Analysis: China’s DeepSeek Proves Limited Resources Are No Barrier to Catching Up With U.S. AI Rivals

DeepSeek, a Chinese artificial intelligence (AI) startup, is making waves in the global AI industry by offering powerful and cost-effective AI models, attributed to high-quality engineering and a "fast follower" strategy [para. 1]. The Hangzhou-based company launched its open-source AI model R1 on January 20, demonstrating its capability to rival OpenAI’s o1 model and the upcoming Llama 4 from Meta Platform Inc., using significantly lower resources and costs [para. 2]. The R1 model is an upgrade from DeepSeek-V3, which had 671 billion parameters and required only 2,048 Nvidia H800 chips and $5.58 million for training, far less than what OpenAI and Meta spend for similar-sized models [para. 3].

The Nvidia H800 GPU was designed for the Chinese market following U.S. export bans on advanced semiconductors in October 2022, but further restrictions were imposed by Washington in October 2023, pushing Chinese firms to optimize their use of available chips [para. 4]. Andrej Karpathy, an OpenAI founder, described DeepSeek-V3 on social media as groundbreaking for managing resource constraints, suggesting a model of its scale typically needs at least 16,000 GPUs [para. 5][para. 6].

The ability of DeepSeek, founded by Liang Wenfeng, to achieve competitive performance at a fraction of rivals' costs is seen as a showcase of Chinese tech firms’ resilience against U.S. restrictions [para. 7]. Their success raises questions about the high costs incurred by other AI developers [para. 8]. DeepSeek's approach leverages the fast-follower strategy, which allows them to replicate and improve upon existing AI models at a lower cost [para. 9][para. 10][para. 11][para. 12].

A technical executive from a leading Chinese LLM company and AI startup StepFun's founder, Jiang Daxin, noted that, while U.S.-based companies like OpenAI pursue multiple technical avenues, Chinese companies must follow due to resource constraints, aiming to catch up and innovate eventually [para. 13][para. 14][para. 15]. The "fast-follower" strategy has been effective as the AI industry moves toward the culmination of model capabilities driven by extensive data use [para. 16]. OpenAI co-founder Ilya Sutskever highlighted at the NeurIPS 2024 conference that conventional data-driven advances may be nearing their limits, necessitating new approaches [para. 17].

Efficient resource utilization has become a priority for Chinese AI firms, resulting in significantly reduced training and inference costs compared to international counterparts [para. 18][para. 19]. Inference cost measures the expense of processing data to generate model responses, and DeepSeek's R1 model's inference cost is considerably lower than OpenAI's [para. 20]. Kai-Fu Lee, the former Google China head and founder of 01.AI, commended Chinese companies’ infrastructure abilities, noting substantial cost advantages over international peers and reflecting on his prior prediction of superior Chinese engineering [para. 22][para. 23][para. 24].

Lee's past predictions met skepticism, but the outcomes of DeepSeek’s releases lent credibility to his insights [para. 25]. Chinese AI firms have increasingly focused on leveraging their strengths in cost and resource efficiency, positioning themselves advantageously within the industry [para. 26][para. 27].