DeepSeek CEO says China can break its Nvidia dependence within a year

DeepSeek founder Liang Wenfeng reportedly told investors that China’s move off Nvidia hardware has hit a turning point and that the last thing standing in the way is chip supply, which he believes should no longer be an issue within a year.

DeepSeek is currently raising outside funding for the first time, and its founder’s statement is attributed to minutes from a recent investor meeting. However, neither Liang nor the company has confirmed the document.

Where is China still trailing the West in the chip race?

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

Chinese firms cannot buy Nvidia’s top chips freely; they spend less on AI overall and put a smaller slice of that money toward people.

Speaking on the scale of the gap, Liang put the largest frontier models from the West at about 800 billion active parameters against tens of billions for China’s best, which is roughly a tenfold spread.

DeepSeek reckons it sits 12 to 18 months behind the US leaders while matching their output on something near one-twentieth of the compute. It says its goal is to pull that lag down to three to six months.

Why is DeepSeek betting on Huawei?

For now, it seems DeepSeek is leaning on Huawei. The company has lined up about 16,000 of Huawei’s Ascend 950 chips to test how well its models run on domestic hardware. Liang stated that Huawei’s Ascend 950 supernode matches Nvidia’s GB200 and GB300 systems on both performance and price.

DeepSeek is also reportedly trying to enter the chip design market, as sources close to the matter say that it is designing its own inference chip, which will put it in competition with Huawei and Nvidia.

There is also the software part that DeepSeek and some Chinese researchers are finding a way around to reduce dependence on Nvidia’s programming language CUDA and its toolkit. DeepSeek is currently chipping at it with a homegrown programming language it calls Tile Language.

Liang expects the industry’s view that domestic chips are hard to use to change within a year.

Does this mean a retreat from Nvidia?

Nvidia’s share of AI accelerators sold in China fell as local designers took 41% of the market in 2025, shipping 1.65 million cards, Cryptopolitan reported in April, and recent reports show that it continues to fall.

Alibaba’s chip unit has deployed thousands of its own Zhenwu processors and open-sourced the software stack behind them, part of a coordinated push, alongside Huawei and Moore Threads, to offer an alternative to CUDA.

The Chinese government is also said to be actively in support of developing local solutions to remove the dependence on US tech and also navigate the limitations of Washington’s export control on its AI sector.

At the World AI Conference that was held in Shanghai, Huawei showed its Atlas 950 SuperPoD, a system it says links 8,192 Ascend chips and delivers 6.7 times the compute of Nvidia’s NVL144, with no US-origin parts inside.

If you're reading this, you’re already ahead. Stay there with our newsletter.