China faces uphill battle to wean tech firms off Nvidia chips

Chinese technology companies have bought more than 1mn Nvidia gaming and workstation chips this year, turning to products never designed to power advanced AI systems as the country struggles to satisfy soaring demand for computing power.

The buying spree amid a domestic chip supply shortage, revealed to the FT by multiple industry insiders, shows how the US chipmaker continues to play a significant role in China.

Nvidia’s prominence continues four years after Washington began to restrict sales of advanced AI processors to the country and two years after Beijing instructed technology companies to reduce their reliance on Nvidia.

China has made AI development a national strategic priority and more than doubled the production of domestic AI chips this year, led by companies including Huawei, Alibaba, Baidu and Cambricon.

But demand is growing even faster as the country’s biggest tech groups pour billions of dollars into AI models, computing infrastructure and data centres.

FT series: China’s data centre dreams

This is the second in a three-part series examining China’s breakneck build-out of AI infrastructure as it seeks to become the pre-eminent tech superpower

Part I: The world’s newest computing hub — Inner Mongolia
Coming soon: Part III: China’s forgotten cities look to AI data centres for economic boost

Domestic supply is expected to reach as many as 3.5mn AI chips this year, including about 1mn manufactured overseas, but total demand is estimated to far exceed 5mn, according to industry insiders.

The shortage is one of the stresses emerging from China’s attempt to carry out one of the world’s largest and fastest build-outs of AI infrastructure, as Beijing races to create domestic alternatives to compete with the US.

“We believe China’s AI compute expansion has so far been constrained less by demand than by availability of high-end compute,” said Gary Yu, a Morgan Stanley analyst based in Hong Kong.

To help bridge the gap, Chinese technology groups have turned to Nvidia gaming and workstation processors that comply with US export controls but can be repurposed for some AI workloads.

The chips, mostly RTX Pro 6000D, RTX Pro 5000 and RTX 5090, have inferior performance in most areas compared with leading Chinese AI chips such as Huawei’s flagship Ascend 910C.

But they are advanced enough to be used for simpler AI workloads, such as the prefilling phase of inference — how machines read a user’s prompt before decoding and giving an answer.

Nvidia said: “China’s domestic makers of workstation chips and cards have experienced unprecedented growth since 2022. US firms continue to be restricted by a combination of outdated US export controls, which cover gaming products released nearly a half a decade ago, and China’s own limits on US imports.”

Part of the desire for Nvidia’s workstation chips is its “Cuda” software system, which most Chinese tech companies used before being pushed towards domestic alternatives such as Huawei’s Cann, considered not as easy to work with.

DeepSeek is working closely with Huawei to improve its hardware and software, while Huawei has also utilised AI tools to help accelerate the transition of programming language from Cuda into Cann.

“While no flagship large Chinese model has yet been pre-trained on Huawei’s stack, it was inevitable given the progress that had been made,” said the executive at state-backed cloud provider.

The shortage has intensified as Chinese companies rapidly increase the amount of computing power they consume. Alibaba, ByteDance and Tencent have all pushed heavily to introduce advanced models and agentic tools this year, driving an explosion in demand for tokens — the basic units of data processed and generated by AI models.

The total capital expenditure this year from these three Chinese tech groups combined could more than double to about Rmb700bn ($104bn), according to multiple people with direct knowledge of the matter. Meanwhile, Morgan Stanley analysts have projected China’s total expenditure on computing power this year to be $193bn, about 21 per cent of US spending.

“There is a consensus across the industry that the current shortage in AI compute will not be resolved until at least 2030,” Alibaba’s chief executive Eddie Wu said on its August earnings call.

China’s domestic chipmakers are responding. Huawei will probably meet its target of producing 1.6mn AI chips this year, while chip units from Alibaba and Baidu will together make about 1mn. Cambricon is expected to produce 300,000, according to people familiar with the companies’ plans.

Alibaba last month introduced Zhenwu V900, which it claims to be the most advanced domestic AI processor, while Huawei is rolling out “superclusters” and introducing its first AI training chip, Ascend 960DT, in the first quarter of next year.

Chinese chipmakers are also trying to compensate for weaker individual chip performance compared with global leaders by connecting large numbers of processors using advanced networking technology. Huawei has a roadmap to eventually build data centres with up to one million chips connected together.

ByteDance, which has been late to designing its own AI chips, is also expecting to test advanced samples by the end of this year, which it hopes to mass-produce in 2027, according to two people with knowledge of its plans.

Even the rapid expansion of domestic supply, helped by a ramp-up at fabs led by SMIC, is struggling to keep pace with the increase in demand.

China’s frontier AI labs are adding to the pressure as they raise funds and spend more heavily on their own computing infrastructure. DeepSeek and Zhipu were both building their own data centres with 100,000 or more Huawei chips, people familiar with the matter said.

DeepSeek is close to finalising its second round of financing which would add more than Rmb100bn ($15bn) in total this year to its war chest. Hong Kong-listed Zhipu raised a total of $9bn from sale of shares and convertible bonds.

Moonshot and MiniMax, which are yet to finalise plans to build their own infrastructure, are leasing a significant amount of computing power from various providers both domestically and overseas.

The imbalance is expected to become even more pronounced next year. Domestic chip production is forecast to increase again, but demand could more than double to exceed 10mn units, leaving a significant shortfall.

Chinese tech companies are also increasingly looking overseas to secure the computing power they cannot obtain at home.

Tencent signed with Oracle to lease 100,000 GPUs in various data centres across south-east Asia, the FT reported this month. ByteDance and Alibaba are pursuing similar large-scale and long-term lease agreements in order to support their AI demand not yet met, according to people with knowledge of their plans.

Most of these overseas data centres are filled with Nvidia’s cutting-edge chips such as B300, which can be used to make up for the lack of advanced compute onshore for AI training before Chinese chips improve further to be able to perform such tasks.

The widening gap between demand and domestic supply presents a challenge not only for Chinese tech groups trying to catch up with US rivals but also for Beijing as it seeks to build a self-sufficient AI industry.

It is also creating an opportunity for US chipmakers led by Nvidia, which are seeking to preserve their position in China even as export controls restrict the products they can sell there.

Nvidia is introducing a new workstation processor, RTX Pro 5500, which would exceed the performance of its existing China offerings to be a better option for basic AI workloads, according to two people with direct knowledge of the move.

Chinese companies had already indicated orders of more than 2mn in total for this new Nvidia workstation product, which will only start shipping as early as the end of this year, the people said. If no restrictions are imposed, RTX 5500 could further increase Nvidia’s China market share, which was more than 90 per cent before the introduction of US export controls.

“Regulators are talking with tech companies to find out if their demand for Nvidia’s RTX 5500 is necessary,” said one executive at a leading Chinese tech firm. “The honest conclusion will be ‘yes’.”

Huawei, Cambricon, Baidu, ByteDance, Alibaba, DeepSeek and Zhipu did not respond to requests for comment.