Chinese AI firms push beyond Nvidia as DeepSeek turns to Huawei

[TAIPEI] When Chinese startup DeepSeek released its latest artificial intelligence model last month, it edged Beijing closer to a future that it has spent years trying to build.

In a small but meaningful break from American technology, DeepSeek said for the first time that its new model had been optimised to run on chips made by Chinese tech giant Huawei. This was a milestone in China’s long-running effort to develop advanced technologies at home and reduce its reliance on Western innovation.

While most of the world’s leading AI systems still rely on semiconductors from the US chipmaking giant Nvidia, Chinese AI firms are increasingly turning to homegrown alternatives.

The timing of DeepSeek’s announcement – before this week’s scheduled summit between US President Donald Trump and Xi Jinping, China’s leader – gives Beijing fresh confidence entering trade talks that US export controls on Nvidia chips have not derailed China’s AI development.

Before last year’s meeting between the two leaders, Trump said that he planned to discuss Nvidia’s most powerful AI chips with Xi, fuelling speculation that the United States might ease restrictions on the technology.

But after years of Washington’s preventing Chinese companies from buying certain advanced technology products, firms such as DeepSeek and Moonshot AI are starting to design their AI systems around the constraints rather than waiting for them to disappear. That includes exploring how their models can run on a broader range of processors beyond Nvidia’s.

“US export controls are not freezing China’s AI development,” said Wei Sun, a principal AI analyst at Counterpoint Research in Beijing. “They are forcing China to build an alternative stack.”

DeepSeek has said that its latest model can use Huawei chips for inference, the process that allows an AI system to respond more quickly and accurately to users. Inference generally requires less computing power than training, the demanding process of teaching a model how to function. DeepSeek still relied on Nvidia chips to train its system, according to two sources in the semiconductor industry who were not authorised to comment publicly on the matter.

It was not immediately clear how DeepSeek gained access to those chips, though Chinese companies can still remotely use Nvidia chips housed in data centres outside China. DeepSeek did not respond to a request for comment.

SEE ALSO

Alphabet, Amazon tap overseas debt markets to fund AI infrastructure push

AI upheaval: entry-level ICT job postings in Singapore fall 38% as roles change, says Accenture

Grab seeks AI edge in product development using agentic engineering, vibe coding

AI growth pushes Alphabet closer to overtaking Nvidia as world’s biggest company

More

Huawei has said that it plans to release a chip for training this year. But it also said that it would take another year after that before its products could match the performance of Nvidia’s current offerings.

The growing split between Chinese and American AI infrastructure is a consequence that Nvidia CEO Jensen Huang has long warned would result from rigid export controls.

He has said that the restrictions have only pushed Chinese companies to accelerate efforts to build domestic alternatives, which could lead to a bifurcated market: Chinese AI systems running on Chinese chips while the West sticks with American hardware.

As the world’s dominant maker of AI chips, Nvidia stands to gain from unfettered access to China. But Huang has argued that the strict restrictions will ultimately hurt the United States by diminishing its influence over China’s AI industry.

Two months after his last meeting with Xi, Trump granted Nvidia permission to sell the H200, one of its most powerful chips, to China.

But since then, those chips have been squeezed between lawmakers in Washington, who are seeking closer oversight of their use in China, and Beijing, which has directed Chinese tech companies to buy domestic chips.

Commerce Secretary Howard Lutnick told a Senate Appropriations Committee last month that no H200s had actually gone to China, and Nvidia said in regulatory filings this year that it had yet to generate any revenue from H200 sales there. Before this week’s summit in Beijing, the fate of Nvidia’s chips in China is no clearer than it was at the last meeting between Trump and Xi.

Analysts expect that China’s frustration with US export controls will be part of the discussion when the two leaders meet.

“Chip export controls have consistently been an issue China opposes,” said Jiang Tianjiao, an associate professor at Fudan University in Shanghai. But as China’s chipmaking abilities improve, officials may not want to interfere with efforts to reduce its dependence on American technologies, he said.

While Chinese technology companies have continued to release high-performing AI systems despite export controls, China’s push for technological self-sufficiency in chip manufacturing still faces significant hurdles. Semiconductor Manufacturing International Corporation, or SMIC, the Chinese company making some Huawei chips, has struggled to produce them at scale. The chips it manufactures are more prone to defects and consume more power than those made by foreign rivals.

Huawei’s workaround has been to strap together large numbers of these weaker chips to achieve the computing power of more advanced processors, a strategy that depends on SMIC’s being able to manufacture in large volumes. Yet Chinese chipmakers are still expected to produce only a fraction of the advanced semiconductors made by foreign companies such as Nvidia this year.

Before Washington tightened controls, many of Huawei’s chips were made by Taiwan Semiconductor Manufacturing Company, which produces most of the world’s advanced chips, including Nvidia’s.

Export controls have constrained China’s ability to make the large volumes of advanced chips needed for AI, said Dan Kim, chief strategy officer at TechInsights, a Canadian research firm, and a Commerce Department official during the Biden administration. But he added that those same restrictions had also pushed Chinese tech companies to innovate in new ways.

Chinese companies are trying to redefine what determines success in the race to build cutting-edge AI. For years, the industry’s most advanced systems have come from companies that can afford to spend billions of US dollars assembling vast numbers of powerful chips.

Now, companies such as Huawei are betting that success could someday depend less on amassing the most computing power and more on building an integrated ecosystem of chips, AI models and applications that is good enough for most real-world uses. By working closely with AI model developers, including DeepSeek, Huawei can customise its hardware to better support the software running on it.

When DeepSeek announced its latest model, Huawei said there had been “close collaboration of chip and model technologies from both parties”.

In technical papers describing its models, DeepSeek outlined specific ways chipmakers could modify their products to improve performance with its systems.

“DeepSeek is calling out into the void to Huawei and other companies, ‘Please make these changes so we can get better performance out of your chips,’” said Jacob Feldgoise, an analyst at the Center for Security and Emerging Technology at Georgetown University. NYTIMES