A transcript attributed to DeepSeek founder Liang Wenfeng describes continual learning as central to the company’s AI research, while portraying API and enterprise revenue as a possible source of support for that work.
National Business Daily reported that it obtained a record of a May 2026 investor meeting. The newspaper said an institution involved in DeepSeek’s investment verified that the meeting took place and considered the contents credible. It described a session lasting three hours and 44 minutes and a Chinese-language record of roughly 34,000 characters. The newspaper said DeepSeek had not responded to its inquiry by publication.
The English version dates the recording to May 20, 2026, and says it was transcribed automatically and organized with AI. It warns that speakers are not identified and that names or numbers may be inaccurate. The South China Morning Post reported that DeepSeek had not officially confirmed the transcript.
In the transcript, Liang characterizes continual learning as a major unresolved research challenge and says DeepSeek wants its models to help researchers develop later versions. The transcript portrays continual learning as a step toward AGI and attributes to Liang the view that AGI is DeepSeek’s central goal.
The transcript ties a possible annual recurring revenue (ARR) in the several-hundred-million-dollar range to further growth in API and enterprise demand. The conditional scenario is not confirmation of current revenue or an official company forecast.
DeepSeek’s Jan. 20, 2025, R1 announcement said the model’s code and weights were released under the MIT License, while website and API access were available. In a separate service analysis, the company reported that V3 and R1 processed 608 billion input tokens and 168 billion output tokens from noon Feb. 27 to noon Feb. 28, 2025 (UTC+8).
DeepSeek estimated that charging all of that traffic at R1 prices would imply $562,027 in revenue for the period, compared with $87,072 in GPU leasing costs. The company said realized revenue fell well short of the theoretical figure because V3 was cheaper, web and app access remained free, and nighttime discounts applied.
