DeepSeek Founder Says Computing Power Is China's Biggest AI Gap with the US

DeepSeek founder and CEO Liang Wenfeng speaks at an event. He assessed that the limitation of computing power is the biggest gap between China's AI and the United States. Photo: Yicai Global.

JAKARTA - DeepSeek founder Liang Wenfeng is said to assess the limitations of computing power as the biggest gap between China's AI and the United States, not a lack of talent or technology.

Yicai Global quoted Friday, July 23, saying the valuation was listed in a meeting report for DeepSeek's latest funding round. However, neither Liang nor the Hangzhou-based company has confirmed the authenticity of the document.

"The biggest gap between us and the US lies in resources," Liang said, according to the memo.

According to the document, Liang attributed the difference in talent, model capabilities, and AI implementation to the disparity in computing resources.

China is also facing restrictions on the purchase of advanced graphics processing units or GPUs. The restrictions stem from US export controls that limit Chinese companies' access to Nvidia's most advanced AI chips.

The paper also said that China's AI investment was smaller than that of the US, while the salaries of AI professionals only took a relatively small portion of total spending.

DeepSeek has attracted the world's attention for being able to develop large language models at a lower cost. The company closed its first external funding round at the end of May by raising more than 50 billion yuan or about 7.4 billion US dollars.

The funding was carried out with a pre-investment company valuation of around 367.5 billion yuan or 54.3 billion US dollars.

Investors involved include Tencent Holdings, Contemporary Amperex Technology, NetEase, JD.Com, IDG Capital, and the National Artificial Intelligence Industry Investment Fund.

Liang put in 20 billion yuan, the largest amount from a single investor in the round.

DeepSeek-R1, which was launched in January last year, was referred to by Silicon Valley investor Marc Andreessen as "the Sputnik AI moment". The term is usually used to describe a technological surprise that triggers major competition.

Based on a presentation seen by Yicai, DeepSeek assessed that its capabilities were still 12 to 18 months behind leading US AI companies.

However, the company claims to get comparable results using only about one-twentieth of the computing power of its competitors.

DeepSeek targets that the gap can be reduced to three to six months even though access to advanced chips is still limited.

The company is said to have computing resources equivalent to about 20,000 Nvidia H-series GPUs. Most of the hardware was received in the last two months.

Liang said the largest current AI models have about 800 billion active parameters. While China's most advanced model only has tens of billions of parameters, or about ten times apart.

Active parameters are the part of the model parameters that are used when the AI processes a command.

According to Liang, DeepSeek still can't train such a large model even though all of its new funds are spent on buying computing power.

The paper also mentions that DeepSeek is working with Huawei Technologies to optimize its model for Huawei's Ascend computing ecosystem.

DeepSeek uses its own programming language, Tile Language, to reduce its dependence on CUDA.

CUDA is Nvidia's programming platform that is widely used to run AI computing on GPUs.

The approach is claimed to improve inference efficiency with a performance drop of only 1 percent to 2 percent.

Inference is a process when an AI model uses the learned ability to produce answers or predictions.

The paper says the Huawei 950 SuperPod is comparable to Nvidia's GB200 and GB300 in terms of performance and price.

The main obstacle is said to be in production capacity, not technology or the maturity of the ecosystem. According to the paper, the ecosystem problem is estimated to be largely resolved within a year.

In its technology roadmap, DeepSeek believes the journey to artificial general intelligence or AGI will go through chain reasoning, AI agents, continuous learning, and the "singularity" stage, when the system starts to evolve on its own.

AGI is AI that is designed to have human-level general reasoning abilities.

The next stage is embodied AI, artificial intelligence that interacts with the physical world through robots or devices.

DeepSeek is putting video creation and world modeling outside its main development path. The world model is an AI system that mimics the real environment to predict future events.