JP Morgan Report on DeepSeek
Yesterday, we saw the market experience significant volatility as it reflected the following factors: 1) TXN (-7.5%) earnings and renewed concerns around analog, 2) META (+1.7%) surprising CAPEX increase, and 3) ongoing worries related to CPOs (ALAB -6.6%; COHR -6.9%). However, 95% of the questions we received were about the Chinese AI company DeepSeek, particularly regarding how they managed to develop state-of-the-art models on a modest budget, raising questions about the billions being poured into cutting-edge AI CAPEX (as reflected in NVDA -3% and VRT -5.2% today).
We believe these concerns are overblown and would like to share why. However, since this discussion is unlikely to be resolved quickly, we are interested in hearing your thoughts and counterpoints.
DeepSeek is a two-year-old AI R&D venture startup based in Hangzhou, China, fully funded and operated by High-Flyer, a leading Chinese quant fund. High-Flyer was founded in 2015 by friends from the prestigious Zhejiang University, aiming to use machine learning for stock trading. In 2023, they spun off this independent research division to develop highly efficient, state-of-the-art AI models using open-source models (you can find their company profile and interviews with their founders here). DeepSeek’s stated goal is to pursue AGI rather than application development or monetization, and its founder, Liang Wenfeng, is known as a “maverick.”
DeepSeek quickly gained attention with its DeepSeek-V2, R1 Lite, and V3 models. The release of R1 Lite in late November was particularly notable for outperforming OpenAI’s o1 model in several benchmarks as a test-time scaling model. This Monday, DeepSeek announced its test-time models R1 and R1-Zero, which have shown competitiveness against OpenAI’s test-time models. The R1 model, which can be experienced here, demonstrates impressive capabilities in explaining the process of understanding user intent and generating responses. Meanwhile, R1-Zero is a unique model trained entirely without supervised fine-tuning.
What has drawn the most attention, however, are DeepSeek’s research papers and the efficiency of its models. It is unclear to what extent DeepSeek is utilizing High-Flyer’s approximately 50,000 Hopper GPUs (a cluster comparable in size to the one OpenAI is rumored to be using for GPT-5 training). What is evident, though, is that they are dramatically reducing costs (for example, they claim that the inference cost of the V2 model is 1/7th that of GPT-4 Turbo). This week, their disruptive (but not entirely new) assertion, which has begun to impact U.S. AI companies, is that “more investment does not necessarily equate to more innovation.” Liang Wenfeng stated, “At present, we see no new approaches, but big companies do not hold a clear advantage. While they have existing customers, their cash flow businesses can become burdensome and make them vulnerable to innovation at any time.” Regarding the delay in GPT-5’s release, he added, “OpenAI is not a god and does not need to always remain at the forefront.”
SemiAnalysis’ Dylan Patel has been pointing this out since June. This does not signify the end of scaling, the end of needing more computing power, or that the most capital-intensive player won’t win (remember Mark Zuckerberg announced a sharp increase in AI CAPEX yesterday). Instead, it suggests that competitors across the Pacific have been forced to pursue efficiency due to export restrictions: “DeepSeek V2 demonstrated astonishing training efficiency, achieving better model performance than other open models with 1/5th the computing power of Meta’s Llama 3 70B. For context, training DeepSeek V2 required only 1/20th the FLOPs of GPT-4, while delivering comparable performance.” If DeepSeek can reduce inference costs, others will have to follow suit, and over time, demand may offset this trend.
There are several recent Substack articles about DeepSeek. In the coming weeks, we will learn more about how this company’s success in cost reduction might reshape the outlook for the AI industry (and the trajectory of chip demand for inference).