Beijing appears to be softening its stance on Nvidia\u2019s H200 graphics processing units (GPUs), with Chinese policymakers and commentators shifting their focus from outright rejection to the practical question of how the chips would be governed in use.<\/p>\n
When Reuters reported<\/a> on November 21 that the US Commerce Department was considering allowing exports of the H200 to China, Chinese commentators warned that the artificial intelligence (AI) chips could be \u201csugar-coated bullets\u201d that would entrench Nvidia\u2019s market dominance and slow the long-term development of China\u2019s domestic chip-making sector.<\/p>\n United States President Donald Trump and Chinese President Xi Jinping held a phone call on November 24. Ten working days later, on December 8, Trump announced in a social media post<\/a> that the US would allow Nvidia to ship its H200\u00a0products to China, under \u201cconditions that allow for continued strong national security.\u201d<\/p>\n Trump highlighted that he had informed Xi of this, and said that \u201cXi responded positively.\u201d He also used the chance to criticize his predecessor\u2019s approach.<\/p>\n \u201cThe Biden administration forced our great companies to spend billions of dollars building \u2018degraded\u2019 products that nobody wanted, a terrible idea that slowed innovation, and hurt the American workers,\u201d he said. \u201cThat era is over! We will protect national security, create American jobs, and keep America\u2019s lead in AI.\u201d<\/p>\n The degraded products he mentioned are believed to be Nvidia\u2019s H20 and AMD\u2019s MI308 chips.\u00a0<\/p>\n When asked whether Beijing would allow Chinese firms to purchase Nvidia\u2019s H200 chips, Guo Jiakun, a spokesperson for the Chinese Foreign Ministry, said at a regular media briefing on December 9 that \u201cChina always advocates that China and the US achieve mutual benefit through cooperation.\u201d<\/p>\n Some Chinese commentators have outlined a so\u2011called \u201ctwin track\u201d or \u201cwalk on two legs<\/a>\u201d approach. Under this framework, Chinese firms can use H200 chips for AI model training, which requires high-performance processors to process vast datasets and adjust model parameters. <\/p>\n Once a model is deployed, it can run inference tasks using Chinese chips. When a person asks DeepSeek to generate responses or predictions, the AI model performs inference, which requires\u00a0less powerful but more energy\u2011efficient AI chips than those used for AI training.\u00a0 \u00a0<\/p>\n The term \u201csugar\u2011coated bullet<\/a>\u201d dates back to March 1949, when Mao Zedong cautioned Communist Party cadres that although military victory in the civil war was near, they could still be defeated by subtler temptations. Mao used the phrase to describe the allure of comfort, flattery and material inducements offered by capitalists, warning that such temptations could quietly erode political resolve.<\/p>\n Nearly a decade later, as the Great Leap Forward<\/a> began in 1958, Mao promoted the concept of \u201cwalking on two legs,\u201d calling for the full mobilization of all positive forces and warning against one\u2011sided or unbalanced development. In practical terms, the slogan urged China to advance heavy industry alongside manufacturing and agriculture, and to pursue production growth while also safeguarding people\u2019s livelihoods.<\/p>\n H200\u2019s strategic value<\/p>\n A Chinese columnist writing under the pen name Tangyipao gives a detailed explanation of the \u201ctwo-track\u201d approach, highlighting the H200\u2019s strategic value for China.<\/p>\n \u201cFirst, we need to understand what the H200 really offers,\u201d he writes. \u201cThe H200 is Nvidia\u2019s second-tier flagship chip, just below its latest Blackwell series, with its core advantage lying in high-bandwidth memory that significantly boosts data throughput.\u00a0The H200 delivers nearly six times the performance of the H20.\u201d<\/p>\n \u201cFor AI research, high-end AI chips are essentially engines of computing power. Access to the H200 will enable Chinese laboratories to build systems that approach top-tier US performance for large-language model (LLM) training and complex data processing,\u201d he says. Famous LLMs include ChatGPT and DeepSeek.\u00a0<\/p>\n \u201cWhile benefiting from the relaxation of the H20 export,\u00a0we must stay rational,\u201d Tangyipao says. \u201cChina can make reasonable use of advanced foreign technology to promote industrial development, but it must also remain alert to the trap of technological lock-in. After all, the US government\u2019s 25% cut in Nvidia\u2019s H200 revenue from China is actually a tax on Chinese buyers. China\u00a0needs to press ahead with domestic chip development.\u201d<\/p>\n \u201cWith the H200, China can accelerate technological iteration and better pursue a \u2018two-track\u2019 approach.\u00a0Even without it, China\u2019s AI industry can continue to move forward on a path of independent development,\u201d he says. \u201cAs domestic processors such as Huawei\u2019s Ascend mature, China\u2019s reliance on foreign high-end chips will gradually decline.