The Takeaway
- Chinese regulators blocked ByteDance from using Nvidia chips in new data centers.
- The move is part of Beijing’s strategy to force adoption of domestic chips.
- Beijing is building an alternative AI ecosystem using homegrown chips and open-weight models.
ByteDance bought more Nvidia chips than any other Chinese company in 2025, according to three people with direct knowledge of its purchases. Racing to secure computing power for its billion-plus users and fearing Washington might suddenly cut off supply, the TikTok parent hoarded its stock of the chips.
ByteDance’s fears about a supply crunch were well placed—but when the clampdown came, it wasn’t from Washington. In recent months, Chinese regulators blocked ByteDance from using Nvidia chips in new data centers, according to two ByteDance employees, which means it is sitting on chips it can’t use.
The ban represents a major escalation from August, when Chinese regulators asked local companies to halt new orders of Nvidia AI chips. It is part of Beijing’s strategy to foster its own AI ecosystem as an alternative to that of the U.S., including by forcing local companies to rely on homegrown chips, according to multiple people close to Chinese tech regulators.
The strategy is designed to fit in with Chinese chipmakers’ current capabilities, which are limited to making chips for use in the AI process known as inference, when AI models generate answers or perform tasks. Chinese-made chips are still not good enough to use for AI model training, a process in which models ingest vast amounts of information from the web and elsewhere to understand connections between the data.
As a result, China doesn’t entirely prohibit local companies from buying Nvidia’s more powerful chips for training, according to two of the people close to the regulators. But the U.S. government doesn’t allow Chinese firms to buy those more-advanced chips. Until recently, it only allowed Nvidia to sell those companies a scaled-back version of its more powerful chips, the H20—useful just for inference. But even sales of those chips have stalled for the past six months.
That may be changing a little. The Trump administration is considering approving sales of Nvidia’s H200 chips to China, according to two people familiar close to the U.S. government. Like the H20, the H200 belongs to Nvidia’s last generation of processors, Hopper. But the H200 is roughly twice as powerful as the H20, and the U.S. has never allowed its export to China due to its strong capabilities in training AI models.
If Washington and Beijing can reach an agreement on the H200, that would be a boon for Nvidia. The company saw its revenue from China plunge 63% to about $3 billion in the three months ending October compared with the year-earlier period. The decline was largely due to Nvidia’s inability to sell the H20. Nvidia took in just $50 million in revenue from the H20 in the October quarter, compared to $4.6 billion between February and April. (The vast majority of Nvidia’s revenue from China now comes from gaming chips.)
An Nvidia spokesperson said the company has no plans to resume production of H20, adding that “the regulatory landscape does not allow us to offer a competitive data center [graphics processing unit] in China, leaving that massive market to our rapidly growing foreign competitors.”
Chinese companies are moving faster than those in other markets to reduce their dependence on Nvidia chips, said Brady Wang, a semiconductor analyst at Counterpoint Research. “This is not something they want to do, but something they have to do,” he said.
Beijing’s Learning Curve
The Chinese government took a few years to figure out the right approach to cultivating a local AI industry. When the U.S. government began its crackdown on China’s semiconductor industry in 2022—restricting exports of the equipment used to manufacture advanced chips and the software used to design them—the global AI arms race was just taking off.
Chinese firms were left unable to buy the cutting-edge GPUs that underpinned the first wave of large-model development. Domestic chipmakers lacked both the expertise and equipment to produce anything close to Nvidia’s top-end parts.
Many Chinese officials struggled to grasp the semiconductor industry’s technical and economic complexity. Early policies were often broad, slow and poorly targeted, according to executives who participated in those early consultations. Three years later, the government appears far better informed.
Since last year, the Ministry of Industry and Information Technology has repeatedly summoned China’s largest tech companies, including ByteDance, Tencent Holdings and Alibaba Group, to detail exactly how many Nvidia chips they had purchased, how they were using those chips and how much room existed for domestic substitution, according to four people involved with those meetings. Officials from the National Development and Reform Commission, China’s powerful economic planner, also joined those meetings, signaling that buying more made-in-China chips was no longer just an industry policy but had become a macroeconomic and strategic priority.
Those meetings alone, however, were not enough to move the market. Nvidia’s chips remained superior to Chinese alternatives in both performance and software ecosystem, making them the default choice for internal AI research and external cloud services. Even as they participated in government working groups, China’s tech giants quietly tried to secure as many Nvidia chips as export rules would allow.
