(Yicai) July 23 -- China’s main disadvantage in artificial intelligence compared with the United States is not a shortage of talent or technology, but limited access to computing resources, according to the founder of DeepSeek, one of the country's most prominent AI startups.

“The biggest gap between us and the US lies in resources,” Liang Wenfeng, who is also chief executive, said at a recent investor meeting for DeepSeek's latest financing round, according to meeting minutes seen by Yicai. “All the differences we see in talent, model capabilities, and applications can be attributed to differences in computing power resources.”

Neither the Hangzhou-based firm nor Liang has yet confirmed the authenticity of the minutes.

According to the minutes, he also noted that China faces restrictions on buying advanced graphics processing units and invests less capital in AI than the US, while salaries for AI professionals represent a relatively small share of overall expenditure.

The purchase restrictions refer to US export controls that limit Chinese access to Nvidia's most advanced AI chips. 

DeepSeek, which has gained global attention for its low-cost large language models, closed its first ever outside funding round in late May, raising more than CNY50 billion (USD7.4 billion) at a pre-money valuation of about CNY367.5 billion (USD54.3 billion). Investors included Tencent Holdings, Contemporary Amperex Technology, NetEase, JD.Com, IDG Capital, and the National Artificial Intelligence Industry Investment Fund. Liang contributed CNY20 billion, the single largest amount.

The release of the DeepSeek-R1 LLM in January last year was famously described by Silicon Valley venture capitalist Marc Andreessen as an “AI Sputnik moment.”

Based on the meeting minutes, DeepSeek believes it trails leading US AI companies by about 12 to 18 months but has achieved comparable results using only around one-20th of their computing power. The company aims to narrow that gap to three to six months despite ongoing constraints on access to advanced chips.

DeepSeek has computing resources equivalent to about 20,000 Nvidia H-series GPUs, with most of the hardware delivered over the past two months. Liang said that today's largest AI models have roughly 800 billion active parameters, compared with 10s of billions for China's most advanced models -- a roughly 10-fold gap. He added that DeepSeek could not afford to train a model at that scale even if the firm spent all of its new funding on computing power. 

Huawei Partnership and AGI Roadmap

The minutes also said DeepSeek is working closely with Huawei Technologies to optimize its models for Huawei’s Ascend computing ecosystem while using its self-developed high-level programming language called Tile Language to reduce reliance on Nvidia’s Compute Unified Device Architecture ecosystem. This approach is said to significantly improve inference efficiency while limiting performance losses to 1 percent to 2 percent.

The minutes added that Huawei’s AI supernode 950 SuperPod is considered comparable with Nvidia’s GB200 and GB300 systems for both performance and price. The main constraint is production capacity rather than technology or ecosystem maturity, with related ecosystem issues expected to be largely resolved within one year.

Regarding DeepSeek’s technological roadmap, the minutes said the company believes the path toward artificial general intelligence -- AI with human-level general reasoning capabilities -- will progress through chain-of-thought reasoning, AI agents, continuous learning, and a self-evolving "singularity" stage before ultimately leading to embodied AI.

Areas such as video generation and world models, which simulate real-world environments to predict future events, are considered peripheral to DeepSeek's core development path, according to the minutes.

Editor: Emmi Laine