The Circle of Life for LLMs. Was the Reaction to DeepSeek Justified?

Since the release of DeepSeek Large Language Models (LLMs) and free desktop and mobile apps, the industry, the investors, and the media have reacted with alarm, surprised that a Chinese startup—despite operating on a low budget and with limited access to specialized AI hardware—could surpass the latest ChatGPT models with reasoning capabilities. This has led to geopolitical concerns about threats to U.S. technological dominance, and the effectiveness of AI chip sanctions imposed by the U.S. on China. Investor confidence in leading U.S. tech companies involved in AI, AI hardware, and AI/cloud hosting has been shaken, contributing to a significant stock market drop on January 27, 2025. In this paper, we argue that while the success of DeepSeek V3 and R1 is remarkable, it does not signal the decline of any major player. Instead, it is a natural progression of how LLMs and generative AI function. Most LLM providers, of a same LLM generation, rely on similar algorithms, big-data pools, and development techniques, meaning that models tend to converge in performance once their methodologies become public. Whether using proprietary or open source foundations, different starting points often lead to LLMs of comparable capabilities for a same generation. Techniques such as model distillation and reinforcement learning further enable the reduction of model size, data requirements, and hardware constraints. As a result, each time a model is developed, it can be replicated, closely matched, or even surpassed soon after—sometimes with significantly lower effort than the original, or with a significantly smaller set of parameters. This cycle of life will continue as long as LLMs remain a competitive field, by opposition to a commodity, and until new AI approaches beyond generative AI emerge, or the old AI reemerges. We anticipate such a pattern to continue: new models will be matched and overtaken by (nimbler) competitors, while major providers respond with the next iteration of improvements—repeating the cycle. Open source models, in particular, have the advantage of drawing from broader communities and collective innovation, making it increasingly difficult for proprietary models to maintain a lasting edge. As development costs rise, it will be interesting to see whether proprietary models can sustain their dominance or whether they, too, will need to integrate open source strategies. Ultimately, there is, and was, no reason for panic or hasty divestment. AI may be in a bubble, but if it bursts, it will not be because DeepSeek outperforms OpenAI’s latest model. Instead, the real challenges facing LLMs and GenAI lie elsewhere. The path to AGI is likely beyond current LLM-based approaches. While AI agents may extend the viability of generative models for some time, factors such as the finite availability of high-quality digitized training data and the risks of model collapse due to synthetic data contamination pose more significant long-term threats. That said, if LLMs are not the future of AI, there is little reason to be concerned about new players mastering them.