Nvidia has reportedly halted production of its H200 artificial-intelligence (AI) chips intended for China amid rising political tensions between the United States and China and Beijing’s policy of supporting domestic chips.
The company has reallocated manufacturing capacity at Taiwan Semiconductor Manufacturing Company (TSMC) away from producing H200 chips to its next‑generation Vera Rubin hardware, the Financial Times reported, citing two people familiar with the situation.
The report said Nvidia decided it could not remain in regulatory limbo between the US and China and had to focus on products with clearer market prospects, particularly as demand for its most advanced chips remains strong.
US President Donald Trump said on December 8 last year that Washington would allow exports of the H200 chips to China for civil use. Media reports said Chinese internet giants Alibaba, Tencent and ByteDance were seeking to purchase 400,000 units.
However, after Trump formally approved limited H200 exports to China on January 13, Chinese customs authorities informed Nvidia that the chips would not be allowed to enter the country. In late January, Beijing fine-tuned its line by saying that Chinese companies can purchase the H200 but should consider local chips first. As of now, no H200 chips have been sold yet to Chinese customers, according to US officials.
Some Chinese commentators said Beijing’s decision to curb H200 imports was due to the United States’ interception of 1.8 million barrels of Venezuelan oil destined for China on December 20 last year.
“A tanker that had just left Venezuela was intercepted on the high seas, with armed personnel boarding and redirecting it,” Xia Yuanqi, a Shanghai‑based columnist, writes in an article. “The ship carried a Panamanian flag, the operator was based in Hong Kong and the cargo owner pointed to a Chinese petrochemical company. The interception was less about the oil and more about demonstrating US law enforcement’s power.”
“The seized heavy crude may be worth more than a billion dollars at market prices, but Nvidia’s H200 chips cost 10 times more,” he says. “Nvidia wanted Chinese firms to buy its H200 chips. It definitely felt the pressure.”
“By seizing Venezuela’s oil, the US wants to show its muscle to the whole world,” a Guizhou-based writer says. “But China no longer accepts this. Over the past two years, China’s chip technology has advanced rapidly. Besides, we also hold the rare earth card.”
“The US wants to use the H200 chips to squeeze the final profits from China, but the Chinese market has changed. The global supply of AI computing power has also been diversified,” he says.
Ascend 950PT
Chinese media and commentators had previously welcomed the prospect of Nvidia’s H200 chips entering the Chinese market. They said the H200 would be used for training AI models such as DeepSeek, while domestic chips would handle AI inference, a stage when trained models respond to user queries or perform real‑time tasks.
Jin Canrong, a professor and vice dean of the School of International Relations at Renmin University of China, said on December 19 that Trump had given China a good opportunity to upgrade its AI sector.
After the seizure the following day of the Venezuelan tanker, Jin changed his views. He said on December 25 that China might ultimately not need the H200 chips, as domestic chipmakers were preparing substitutes that could soon enter the market.
On the same day, the Fudan Development Institute published a research report stating that Trump’s approval of H200 exports was primarily a strategy to help Nvidia preserve its dominance in the AI chip market, rather than a sign that Washington was easing its technological blockade against China.
Some Chinese commentators said banning Nvidia’s hardware could provide more market space for Huawei Technologies’ Ascend AI chips.
“China’s ban on Nvidia’s H200 chips seems to be a single issue, but it reflects the broader technology competition,” says a Guangdong-based columnist writing under the pen name “HY Skywalk.” “Washington’s approach is to offer China advanced technology and maintain its chipmakers’ market share. Beijing’s response is to endure short‑term pain while accelerating development of core technologies so the country controls its own technological future.”
He points out that Huawei’s Ascend 910B, Cambricon’s Siyuan 590 and Biren Technology’s BR100 are now being used for AI training and inference. He says Huawei’s Ascend roadmap includes the Ascend 950PR planned for 2026, followed by the Ascend 960 in 2027 and the Ascend 970 in 2028.
High‑bandwidth memory
Nvidia’s Hopper AI chips, including the H100 and H200, are more powerful than most Chinese alternatives largely because they rely on high‑bandwidth memory (HBM), a type of dynamic random access memory (DRAM) produced by South Korea’s SK Hynix. In a personal computer, insufficient DRAM can cause slowdowns or system freezes.
Hopper chips use SK Hynix’s HBM3 and the enhanced HBM3e, while newer Blackwell processors use HBM3e. The next‑generation Vera Rubin architecture is expected to adopt HBM4. The higher the bandwidth of these memory chips, the faster the data can move between AI processors during training.
SK Hynix began mass production of HBM3 in June 2022 for Nvidia’s H100 GPU. It was only in December 2024 that the Biden administration blocked South Korea from exporting HBM chips and related manufacturing equipment to China.
Huawei said last September that it developed its own HBM chips, called HiBL 1.0 and HiZQ 2.0. It said HiBL 1.0 is expected to be used in the Ascend 950PR chip, scheduled for release in the first quarter of 2026.
It remains unclear how Huawei and its supplier, ChangXin Memory Technologies (CXMT), could advance HBM technology so quickly.
South Korean police have repeatedly arrested former employees of chipmakers for stealing semiconductor secrets and passing them to China.
In April 2024, a Chinese national who had worked at SK Hynix for about a decade was arrested at a South Korean airport after allegedly printing 3,000 pages of the company’s confidential documents for Huawei.
Last May, a former SK Hynix employee, surnamed Kim, was accused of copying more than 11,000 documents at the company’s China office and later attempting to apply for a job at Huawei. Last December, South Korean prosecutors indicted 10 people suspected of leaking Samsung’s HBM technology to CXMT.
On January 19 this year, South Korean police said they arrested 378 suspects linked to the leaking of HBM‑related technologies and detained six of them in 2025. They said more than half of last year’s overseas technology leak cases were linked to China.
Read: Beijing to approve Nvidia H200 imports, flagging overreliance
Follow Jeff Pao on Twitter at @jeffpao3