Failing to tap into China’s artificial intelligence market would represent a tremendous loss, NVIDIA CEO Jensen Huang made an alarming remark in an interview with CNBC on Tuesday, adding that the East Asian country’s artificial intelligence market is projected to hit around $50 billion in the next two years.
In recent years, considering the US’ stance on China and its subsequent action restricting both import and export, NVIDIA faced significant challenges. The company has also come under scrutiny for alleged chip smuggling to China.
However, Huang is trying to balance with both countries. On the one hand, he says NVIDIA will support whatever is in the best interest of the US. On the other hand, he notes they hope to continue cooperating with China.
The United States has consistently restricted exports of high-end AI hardware to China, fearing China might take over the AI race and create outcomes unfavourable to its interests.
This, in turn, has affected the company’s business operations and new deals. According to The Information, the company has reportedly received orders from major Chinese tech firms, such as ByteDance, Alibaba, and Tencent, which have collectively placed orders totalling $16 billion for their H20 GPUs.
However, NVIDIA is unable to fulfil these orders owing to government restrictions. The company acknowledged in an exchange filing that it could also face a potential charge of $5.5 billion for exporting H20 GPUs without the required permission. Meanwhile, it is reportedly working on a modified GPU that could legally comply with export restrictions.
The Ban Loop
For context, the H20 GPUs were a modified version of NVIDIA’s H800 GPUs, designed to meet the US export rules so they could be sold to China. These restrictions were introduced during former President Joe Biden’s tenure in 2023. The H800, in turn, was a modified version of the H100 GPU, offering the same computational power but with reduced bandwidth for interconnection.
The H100 features a transfer rate of 600 gigabytes per second. In contrast, the H800 offers 300 gigabytes per second, after the regulations restricted the bidirectional transfer rate, allowing only ones under 600 GB/s—thereby leading to the creation of H800 and banning H100.
Nevertheless, the U.S. government removed the transfer rate as a criterion for imposing restrictions, considering it too easy to bypass. Subsequently, the total computational power threshold was decreased. This meant that the H800 could no longer be exported, which led to the creation of the H20 GPUs.
Source: Semi Analysis.
NVIDIA will not want to forgo its Chinese customers. According to a report last December, technology consulting firm Omdia estimated that ByteDance and Tencent each ordered approximately 2,30,000 NVIDIA chips for 2024. Moreover, in the 12 months ending January 26, 2025, NVIDIA generated $17 billion in sales from China, accounting for 13% of its total revenue.
While NVIDIA continues to battle the United States government and figure out ways to export GPUs to China, the country also has its own way of getting its hands on these chips.
U.S. Cracks Down on Chip Smuggling
In addition to purchasing GPUs that NVIDIA is legally allowed to export, companies in China have allegedly gained access to these chips through data centres and subsidiaries in neighbouring countries.
For instance, Malaysia’s semiconductor industry will closely track the movement of chips across borders, following the US push for tighter regulations on suspected stock movement through the Southeast Asian country.
In addition, authorities in Singapore and the U.S. are also investigating allegations that $390 million worth of NVIDIA chip servers sent to Malaysia were fraudulently redirected to China.
Malaysia and Singapore are ‘Tier 2’ countries in the U.S. government’s Framework of Artificial Intelligence Diffusion, which allows limited chip exports without a license.
The first tier includes 17 countries, including Taiwan, which can access unlimited AI chips. About 120 countries are in the second tier with restricted chip access. Exports are banned to the third group, which includes China, Russia, Iran, and North Korea.
Companies can import up to 1,700 GPUs, valued at $40-50 million, without a license to ‘Tier 2’ countries. Larger imports, worth up to $1 billion, require a license review.
For instance, DeepSeek officially revealed that it had trained its V3 model using over 2,000 NVIDIA H800 GPUs, even as contrasting reports emerge to claim that the company (High Flyer)
has access to around 10,000 H800s and H100s GPUs each.
Amidst rising competition, Chinese technology company Huawei has designed the Ascend 920 AI chip, which seeks to match NVIDIA’s H100 GPUs. The company has already shipped more than 8,00,000 units of its Ascend 910B and 910C chips to clients this year. This includes state-owned telecommunications carriers and private AI developers such as ByteDance. The mass production of the 920 chip is set to begin in the second half of this year.
Besides, a Financial Times report indicated that Huawei is building an advanced chip production facility to deliver multiple parts for the AI supply chain.
“Not only is China on the verge of recreating the entire advanced semiconductors ecosystem domestically–which would rank as one of mankind’s most impressive technological achievement ever–but it’s increasingly looking like a single company–Huawei–will on its own have the entire ecosystem in house,” said Arnaud Bertrand, entrepreneur and founder of MeAndQi.com.
