DeepSeek Suspends Second Funding Round After Founder Liang Wenfeng's Remarks Leak; Valuation at About RMB 500

Chinese artificial intelligence startup DeepSeek has suspended its second funding round after founder Liang Wenfeng's remarks at a closed-door meeting were leaked and circulated online. At a May 20 investor meeting, Liang said the company could buy some "non-compliant" processors but did not specify chip types, suppliers, procurement channels, or relevant rules. The meeting minutes were uploaded to GitHub on July 22, and DeepSeek has not publicly confirmed their authenticity. Participating investment institutions confirmed that the meeting was genuine and the record accurate. In its first external funding round, DeepSeek raised about US$7.4 billion, giving it a valuation of approximately RMB 500 billion.

During the meeting, Liang described advanced processors as the most valuable use of DeepSeek's funds, saying the company was willing to convert all available capital into Nvidia chips and pay a premium to secure sufficient hardware. He estimated that spending RMB 20 billion on processors in 2025 would be an excellent performance by the procurement team. Nvidia B200 processors were worth buying in any available quantity at a reasonable price. These remarks come as the US continues to investigate how DeepSeek obtained its computing hardware. In April 2025, the US House Select Committee on China said DeepSeek appeared to have used tens of thousands of Nvidia processors subject to export restrictions to China and separately requested Nvidia CEO Jensen Huang to provide records related to DeepSeek's acquisition of export-controlled chips. On May 31, 2025, the US Department of Commerce's Bureau of Industry and Security issued guidance requiring licenses for exports of controlled advanced computing products to companies headquartered in China or with parent companies in China.

Liang revealed that DeepSeek's computing power is equivalent to about 20,000 Nvidia H-series processors, though he did not specify the exact model, and that most of the equipment had only recently arrived or was still awaiting installation. He estimated that even with an investment of RMB 50 billion, DeepSeek could not train a US-scale system-level model; completing such a training run would require about 50,000 Nvidia GB300 processors or 200,000 Huawei Ascend 950 processors, not including the extensive experiments needed before final model training. He believes Chinese developers are one to two years behind US competitors, with computing resources only about one-twentieth of theirs. Through engineering efficiency, Chinese companies can outperform US rivals on specific tasks, but cannot achieve overall advantage given such a large hardware gap. Liang also said that about half of DeepSeek's core researchers are engaged in data labeling, while US developers OpenAI and Anthropic have built an advantage in high-quality post-training data, and additional funds cannot quickly close that gap.

Liang said Huawei has allocated about 16,000 Ascend 950 processors to DeepSeek, while major Chinese internet companies have received hundreds of thousands. By his estimate, this allocation provides computing power equivalent to about 4,000 Nvidia B-series processors, sufficient for current models but not for next-generation ones. He estimated that four Huawei processors are needed to match the computing power of one Nvidia processor, with Nvidia hardware having an economic life of about five years and Huawei processors no more than three years, and that Huawei systems incur higher operating costs over time due to power consumption. Huawei's production capacity also cannot meet the needs of both DeepSeek and major internet companies simultaneously. DeepSeek buys Huawei processors partly to help Huawei improve its chips and software ecosystem, although the allocation is too small to train next-generation models. DeepSeek is also developing software to make its models easier to migrate between processors from Nvidia, Huawei, and others.

Liang also described lowering prices as a way for Chinese developers to compensate for hardware disadvantages, saying DeepSeek could accept lower profit margins than US competitors, and that companies trying to independently run its publicly released models would find it difficult to compete on cost.