Leang Wenfong, founder of Deepseek, a Chinese AI lab, remains an elusive figure despite his company’s significant impact on the AI industry. Deepseek’s R1 model notably caused Nvidia to lose $593 billion in a single day, and their latest V4 model outperforms competitors like Claude and GPT-5 at a fraction of the cost. Wenfong’s background includes topping his college entrance exam, studying engineering, and founding a successful quant fund that leveraged machine learning for trading. His transition from finance to AI is reflected in his business-minded approach to AI development, focusing heavily on cost, pricing, and operational discipline.
Deepseek’s pricing strategy is notably pragmatic: they price their API to recoup server costs within ten months, aiming for a sixfold profit over the server’s lifespan. Wenfong emphasizes volume and usage over maximizing margins, drawing parallels to Amazon Web Services’ strategy of cutting prices to outpace competitors. He also argues that open-sourcing model weights does not threaten their business because the real challenge lies in serving the models efficiently and cost-effectively, something rivals struggle to do at Deepseek’s low prices.
The company deliberately avoids chasing consumer markets or overextending its resources. When their R1 model went viral, Deepseek made minimal efforts to retain or monetize the influx of users, focusing instead on their core mission. Wenfong views AI development as a stepwise process, with language models as the foundation of general intelligence, followed by agents and continual learning. He believes the next breakthrough will be models capable of continuous learning, akin to how new employees quickly adapt to company culture and knowledge without needing everything spelled out repeatedly.
Despite the cutting-edge nature of their work, much of Deepseek’s current effort involves labor-intensive data labeling, reflecting the reality of AI development as “Software 2.0,” where data is the new source code. Wenfong plans to invest heavily in GPUs to accelerate research, viewing hardware acquisition as a more profitable and urgent investment than holding cash. He acknowledges the significant compute gap between China and the US but is optimistic about overcoming software barriers with innovations like Tileang, a language that simplifies GPU kernel programming and reduces reliance on Nvidia’s CUDA.
Ultimately, Wenfong stresses that the key to achieving AGI lies in maintaining a strong, cohesive team rather than focusing solely on compute or models. Deepseek enforces strict policies against poaching staff and spin-offs to protect its talent pool. The company fosters a flexible research environment with minimal KPIs and encourages self-directed work, though it is currently building more formal hierarchies. Wenfong promises that Deepseek will always release its strongest models openly, ensuring transparency and contributing to the broader AI ecosystem.
