DeepSeek CEO Liang Wenfeng Outlines AGI Vision and China’s AI Chip Strategy

Liang Wenfeng, CEO of DeepSeek, recently shared his vision for the company and China's AI ambitions during a nearly four-hour discussion with investors. Unlike leaders at OpenAI or Anthropic, Liang emphasized that DeepSeek’s primary goal is achieving artificial general intelligence (AGI), not profit or revenue. He stated that DeepSeek exists to benefit humanity, not to pursue an initial public offering (IPO) or maximize earnings.

Key Highlights

• DeepSeek CEO Liang Wenfeng prioritizes AGI development over revenue or IPO plans.
• US chip restrictions push DeepSeek and China to develop domestic AI hardware.
• DeepSeek collaborates with Huawei to design advanced AI chips for model training.
• DeepSeek’s open-weight AI models offer lower operational costs than closed-source competitors.

DeepSeek’s Mission and Approach

Liang Wenfeng made it clear that DeepSeek is focused on long-term goals rather than immediate financial gains. He described revenue as "sesame seeds"—small and not worth prioritizing over the broader mission. Instead, he compared the potential of AGI to "watermelons," representing a much larger and more significant achievement.

He explained, "Our original intention was not to make a certain amount of money or to go public. We are doing this with goodwill toward the world, aiming to be useful for humanity." Liang stressed that a company’s vision must be reflected in its actions, not just in statements. He said, "Vision is not a slogan; it is how you operate."

While US-based AI labs focus on attracting enterprise clients and generating returns, DeepSeek’s sole objective is to reach AGI. Liang acknowledged that definitions of AGI vary but maintained that this goal guides DeepSeek’s work. He compared AI development to climbing a staircase, where each step builds on the previous one. He believes that as AI agents improve, they will help create even more advanced models, moving closer to AGI.

Liang identified continuous learning as the key technical barrier for AI. He said that once an AI system can learn and iterate on its own, it will reach a singularity, enabling it to develop increasingly advanced versions without human intervention.

China’s AI Chip Development

Training advanced AI models requires powerful GPUs, but US trade restrictions limit DeepSeek’s access to Nvidia chips. Liang argued that these restrictions are pushing China to develop its own AI chip ecosystem. He stated, "When Nvidia cards can’t be bought, everyone is forced to develop domestic chips."

Liang believes that Nvidia’s position is weakening as Chinese companies invest in alternatives. He highlighted Huawei’s 950 supernode, which he claims can match Nvidia’s GB200 and GB300 chips in both performance and price. DeepSeek is already collaborating with Huawei to design advanced chips for AI training. Liang noted, "We mainly cooperate with Huawei and will participate deeply in their ecosystem."

DeepSeek has also reduced its reliance on Nvidia’s software ecosystem by developing its own high-level compiler, TileLang. However, Liang admitted that Huawei’s chips currently lag behind Nvidia’s, with four Huawei cards equaling the performance of one Nvidia card and a two-year technology gap.

Despite these challenges, Liang is optimistic that Chinese AI labs will produce powerful and cost-effective models. He expects Chinese-made AI to become systematically cheaper, similar to trends in other industries. DeepSeek’s open-weight models are already less expensive to use than closed-source models from companies like OpenAI and Anthropic.