By Britney Nguyen
The stakes are high for Nvidia as it looks to restart China chip sales in the face of fast-evolving geopolitical tensions
Nvidia is working out a deal with the Trump administration to sell its H20 chips in China again after those sales were effectively banned in April.
Nvidia Corp. is at the cutting edge of chip design, keeping competitors on their heels with a fast-moving roadmap of innovations that are powering the artificial-intelligence wave. But lately Washington and Wall Street have become preoccupied with a two-year-old product with far less advanced features.
The chip at the center of the fascination is the H20, a downgraded model from Nvidia’s (NVDA) older Hopper lineup that was specially designed to comply with export restrictions on sales of AI hardware to China. While it’s less powerful than Nvidia’s most sophisticated chips, the H20 has some features that have become more crucial in the AI race since U.S. export controls around them were implemented.
A few months ago, President Donald Trump effectively banned Nvidia from selling the chip at all, citing national-security concerns. Then he announced this week that he’d allow the company to sell the H20 again if it gave the U.S. government a 15% cut of related revenue – an arrangement some experts say is unprecedented.
See more: Nvidia and AMD reportedly strike deals with Trump – but analysts see a ‘slippery slope’
Nvidia did $17.1 billion in revenue in China last fiscal year, representing 13.1% of total sales. The financial stakes are high, as the company has been effectively banned from selling in the market since Trump imposed restrictions in April. Chief Executive Jensen Huang has predicted the China AI market could be worth $50 billion in two or three years.
Investors have been eager for Nvidia to get back to selling in China again, and Huang said demand for the H20 remains high, even though it is an older model. The company is in the process of developing a downgraded version of its new Blackwell offering that it hopes to win clearance to sell in China down the line. Nvidia reportedly ordered 300,000 H20 chipsets from Taiwan Semiconductor Manufacturing Co. (TW:2330), according to a Reuters report in July, suggesting the company won’t simply look to draw down inventory of the H20 but will aim to produce more.
According to the Reuters report, Nvidia has a stockpile of 600,000 to 700,000 H20 chips. Jefferies analysts said at the time that they estimated Nvidia’s H20 inventory to be between 600,000 and 900,000.
But now China is reportedly pushing back, with Bloomberg News writing earlier this week that the government is urging local companies to ditch the H20. Chinese tech firms have received notices from the government in recent weeks to not use the chip for governmental or national-security-related purposes, the report said, citing unnamed people familiar with the matter.
An Nvidia spokesperson told MarketWatch that both the U.S. and Chinese governments recognize that the chip “is not a military product or for government infrastructure.”
China itself “has ample supply of domestic chips to meet its needs” and “won’t and never has relied on American chips for government operations, just like the U.S. government would not rely on chips from China,” the spokesperson said.
Still, the risk is that the Chinese government could block the chip from coming back in an effort to support homegrown semiconductor competitors like Huawei Technologies Co. as they look to make inroads, even if it slows down the country’s AI industry overall.
The Chinese government has questioned the security of the Nvidia chips and said they are not good for the environment. Nvidia’s spokesperson said its chips don’t have “back doors” that would allow a user to remotely access or control them.
What is the H20?
Nvidia introduced the H20 chip in early 2024 after a previous control-compliant chip, the H800, was banned from being sold to China under the Biden administration’s update of chip-export controls in late 2023. The H20 is part of Nvidia’s Hopper family of chips and is a reduced-power version of its H100 chip. OpenAI’s latest GPT-5 AI model was trained on H100 and H200 chips.
The H20 is considered a less powerful chip because of its performance constraints, which can be determined by a few different metrics, as Janet Egan, a senior fellow at the Center for a New American Security, explained to MarketWatch. For one, chips are measured by floating point operations per second, or FLOPS, which is the amount of computational operations the chip can do and how quickly.
While the H20 has less compute power than the H100, making it a less attractive chip for AI training, Nvidia had to keep the chip competitive in the Chinese market, and that came in the form of more high-bandwidth memory, Egan said. The H20 uses HBM3 memory, which is a few generations behind the latest HBM4 that is currently being developed.
