Trump wants tighter AI chip export restrictions but may face staffing shortage and other issues

Nvidia Hopper HGX H200
(Image credit: Nvidia)

The Trump administration is looking to tighten restrictions on exports of advanced AI processors from the country as well as persuade allies in Japan and the Netherlands to make Tokyo Electron and ASML stop servicing their tools installed in China, reports Bloomberg. While the move will certainly make the lives of Chinese chipmakers harder, it will affect U.S. allies and not in a good way.

Fewer AI processors for everyone

When President Joe Biden was prepping to leave the White House in early January, his administration proposed its final set of export restrictions on advanced AI processors. Under the new guidelines, called the AI Diffusion Rule, only entities from the U.S. and 18 allies (Tier 1 countries) can obtain powerful AI processors — such as Nvidia's H100 GPUs — without limits. Other foreign entities (Tier 2 countries) will have restricted access to AI GPUs unless they secure validated end user (VEU) status. Countries under arms embargoes (Tier 3 countries), including China, Russia, and Macau, will be almost entirely prohibited from importing AI GPUs. Trump's government is reviewing this framework to streamline and reinforce its efficiency, according to the report.

Under the Biden's AI Diffusion Rule, companies from Tier 2 countries can get up to 1,700 Nvidia H100 GPUs (or equivalent) without any export license. These orders will also do not count toward the national AI processor limit of around 50,000 units. One proposed change is reducing the number of AI chips that can be exported without government approval (i.e., from 1,700 to a lower number). Some people in the government want to lower this threshold and increase oversight over AI processors exports.

The move will not be exactly welcome by the industry as tech industry leaders were pushing back against AI Diffusion Rule restrictions. Nvidia's CEO, Jensen Huang, has expressed optimism that the Trump administration would adopt a more flexible approach. Companies worry that stricter limits on chip exports could harm business interests while encouraging Chinese firms to develop domestic alternatives faster.

Another hit on Chinese semiconductor sector

Officials from the U.S. government have held meetings with representatives from Japan and the Netherlands to push for restrictions on maintenance services for semiconductor manufacturing equipment in China. The goal is to prevent companies like Tokyo Electron and ASML from servicing chipmaking tools in China, mirroring U.S. restrictions on domestic firms such as Applied Materials, KLA, and Lam Research.

Without regular maintenance, tools from ASML and Tokyo Electron can quickly become obsolete. If allies agree to U.S. demands, Chinese semiconductor production could face serious disruptions. However, ASML and TEL will also lose tens of millions of dollars that Chinese companies pay for their services.

In addition, there are discussions underway in Washington about imposing sanctions against specific Chinese semiconductor firms. One proposal targets ChangXin Memory Technologies (CXMT), a company that was considered for restrictions by the previous government, but was spared due to Japan's opposition. Trump officials are now revisiting the possibility of blocking it from acquiring American chipmaking tools completely by adding it to the U.S. Department of Commerce Bureau of Industry and Security's (BIS) Entity List.

The administration is also looking to impose tighter restrictions on Semiconductor Manufacturing International Corp. (SMIC), which makes chips for blacklisted Huawei. Biden had restricted shipments to certain SMIC facilities while allowing case-by-case reviews for others. However, this approach allows SMIC to obtain U.S. tools that could be transferred from one facility to another to support production of processors on advanced nodes.

Despite these plans, it may take several months before any new regulations take effect. Trump's administration is still in the process of staffing key federal agencies, and allies may not immediately support these initiatives. To that end, Biden's AI Diffusion Rule will come into effect on May, whereas ASML and TEL will continue to service their tools at Chinese fabs.

Anton Shilov
Contributing Writer

Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.

Read more
Nvidia Hopper HGX H200
Nvidia and SIA fire back at US gov's new export restrictions on AI GPUs to China
GlobalFoundries
In a private letter to U.S. President Biden, semiconductor industry groups blast incoming export rules
Nvidia AI
Nvidia blasts Biden’s new AI restrictions as ‘misguided’ – GPU giant praises Trump’s hands-off approach to tech industry
Nvidia Grace Hopper superchips
New export regs may see Israel requiring a license to buy U.S. chips developed in the country
EU Flag
EU protests new US AI chip restrictions — some countries face GPU caps
Nvidia
US AI Diffusion Policy may harm Nvidia's sales — most of the chipmaker's AI GPUs are affected
Latest in Tech Industry
Inspur
US expands China trade blacklist, closes susidiary loopholes
Qualcomm
Qualcomm launches global antitrust campaign against Arm — accuses Arm of restricting access to technology
ChatGPT Security
Some ChatGPT users are addicted and will suffer withdrawal symptoms if cut off, say researchers
Intel
Pat Gelsinger supportive of Lip-Bu Tan, warns him about 'the short-termism of Wall Street'
TSMC
Producing wafers at TSMC Arizona is only 10% more expensive than in Taiwan: TechInsights
Intel
Pat Gelsinger becomes executive chairman, head of technology at church-focused platform Gloo
Latest in News
Inspur
US expands China trade blacklist, closes susidiary loopholes
Qualcomm
Qualcomm launches global antitrust campaign against Arm — accuses Arm of restricting access to technology
Nvidia Ada Lovelace and GeForce RTX 40-Series
Analyst claims Nvidia's gaming GPUs could use Intel Foundry's 18A node in the future
Core Ultra 200S CPU
An Arrow Lake refresh may still be on the cards with only K and KF models, claims leaker
RX 9070 XT Sapphire
Lisa Su says Radeon RX 9070-series GPU sales are 10X higher than its predecessors — for the first week of availability
RTX 5070, RX 9070 XT, Arc B580
Real-world GPU prices cost up to twice the MSRP — a look at current FPS per dollar values
  • bit_user
    The article said:
    Companies worry that stricter limits on chip exports could harm business interests while encouraging Chinese firms to develop domestic alternatives faster.
    That's already been happening for quite a while. I doubt they can develop any faster.
    Reply
  • JTWrenn
    The biggest issue I have with this is from everything I can see...it just doesn't stop the chips from getting there. It still happens just through black market means. It's one of those waste of money ideas that seems like it makes sense. We are giving away our market share for chips to keep AI services ahead and I just don't think it works out. I think you can draw a directly line of this sort of thing to DeepSeek pushing efficiency so much, and I don't think the Chinese government has any problem getting around these blocks for their higher end government programs.

