New U.S. Ban on Nvidia’s H20 Chip: A Gift to Investors and a Kick in the Pants for U.S. Regulators
By George Gilder and Robert Castellano
Earlier this month, on April 9, Nvidia disclosed that the U.S. government now requires an export license for its H20 processors and any chips matching their bandwidth profile. This was unwelcome news. The H20 was deliberately designed and produced to meet the fears of U.S. regulators who had banned NVDA’s high-performance, AI-adept hardware such as the A100 and H100 from being exported to China.
Naturally, the bans were only imperfectly enforced. The U.S. repeatedly complained that these high-performance chips were getting to the Chinese anyway. But why respond to leakage in the ban on high performance chips by banning chips with low performance imposed specifically to meet U.S. fears?
Not, apparently, to punish Nvidia, though that company had to take a $5 billion write down to account for H20 chips that may now have no market.
The real target seems to be DeepSeek and other Chinese companies that have succeeded in training AI models using low-powered chips such as the H20. This was not supposed to be possible. Surprise! The Chinese figured it out. Add this to the list of impossible things the Chinese do before breakfast, such as recently producing “7nm” and now “5nm” processors without banned EUV (extreme ultraviolet) photolithography machines from ASML, deemed essential by “experts.”
Table 1 shows the capacity of NVDA’s A100, H100, H20 and L20 processors, with the H20 and L20 having been designed to conform to U.S. sanctions. On memory bandwidth and interconnect bandwidth—crucial metrics for AI functionality—the H20 has less than half the capacity of the H100, but was banned anyway by the April 2025 order.
Source: The Information Network
Table 2 shows which chips several Chinese companies, plus G42 from the UAE and America’s OpenAI have used to support their AI models. Both DeepSeek and Baidu are using the H20 for AI, though apparently in combination with other, more powerful NVDA chips. (The “risk profile” is defined by U.S. agencies, principally the Department of Commerce, based on several factors including what the hardware can be used for, its intended use, and to where it is being shipped.)
Still a few bugs in the system
The higher-powered chips listed on the table have been somewhat available in China despite export rules in place since 2022, because Nvidia continued to ship those chips to foreign distributors, which then re-sold them into China. NVDA did little to track end-users. U.S. intelligence officials began flagging this activity in 2023, particularly as AI computing capacity in China accelerated beyond what should have been possible under export constraints.
The newly imposed restrictions on the H20 or any chip with comparable performance effectively rips up Nvidia’s carefully drawn China export roadmap.
The actual aim of this policy, however, is not to cripple Nvidia, but to stem China’s AI development. DeepSeek, China’s most advanced large language model (LLM) initiative, has leveraged Nvidia hardware—likely including some H100-class processors acquired indirectly through third parties—to build GPT-level systems. Supplementing hard to get H100s with legal H20s has been a great strategy for DeepSeek
The U.S. intelligence community now sees DeepSeek as a potential threat, akin to a digital surveillance balloon—an AI platform that could serve military, espionage, and cyberwarfare functions under a civilian guise.
Investor Takeaway
Nvidia’s $5.5 billion write-down has dominated headlines and further cratered the stock, now selling at more than 35% off its 52 week and all-time high. One-time charges of this kind are generally a gift to investors as Mr. Market almost always over-reacts on the downside.
Meanwhile, healthy companies are more likely to overstate than understate their real losses. That’s almost certainly the case here as inventory tied up in H20-class parts may be repurposed or delayed—not lost entirely. The gift is a little less here, only because the stock had already been so badly battered.
In this free letter, we do not share our formal valuations of the companies we cover. We hold that stuff for subscribers to the Gilder’s Technology Report. But if on your very own you conclude than NVDA is shockingly undervalued, we’re not “sayin’ nuthin’.”