Commentary: Answering the Nine Crucial Questions Raised by DeepSeek

DeepSeek, a Hangzhou-based AI startup, has created significant discourse regarding the resilience of Chinese AI firms under resource constraints and the implications of their innovative, cost-effective model training on U.S. export policies. As its AI applications grow, the company's need for additional computational power will likely increase, even though presently, its training cost-efficient methods are working well with limited resources [para. 1][para. 2].

DeepSeek's AI model, DeepSeek-V3, trained using 2,048 Nvidia H800 GPUs, achieved a lower training cost of $5.576 million compared to foreign peers like Meta's Llama 3, which required much greater computational resources. However, as its models become more popular, the demand for AI inference is expected to necessitate additional computational infrastructure [para. 3][para. 4].

Globally, tech giants continue to significantly invest in AI, with Meta Platforms Inc. anticipating $60-65 billion in capital expenditure mostly for AI infrastructure, and OpenAI, in partnership with Oracle and SoftBank, planning a $500 billion investment for AI development in the U.S. This indicates that the current AI computational power growth paradigm, driven largely by the exploration of AGI applications, remains unchanged [para. 4][para. 5][para. 6].

The evolving AI market may shift focus from a "computational power arms race" to improving "algorithm efficiency," driving innovations in developing models. Open-source licensing will facilitate wider innovation, potentially altering the traditional large model company dynamics. AI commercial deployment at an enterprise scale is expected by 2025 [para. 7][para. 8].

DeepSeek's effective model training using lower-end Nvidia chips, like the A100 and H800, could impact Nvidia's high-end GPU demand, reshaping the GPU market landscape. Their approach demonstrates that Chinese companies can competitively utilize lower-end chips to rival Western-developed models in consumer sectors [para. 9].

A January report highlighted that DeepSeek's reported $5.57 million training cost does not account for the total expenses on research, development, and hardware, which could have reached $500 million. Total ownership costs might reach $2.57 billion over four years, primarily due to their efficient training methods and innovative architectures, such as their model's reduced inference costs [para. 10][para. 11].

DeepSeek's innovations in areas like model architecture and inference efficiency employing technologies such as Mixture of Experts and Multi-Head Latent Attention have notably enhanced its performance. Particularly remarkable is its R1-Zero model's use of pure reinforcement learning without supervised fine-tuning, effectively doubling the speed of ChatGPT and providing superior results in computational tasks [para. 12][para. 13].

The company is also poised to accelerate end-device AI development through its cost-efficient models, promoting businesses to integrate AI applications directly into devices, though this evolution will proceed gradually due to ecosystem challenges [para. 14][para. 15].

Following its global impact, there are concerns that the U.S. might tighten technology export restrictions, potentially limiting high-end AI chip exports, restricting open-source large models, and blocking access to significant datasets for Chinese companies [para. 16][para. 17].

Lastly, while traditionally open-source models lagged behind closed-source ones in performance, DeepSeek's success with open-source models could democratize AI technology development, encouraging other firms to reconsider their business strategies [para. 18].