China having second thoughts on Trump's Nvidia chip deal

AllAfrica

4 days ago

China has reportedly urged local firms, especially those for government-related purposes, to avoid using Nvidia's H20 chips for weeks due to national security concerns.
Citing people familiar with the matter, Bloomberg reported on Tuesday that Chinese authorities have sent notices to companies over the past few weeks, discouraging them from using the less-advanced semiconductors.
The report stated that Chinese officials are concerned that Nvidia chips may have location-tracking and remote shutdown capabilities. Nvidia has already denied this.
The Financial Times also reported on the same day that Beijing is demanding technology companies explain why they need to order H20 chips, but not Chinese ones.
The Information reported that the Cyberspace Administration of China (CAC) ordered companies, such as ByteDance, Alibaba, and Tencent, to suspend orders of Nvidia's H20 chips.
According to the report, the government is investigating any potential security risks stemming from H20, while encouraging Chinese companies to purchase chips from local manufacturers, such as Huawei Technologies and Cambricon.
Nvidia told the media in a statement that, as both the US and Chinese governments recognize, the H20 is not a military product or for government infrastructure.
'China has an ample supply of domestic chips to meet its needs. It won't and never has relied on American chips for government operations, just like the US government would not rely on chips from China,' said the company.
On July 15, Nvidia's Chief Executive Jensen Huang said the company would soon resume H20 chip sales in China.
On the same day, US Commerce Secretary Howard Lutnick said the H20 is only Nvidia's 'fourth-best' AI chip, which is slower than the fastest chips American companies use.
He said the US wants to sell the Chinese enough AI chips that their developers 'get addicted to the American technology stack.' He said the United States' strategy is to stay one step ahead of what the Chinese can build.
At that time, some pundits in China described the H20 chips as an irresistible 'poison wine' that would slowly kill Chinese AI chip makers by taking away their customers. Some others criticised the US for 'dumping' the H20 to China.
Reuters reported on July 29 that Nvidia had placed orders for 300,000 H20 chipsets (worth about US$3.6 billion) with the Taiwan Semiconductor Manufacturing Company (TSMC) a week earlier.
On July 31, the CAC said it summoned US tech giant Nvidia over security risks related to its H20. It said some US lawmakers have called for advanced chips exported abroad to be equipped with 'tracking and positioning' functions.
'The US did not unconditionally relax its export controls for the H20 chips,' a writer with Mydrivers.com says in an article published on July 23. 'It still does not allow the export of the high-end Blackwell chips. It seems to be saying that, 'I give you a sweet, but this is the last one.''
The writer says that the Trump administration has not yet altered its strategy to counter China, but has instead employed an alternative approach. He says the US side must demonstrate sincerity if it wants to maintain dialogue with China, obtain Chinese rare earths, and push forward with Trump's visit to Beijing in September.
In an article, a Beijing-based financial columnist says that the US does not want to miss the opportunity to capitalize on the massive demand for AI chips in the country.
'The US government's lifting of the ban on the H20 isn't unconditional goodwill,' he says. 'Lutnick has explicitly stated that the US policy goal is to get Chinese developers addicted to American technology.'
He says Lutnick's comments bluntly revealed the United States' intention to profit from China, while controlling its technological development. He says Nvidia will reportedly sell the B20, B30, and B40 chips to China in September, fueling the already-drastic competition between Nvidia and Huawei Technologies in China.
'In Huawei's labs, engineers are now testing the performance of Ascend 910B chip clusters,' he says. 'This chip can reach 320 TFLOPs (trillion floating-point operations per second) using FT16 (16-bit floating-point numbers) in speed to match the H20, while its price is 40% lower.'
He says China will have more bargaining power in the global chip sector once its chips account for half of the Chinese AI computing market.
In a press briefing on Monday, United States President Donald Trump explained why he agreed to restart the export of the H20 chips to China.
He said that he had asked Nvidia for a 20% cut of the company's H20 chip sales to China, but ultimately, he agreed on only 15%. He stressed that the money will go to the US government, not himself. The same charge will also be applied to AMD, which sells MI308 to China.
Trump also said he would discuss with Nvidia's Huang if Nvidia could reduce the performance of its Blackwell system by 30 to 50% for the Chinese market.
'That would be an unenhanced version of the big one [the most advanced Blackwell chips,' he said in a press briefing. 'We still sometimes sell fighter jets to a country, and we'll give them 20% less than we have.'
In a February article, an IT columnist named Uncle Biao writes that Huawei planned to sell 100,000 Ascend 910C chips and 300,000 910B chips in 2025.
'Ascend 910C is a new generation AI chip, with its inference performance reaching 60% of Nvidia's H100,' he says. 'Besides, an inference service using DeepSeek R1/V3 and Ascend 910B has its performance reaching 80% of the A100.'
'Huawei uses its Ascend 910 series to break the Chinese market's dependence on Nvidia chips and become a significant competitor in the domestic market,' he says.
He adds that Huawei is also building an AI ecosystem with its software tools, including MindSpore, the Atlas computing platform, ModelArts AI development platform, algorithm libraries, and Compute Architecture for Neural Networks (CANN).
Huawei's CANN is a heterogeneous computing architecture for AI, aiming to break the monopoly of Nvidia's Compute Unified Device Architecture (CUDA). Key foundries (such as TSMC) and electronic design automation (EDA) software suppliers (such as Cadence, Synopsys, and Siemens) have already adopted the Nvidia CUDA-X and Blackwell platforms.
On August 5, Huawei announced the full open-source release of CANN. It said the software serves as a bridge between high-level AI training frameworks and Ascend chips.
Huawei's Rotating Chairman Xu Zhijun said that after seven years of development, CANN has achieved key breakthroughs in computing optimization, communication efficiency, and memory management. He said CANN can now provide computing power throughout the entire AI-model training and deployment process.
Read: China fears Nvidia chips could track, trace and shut down its AIs