The United States still requires a licence for every advanced artificial-intelligence chip sold to China, and since January 2026 it has been willing to grant one for a limited class of older chips. On 9 November it also postponed, to 10 January 2027, a rule that would extend its blacklist to subsidiaries of listed Chinese firms. Chinese-language reports, meanwhile, describe a home market in which Huawei's shipments have nearly caught up with Nvidia's.1 2 3
What the US rules allow
The rule in force was published by the Bureau of Industry and Security (BIS), the Commerce Department office that administers export controls, and took effect on 15 January 2026. It moved certain chips bound for China and Macau from a presumption that a licence will be refused to review case by case.1
The chips covered are those below two technical limits: a total processing performance score under 21,000 and memory bandwidth under 6,500 gigabytes a second. Nvidia's H200 and AMD's MI325X are the named examples. The exporter must certify that the chip is on sale in the United States, that supply there is sufficient, that shipments to China will not exceed 50% of the volume shipped to American customers, and that the chips have been tested by an independent laboratory in the United States.1 4
Chips above those limits remain under a presumption of denial, which keeps Nvidia's newer Blackwell line out of China. So do sales to companies headquartered in China, or with a Chinese parent, wherever the buyer is located. BIS restated that last point in guidance dated 31 May 2026, after reports that Chinese-owned subsidiaries in South-East Asia had been receiving restricted chips.1 5 6
A separate measure concerns the entity list, the Commerce Department's register of foreign firms that need a licence for almost any American technology. A rule of September 2025 would apply the same restriction to any company at least 50% owned by a listed firm. It was suspended in November 2025 as part of the trade truce with China, and BIS extended the suspension on 9 November 2026 to 10 January 2027. The law firm Dechert notes that the rule was neither repealed nor amended, and returns after that date unless BIS acts again.2
What Beijing says and does
China's Commerce Ministry rejects the controls as a whole. Asked on 4 June about the May guidance, its spokesperson He Yongqian said:7
近年来美方不断以国家安全为由滥用出口管制,严重损害中国企业正当权益,严重破坏国际经贸秩序。
In recent years the US side has repeatedly abused export controls on national-security grounds, seriously harming the legitimate rights of Chinese companies and seriously disrupting the international economic and trade order.7
The opening Washington created in January has been little used, and American sources attribute that to Beijing. The law firm Covington reported at the time that Chinese authorities had told companies they would not approve H200 imports except in particular cases such as academic research. A tracker of the rules records that Nvidia received a first licence in February and reported in an August filing that China's government had restricted purchases of the licensed chip. China's government has not published such an instruction, so these accounts rest on American and press sources.4 5
What Chinese sources say about substitutes
The figures most cited in China come from the research firm IDC, as reported by Chinese outlets rather than published in full. For 2025, the business site 36Kr reported that Chinese vendors shipped about 1.65 million of roughly 4 million AI accelerator cards sold in China, a 41% share, with Huawei's Ascend line at about 812,000.8
For the first half of 2026, Nanfang Metropolis Daily reported IDC figures of 678,000 chips for Nvidia, down 44% from a year earlier, and 572,000 for Huawei. Alibaba's chip unit T-Head shipped 133,000, and Cambricon, Baidu's Kunlunxin, Hygon and Moore Threads between 50,000 and 100,000 each. The paper quoted Huawei's rotating chairman Xu Zhijun as saying in September:3
如果从我们能够统计到的数据看,昇腾应该已经超过了英伟达。
Judging by the data we are able to count, Ascend should already have overtaken Nvidia.3
These counts do not agree with Nvidia's own account. Its chief executive Jensen Huang said in May that the company's direct share of China's AI accelerator market had fallen to zero. The IDC figures as reported do not say how the Nvidia chips it counted reached Chinese data centres.9 3
Chinese commentary is also frank about limits. The same newspaper wrote that a shortage of production capacity had left AI computing power in short supply, and cited a Morgan Stanley forecast that Chinese-made high-bandwidth memory would improve only around 2027. The technology site QbitAI, reporting from the World AI Conference in Shanghai in July, argued that the harder obstacle is Nvidia's CUDA programming software, which developers are used to, and not the hardware.3 10
比英伟达GPU更难跨过去的,是CUDA。
What is harder to get past than Nvidia's GPU is CUDA.10
The next fixed date is 10 January 2027, when the suspension of the affiliates rule and the wider truce both run out. Whether Beijing lets licensed H200 chips through customs in any quantity before then is not known.2