DeepSeek CEO cites compute gap, pledges open‑source AI models

DeepSeek founder and CEO Liang Wenfeng told investors that the Chinese AI lab’s computing resources are far behind those of leading U.S. firms. He said the current infrastructure equals roughly 20,000 Nvidia H100 GPUs, while training a model comparable to the largest U.S. systems would require about 50,000 Nvidia GB300 GPUs or 200,000 Huawei 950 accelerators. He described this resource shortfall as the chief obstacle to competing with U.S. AI. Despite the gap, DeepSeek is expanding its compute capacity rapidly and plans to build its own large AI clusters. The company, which recently closed a fundraising round valuing it at about $52 billion, said it will keep its most advanced models open‑source, arguing that openness and commercial monetisation are not mutually exclusive. Wenfeng emphasized that DeepSeek’s priority is developing artificial general intelligence (AGI) rather than maximising profit, and that open‑source development serves as a strategic moat. The firm also highlighted a recent capital increase and its focus on high‑quality data annotation to support future model development.