DeepSeek founder Liang Wenfeng reportedly told investors that China’s shift away from Nvidia hardware has reached a turning point, with chip supply now the main remaining obstacle. Liang said he expects that constraint to no longer be a major issue within a year.

DeepSeek is currently seeking outside funding for the first time. Liang’s remarks were attributed to minutes from a recent investor meeting, but neither Liang nor DeepSeek has confirmed the document.

China’s remaining gap with the United States

According to Liang, “The biggest gap between us and the US lies in resources,” linking differences in talent, model quality, and applications to the availability of raw computing power.

Chinese companies are restricted from freely purchasing Nvidia’s most advanced chips under US export controls. They also spend less overall on AI and allocate a smaller portion of that spending to personnel, making access to compute one of the main constraints on training and deploying larger models.

Liang described the scale of the gap by saying the largest frontier models in the West have about 800 billion active parameters, compared with tens of billions for China’s leading models, a difference of roughly tenfold.

DeepSeek estimates that it is 12 to 18 months behind leading US companies while producing comparable output with close to one-twentieth of the computing power. The company says its goal is to reduce that gap to three to six months.

DeepSeek’s focus on Huawei hardware

DeepSeek appears to be relying on Huawei as part of its domestic hardware strategy. The company has arranged for about 16,000 Huawei Ascend 950 chips to test how effectively its models can run on Chinese-made systems. Liang said Huawei’s Ascend 950 supernode matches Nvidia’s GB200 and GB300 systems in both performance and price.

DeepSeek is also reportedly preparing to enter the chip design market. Sources close to the matter have said the company is designing its own inference chip, which would place it in competition with Huawei and Nvidia. Inference chips are used to run trained AI models, a different workload from the large-scale training clusters typically associated with frontier model development.

The company and some Chinese researchers are also working on software approaches intended to reduce reliance on Nvidia’s CUDA programming language and toolkit. CUDA has been a key part of Nvidia’s AI advantage because it gives developers a mature software layer for programming GPUs, so replacing it requires both hardware capability and developer tooling.

Liang said he expects the industry’s perception that domestic chips are difficult to use to change within a year.

Nvidia’s position in China

Nvidia’s share of AI accelerators sold in China has declined as local chip designers gained ground. Cryptopolitan reported in April that domestic suppliers held 41% of the Chinese market in 2025, shipping 1.65 million cards, and recent reports indicate that Nvidia’s share has continued to fall.

Alibaba’s chip unit has deployed thousands of its own Zhenwu processors and open-sourced the software stack behind them. The move is part of a broader effort, alongside Huawei and Moore Threads, to provide an alternative to CUDA.

The Chinese government is also said to be actively supporting domestic technology development to reduce reliance on US technology and navigate restrictions imposed by Washington’s export controls on China’s AI sector.

At the World AI Conference in Shanghai, Huawei displayed its Atlas 950 SuperPoD, a system the company says connects 8,192 Ascend chips and delivers 6.7 times the compute of Nvidia’s NVL144, with no US-origin components inside.