Enforcement Focus
In recent months, after closely assessing domestic chip performance and Nvidia’s grip on the market, the Chinese government shifted from information gathering to regulatory enforcement. The Cyberspace Administration of China ordered firms, including ByteDance and Alibaba, to stop testing and purchasing all of Nvidia’s export-compliant chips, despite the fact that many companies had already ordered tens of thousands of them.
At the same time, the Chinese government instructed companies to redirect budget and engineering resources toward domestic alternatives from Huawei Technologies and Cambricon Technologies, according to multiple people close to the authorities. That forced Chinese tech firms to adopt domestic chips.
Many major Chinese AI models are now running soon after their release not only on Nvidia chips but also on Huawei or Cambricon chips. Tencent and Alibaba have publicly said they are rebuilding more of their cloud infrastructure around Chinese chips.
But in private, engineers from these tech giants said they would still prefer using Nvidia chips.
Even with forced adoption, a brutal reality remains: There simply aren’t enough Chinese chips to go around.
For now, many of the chips Chinese firms use as alternatives to Nvidia are still made overseas. Taiwan Semiconductor Manufacturing Co. shipped nearly 3 million Huawei chips into China through intermediaries before Washington opened an investigation, according to Washington-based think tank Center for Strategic and International Studies. The in‑house AI accelerators now in use at Alibaba were also built at TSMC rather than at Chinese plants. However, Alibaba stopped placing new orders after U.S. officials tightened controls and scrutiny of Chinese AI chip designs produced by foreign fabricators.
“China’s AI chip products are all designed around specific AI operational tasks,” said Wang of Counterpoint Research. “For these chips, increasing coordination between software and hardware to improve performance in AI workloads is exactly what Google is doing, but Google can use TSMC’s most advanced manufacturing technology to produce its own chips, while Chinese companies cannot.”
Chinese AI Stack
Beijing’s ambitions go beyond merely replacing Nvidia. It wants to build a fundamentally different AI ecosystem that not only can withstand future U.S. technology embargoes but that other countries will one day widely adopt, according to one of the people close to the Chinese regulators.
That distinction is already showing up in the hardware. Instead of nurturing one clear winner to one day rival Nvidia, China has allowed a dozen companies to compete and flourish—from the likes of Huawei and Cambricon to the chip divisions of tech incumbents including Alibaba Group, ByteDance and Baidu.
Since last year, provincial governments have set up funds and pilot projects to subsidize companies in building mixed‑chip clusters, according to people involved in those efforts. The goal is to produce smaller, diverse systems that work well with a hodgepodge of homegrown and Nvidia chips.
At the model layer, China has leaned into open‑weight systems to stretch scarce top‑end computing capacity. Developers can download, fine-tune and deploy the open-source versions of Alibaba’s Qwen series and all of DeepSeek’s models on cheaper domestic chips.
For policymakers, those models also steer buying decisions. The China Academy of Information and Communications Technology, a government‑affiliated agency, has begun issuing adaptation certificates that test whether local chips can run inference on models like DeepSeek R1. That gives state firms and less technical buyers a clear guide to which locally made chips they can use in their AI projects.
This is a fundamentally different approach than in the U.S., where companies train proprietary models from OpenAI and Anthropic on tightly controlled Nvidia superclusters owned by a handful of cloud giants. China’s stack is more fragmented but also more accessible, with open-weight models, multiple chip vendors and government-coordinated testing standards.
Chinese processors still trail Nvidia’s older export-grade parts, but they’re moving beyond domestic use. For Beijing, the goal is to create full AI packages—hardware, software and models together—that companies can sell overseas as a Chinese alternative, especially in emerging markets.
Those export ambitions are already creating friction. In May, Malaysia announced the country would deploy 3,000 Huawei chips for a national AI initiative. Within 48 hours, after the U.S. Commerce Department warned that using the Ascend chips could violate U.S. export controls, Malaysia backtracked.
The episode exposed the central tension: Washington’s fear is not just that Beijing will achieve semiconductor self-sufficiency, but that China will export an entire alternative technology stack that weakens U.S. leverage and offers countries a way to opt out of U.S.-led systems.
As Chinese chips improve and Beijing packages them with open-weight models and government-backed financing, more countries will face Malaysia’s dilemma: having to choose between two competing AI ecosystems and two superpowers.