But according to independent research firm SemiAnalysis, the move actually made the H20 20% faster at AI inference than the banned H100. Inference comes after training when an AI model makes its own predictions.
When initial export controls were introduced, the tech community did not yet fully understand the role of inference in frontier AI development, Egan said. At the time, most companies were still mostly focused on AI training, which depends more on computational power.
“With the emergence of the reasoning paradigm, we’re seeing that you do want to be running inference on these models,” Egan said. “In order to refine models and run experiments – all that requires a lot of high-bandwidth memory.”
Now that inference is crucial for AI development, Egan said it should be a concern for U.S. strategic interests.
At the same time, the AI industry in the U.S. is focused on getting more HBM chips for processing increasingly vast amounts of data more quickly as AI demand continues to grow.
“HBM is expensive – it’s also harder to get, just because there isn’t as much of it,” Jack Gold, founder and principal analyst of J.Gold Associates, told MarketWatch. And there are differences between the types of memory chips being made by leaders such as Micron Technology Inc. (MU) and South Korea’s SK Hynix Inc. (KR:000660).
How is Nvidia making the case for selling its H20 chips?
Nvidia and smaller rival Advanced Micro Devices Inc. (AMD) have been sitting on inventory of their China-compliant chips, and they took big write-downs ahead of Trump’s decision to again allow sales in exchange for a cut of the revenue. Nvidia said that it took a $4.5 billion charge in its April quarter due to “excess inventory and purchase obligations” for the H20. AMD said earlier this month that inventory and related charges for its MI308 chip amounted to about $800 million for the June quarter.
The companies have missed out on business in the wake of the effective ban, but Nvidia has argued that it’s about more than just money. The company thinks it should be able to keep its presence in the Chinese market, or else the demand will go back to Chinese providers that are offering increasingly better chips, Egan said. And more demand for Chinese chips means more investment in the country’s research and development system, which Nvidia says is even more reason to keep developers in China hooked on U.S.-made technology.
However, Egan said there are still limitations and constraints to how much the Chinese domestic chip industry can actually make under export controls to date. Along with chips, the U.S. has imposed restrictions on semiconductor manufacturing equipment and chip-design software.
Even though China has been able to stockpile some components, “they don’t yet have the domestic manufacturing capability to really challenge U.S. leadership in this market,” Egan said.
China is also resource-constrained when it comes to chip-making partners, Egan said.
The country has invested heavily for years in homegrown hero Huawei, which landed on the U.S. Entity List in 2019 and is believed to be China’s closest competitor to Nvidia. That list, maintained by the U.S. Bureau of Industry and Security, includes foreign companies, individuals and other entities that are subject to certain restrictions due to national-security concerns.
Huawei’s recent Ascend 910C chip, which is essentially two Ascend 910B chips put together, can deliver about 60% of the performance of its Nvidia competitor, the H100, for AI inference, according to researchers at Chinese AI startup DeepSeek. However, the Ascend 910B chips were found to include components made by TSMC in violation of U.S. export controls prohibiting the Taiwanese chip manufacturer from shipping to Huawei, the New York Times reported.
“So there’s a fundamental question as to how many Ascend 910Cs can China make by itself,” Egan said. “They will face real constraints, so it doesn’t matter how much domestic demand you get. The constraints will be China’s ability to manufacture them.”
See more: Why Nvidia investors shouldn’t worry about Huawei’s new AI chip
Even though Nvidia says its weaker chips won’t pose a risk to U.S. AI dominance, “if you put enough H20s together, or even some of the other GPUs that Nvidia sells, you can still do some pretty powerful things,” Gold said.
Meanwhile, Nvidia can see that it won’t be able to sell its H20 chips in China for much longer because they are becoming subpar compared with ones being offered by homegrown companies, Gold said. Now the company is asking the U.S. government to allow it make whatever profit it can.
The matter goes beyond chips and gets at the heart of why Nvidia has become so dominant.
(MORE TO FOLLOW) Dow Jones Newswires
08-14-25 0830ET
Copyright (c) 2025 Dow Jones & Company, Inc.