\u201d<\/p>\n Mao Yu, a columnist at the Shanghai-based IT Times, has this to say:\u00a0\u201cCompared with the degraded H20, the H200 offers a significant performance jump. But its real appeal is not just hardware. It is Nvidia\u2019s deeply entrenched CUDA ecosystem.\u201d Citing an unnamed industry expert, Mao says several major Chinese technology firms are preparing to order close to 100,000 H200 chips.<\/p>\n \u201cThe reason big Chinese companies are willing to buy is that the H200 is essentially plug-and-play,\u201d he says. \u201cAfter more than a decade of development, the CUDA architecture has formed a complete closed loop, from development tools to application scenarios. It has become the infrastructure for AI training, an advantage domestic chips cannot replicate in the short term.\u201d<\/p>\n He says people should not overworry that the H200 will dominate the Chinese market.<\/p>\n \u201cOnly a handful of large Chinese companies are engaged in LLM training, and many of them have already stockpiled key resources such as HBM memory chips, with the H200 to be serving mainly as a supplement to existing clusters,\u201d he says. \u201cMost small and medium-sized enterprises focus on inference workloads with limited need for cutting-edge computing power, while government-backed cloud providers and users in specific sectors continue to favor domestic AI chips.\u201d<\/p>\n \u201cWhile\u00a0domestic chip makers still need some time to catch up to Nvidia in terms of high-end AI chip designs and production capacity, the H200 will help China ease its computing power bottlenecks and improve research efficiency,\u201d he says.\u00a0<\/p>\n Reuters also reported<\/a> that Chinese technology firms, including ByteDance and Alibaba, are keen to purchase the H200 chips.<\/p>\n \u2018No kill switch\u2019<\/p>\n The Trump administration approve<\/a>d exports of Nvidia\u2019s H20 chips to China in July, but the decision was quickly met with skepticism from Chinese commentators. Many dismissed the chips as \u201cpoisoned wine,\u201d arguing that even with reduced performance, they would further bind domestic AI firms to Nvidia\u2019s technology stack and slow efforts to build an independent ecosystem.<\/p>\n Beijing subsequently raised concerns over alleged \u201cback doors\u201d in the H20, with commentators warning that such features could allow Nvidia or US authorities to locate, monitor or even remotely control GPUs deployed overseas.<\/p>\n Nvidia has repeatedly rejected those allegations. In a blog<\/a> post published on December 10, the company said it was developing a software-based monitoring tool designed to help customers manage large GPU deployments, not to exert control over them.<\/p>\n \u201cNvidia is developing a software solution for visualizing and monitoring fleets of Nvidia GPUs, giving cloud partners and enterprises an insights dashboard that can help them boost GPU uptime across computing infrastructures,\u201d the company said.<\/p>\n Nvidia emphasized the service would be optional and software-driven.<\/p>\n \u201cThis optional service provides real-time monitoring by each GPU system communicating and sharing GPU metrics with the external cloud service,\u201d the company said, adding that \u201cNvidia GPUs do not have hardware tracking technology, kill switches and back doors.\u201d<\/p>\n According to Nvidia, the service will feature a client software agent that the customer can install to stream node-level GPU telemetry data to a portal hosted on Nvidia GPU Cloud (NGC).\u00a0<\/p>\n In other words, Nvidia and its clients will be able to locate all the company\u2019s new chips overseas.<\/p>\n Coincidentally, The Information reported<\/a> on the same day\u00a0that Chinese artificial intelligence startup DeepSeek was using smuggled Nvidia\u2019s Blackwell chips to build \u201cphantom data centers\u201d\u00a0to develop new AI models in China. The US has banned the shipment of Blackwell chips to China.<\/p>\n Responding to the report, an Nvidia spokesperson said the company had seen no evidence of large-scale chip smuggling. The spokesperson said Nvidia had not received credible information suggesting that data centers were built to deceive the company or its partners and then secretly dismantled and moved elsewhere, adding that such scenarios were unlikely.<\/p>\n Some observers said Beijing is less concerned about Nvidia\u2019s ability to track the physical location of its chips than about whether the company\u2019s assurances of having \u201cno kill switch\u201d can be independently verified.<\/p>\n Jin Mei, a Guangdong-based writer, says<\/a> any potential back door functionality must be clearly verified, not merely addressed through Nvidia\u2019s assurances. She says Chinese buyers must seek case-by-case regulatory approval and explain why Nvidia chips are needed instead of domestic alternatives.\u00a0<\/p>\n Read: AMD\u2019s MI308 chips may face H20\u2011like headwinds in China<\/a><\/p>\n Follow Jeff Pao on Twitter at\u00a0@jeffpao3<\/a><\/p>\n \t