeepSeek-R1 has triggered two significant new AI industry trends. Even without DeepSeek-R1, these trends would eventually emerge, but its release has accelerated their onset.
Trend 1: As Scaling Law Marginal Returns Slow, AI Computational Power Can Still Grow Through Optimized Training Methods
In the past 1-2 years, investors in the AI server supply chain have been primarily focused on the sustainable growth of AI server shipments under the assumption that Scaling law holds. However, as the marginal returns of the Scaling law gradually decline, the market is beginning to focus on how DeepSeek significantly enhances model efficiency through other methods.
Scaling law states that AI model performance is determined by the number of model parameters (N), training data (D), and computational power (C), and the ideal situation is for all three to increase simultaneously. However, several factors are causing the marginal returns of Scaling law to slow down:
1. Human-created text data (D) is nearly exhausted.
2. Without a significant improvement in computational power (C) or an increase in training data (D), merely increasing model parameters (N) does not help model performance.
3. Computational power (C) is difficult to significantly improve in the short term (e.g., Blackwell system is not yet mass-produced, power supply restrictions, etc.).
From an industry research perspective, DeepSeek-R1 has significantly improved model efficiency by optimizing training methods, showing that when the marginal returns of Scaling law decline, optimizing training methods helps to continue improving AI infrastructure and uncover new applications.
Trend 2: Significant Drop in API/Token Prices Facilitates AI Application Diversification
Currently, the primary profit model in the AI industry is selling tools and reducing costs, rather than creating new businesses or enhancing the value of existing ones. DeepSeek-R1 has adopted an aggressive pricing strategy, offering free usage and setting API/Token prices at less than 1/100 of OpenAI-o1’s. This competitive pressure may drive down AI usage costs, accelerating the diversification of AI applications.
The costs of AI software/services and hardware-based AI have decreased due to lower API/Token prices and optimized training methods. This helps increase the demand for AI computational power and reduces investor concerns about whether AI investments will be profitable. While AI usage has increased due to the drop in prices, whether the increase in usage can offset the decrease in prices remains to be observed.
Conclusion
Scaling law is an empirical rule, and rationally lowering expectations and taking a reasoned view is beneficial for long-term investment trends. Chip upgrades, improvements in power supply restrictions, and the addition of more multimodal data to training will help accelerate the marginal returns of Scaling law again. Large-scale deployers are the ones who will face the slowdown of Scaling law's marginal returns, which again validates Nvidia’s leading position.
Open community resources and China’s highly competitive environment are expected to see other Chinese manufacturers launch LLMs with outstanding scores and more aggressive pricing. At that time, if LLM service providers have not yet started to make steady profits, the pressure to achieve profitability will increase.
Thanks to the significant drop in API/Token prices, AI software/services and hardware-based AI will attract more investor attention. Whether this will become a new long-term investment trend depends on whether profitable business models can be created.
Nvidia remains the winner in the future as Scaling law’s marginal returns accelerate again, but attention should be paid to short-term issues such as the mass production of GB200 NVL72 and any changes in the long-term US semiconductor export ban.