Beijing may let its technology giants buy more Nvidia high-end chips, giving the US chip designer a fresh foothold in China and keeping Chinese artificial intelligence (AI) developers tied to its software after months of stalled trade.
China’s Ministry of Industry and Information Technology (MIIT) has asked major domestic companies, including ByteDance, Alibaba and Tencent, to disclose their plans to purchase Nvidia’s RTX PRO 5500 processors, The Information reported on Sunday, citing people familiar with the matter.
The news follows a meeting between Chinese President Xi Jinping and US President Donald Trump in Washington on September 23-25. Nvidia chief executive Jensen Huang was among the 134 guests at a White House state dinner that Trump hosted for Xi.
According to the report, ByteDance alone is considering buying about one million RTX PRO 5500 chips. Nvidia plans to start shipping the chips to China in late December at roughly 500,000 units per quarter.
The RTX PRO line has been part of Nvidia’s China strategy for more than a year. In July 2025, after meeting Trump in Washington and officials in Beijing, Huang announced a new, fully compliant RTX PRO graphics processing unit (GPU) for China, which Nvidia said was ideal for digital twin AI in smart factories and logistics.
“General-purpose, open-source research and foundation models are the backbone of AI innovation. We believe that every civil model should run best on the US technology stack, encouraging nations worldwide to choose America,” Huang told reporters in Washington at the time.
Nvidia officially launched the RTX PRO 5500 on September 14 this year. The workstation card is powered by a GB202 graphics processor with 21,760 cores, matching the consumer GeForce RTX 5090, and carries 84 gigabytes of memory.
The RTX PRO 5500 is essentially a top-tier gaming card with its memory pushed to the extreme, allowing a company to run open-source AI models on its own servers for inference, but not for training new models. Training builds an AI model by feeding it vast amounts of data, while inference uses a finished model to answer queries.
An Nvidia spokesperson told Reuters on Sunday that US firms remain restricted by a combination of outdated US export controls, which cover gaming products released nearly half a decade ago, and China’s own limits on US imports.
If the purchases go ahead, Nvidia would have the opportunity to maintain its Compute Unified Device Architecture (CUDA) software ecosystem in China, where local chipmakers have been trying to lure AI developers onto their own platforms, some Chinese commentators said.
“The RTX PRO 5500 retains the full CUDA software stack. For Chinese developers, this means existing training frameworks, inference engines and toolchains can be moved across almost seamlessly without adapting the underlying software,” says a columnist with Communications World Web (CWW.net.cn), a telecommunications media platform managed by the MIIT.
“For Nvidia, every chip that enters the Chinese market helps extend the life of the CUDA ecosystem. The RTX PRO 5500 falls just below the export control red line while remaining powerful enough for real applications,” he says. “Whether it is approved offers a perfect case study of how the chip rivalry between China and the US is layered.”
“Officials asked potential buyers directly how many chips they wanted and what they would use them for. This is not a regulator scrutinizing illegal purchases but groundwork for a conditional approval. Such fine-tuned control avoids a hard landing in computing power supply while keeping a firm grip on key technologies and applications,” he writes.
He says that if ByteDance’s order of about one million chips arrived in bulk, it would quickly fill China’s computing power gap but could also delay some companies’ plans to switch to domestic chips. He adds that regulators must strike a balance so that foreign computing power serves as a transitional supplement rather than creating path dependence.
Insufficient local supply
The H200, an older-generation Nvidia AI chip that trails its latest Blackwell processors, has been at the center of a year-long tug-of-war since 2025. Washington barred its export to China until last December, when Trump allowed sales to approved Chinese customers in return for a 25% cut of the proceeds.
Beijing blocked the chip at customs in January and later allowed only limited imports. By August, ByteDance and Tencent had each received about 10,000 units, far short of the 75,000 each was licensed to buy. Beijing has instead steered firms toward Huawei Technologies’ Ascend chips, made by Semiconductor Manufacturing International Corp (SMIC) and run on its Compute Architecture for Neural Networks (CANN) platform.
Huawei’s rotating chairman Eric Xu told the media during the Huawei Connect 2026 conference in Shanghai on September 17-19 that the company’s production of AI computing equipment could not yet meet domestic demand. He said Huawei would not expand into overseas markets on a large scale in the short term.
He said Huawei is focusing on supernodes and large-scale clusters, using chip interconnects to lift the computing power of whole systems rather than relying on advanced process technology and single-chip performance as Nvidia does. Huawei did not disclose its current production capacity or give a timeline for a full overseas push.
On September 14, a few days before Xu made his remarks, Nvidia quietly launched the RTX PRO 5500, which pundits have described as an irresistible product for Chinese AI companies.
Some observers said the RTX PRO 5500 shows where Washington now draws its line: Chinese firms can buy chips powerful enough to run large AI models, but Nvidia’s most advanced Blackwell data center processors, used to train cutting-edge models, remain off-limits.
“What makes the RTX PRO 5500 so special is its 84 gigabytes (GB) of GDDR7 graphics memory with error-correcting code (ECC),” says an IT columnist using the pen name Sophisticated Digital. “The 84GB figure is like a thunderclap. The consumer flagship RTX 5090 has only 32GB, so this is 2.6 times as much.”
“In the past,” the columnist adds, “a large language model (LLM) with 70 billion parameters simply could not fit into an RTX 5090. It had to be split into pieces, with data shuttled back and forth between graphics memory and system memory like luggage. Each transfer took seconds or even tens of seconds, sharply slowing down inference.”
Typical 70-billion-parameter LLMs include Meta’s Llama 3.3 and Alibaba’s Qwen2.5-72B, launched in late 2024. Both are open-source, meaning companies can download and run them on their own computers to answer queries.
The columnist says ECC matters even more. Gaming cards usually drop it to save costs, but a single flipped bit in memory can wipe out days of AI training or scientific computing. By packing ECC and 84GB into one card, Nvidia signals this is a work tool rather than a gaming toy, he says.
He adds that small and mid-sized AI firms could buy a few RTX PRO 5500 cards to build an in-house cluster to deploy their own industry-specific models. He says that, for firms in financial analysis, medical diagnosis and legal document processing, such confidentiality may matter more than computing speed.
Read: Xi backs AI safety with Trump, draws Taiwan line, urges Iran talks
Follow Jeff Pao on X at @jeffpao3