    Feels a lot like the Russian oil embargos of the last few years that just ended up enriching them because there is always some way to do it. Unless we have a lot of countries on our side for this, it just doesn't work out.
    Reply
  • bit_user
    JTWrenn said:
    The biggest issue I have with this is from everything I can see...it just doesn't stop the chips from getting there.
    We don't know exactly what they mean by Validated End User status. That's probably the key.

    JTWrenn said:
    I think you can draw a directly line of this sort of thing to DeepSeek pushing efficiency so much,
    It's not like DeepSeek is better, though. It's competitive in some areas, and turns out actually to have used quite a lot of GPU power to train. There's no doubt in my mind that they'd be further ahead in AI, if they could buy all the compute they wanted. The reason DeepSeek caught so much attention is precisely because it's an outlier, rather than the norm.

    JTWrenn said:
    oil embargos of the last few years that just ended up enriching them because there is always some way to do it.
    No, they weren't counterproductive. That oil had to be sold at a discount, due to the smaller market for it. Just because the market didn't shrink to zero doesn't mean the measures had no effect.

    JTWrenn said:
    Unless we have a lot of countries on our side for this, it just doesn't work out.
    That's why the list of unrestricted countries is so small. It has to be countries with the means and will to minimize trafficking, or else we end up with the black market situation you mentioned.
    Reply
  • TCA_ChinChin
    bit_user said:
    It's not like DeepSeek is better, though. It's competitive in some areas, and turns out actually to have used quite a lot of GPU power to train. There's no doubt in my mind that they'd be further ahead in AI, if they could buy all the compute they wanted. The reason DeepSeek caught so much attention is precisely because it's an outlier, rather than the norm.
    I'd say they are competitive in many areas and still used less GPU power to train compared to western models. And after the release of DeepSeek, Alibaba followed suit with capabilities not far behind (albeit not open source).
    Reply
  • HideOut
    Admin said:
    President Trump's administration believes restrictions on AI processors exports are not strict enough for close allies, also plans to persuade Japan and the Netherlands to curb services of ASML and TEL machines in China.

    President Trump wants tighter AI chip export restrictions but may face staffing shortage and other issues : Read more
    The real problem with this is we dont export them. Our engineers/companies design them but they are made in Taiwan. Its basically china. We are actually importing them, not slowing their exports...
    Reply
  • zsydeepsky
    bit_user said:
    It's not like DeepSeek is better, though. It's competitive in some areas, and turns out actually to have used quite a lot of GPU power to train. There's no doubt in my mind that they'd be further ahead in AI, if they could buy all the compute they wanted. The reason DeepSeek caught so much attention is precisely because it's an outlier, rather than the norm.

    DeepSeek is currently releasing 5 of their in-production repositories. which provides a glimpse into what they use. For example this FlashMLA: 1893836827574030466View: https://x.com/deepseek_ai/status/1893836827574030466

    it's a perfect demonstration of why the US sanctions failed to stop them: they made this FlashMLA specific for H800 (a severely nerfed H100 so it can sell in China) and made it on par with H100 in computing power & Memory bandwidth.

    which means, DeepSeek got their work done on hardware they can acquire legally.

    which also means, there's no room for any further sanctions to stop them: either you also limit their access to already crippled cards to even more crippled cards (something as low as modded 48/96G VRAM 4090) which will lead to a total embargo of consumer level graphic cards, or they simply just switch to Huawei's Ascend 910C which is already more capable than H800.

    and Financial Times just had a report yesterday about 910C manufacturing yield rising to ~40% and aiming for 60%, which means the quantity will also not be a problem as well:
    Huawei Ascend 910C reportedly hits 40% yield, turns profitable; aims for 60% industry standard
    I even wonder, with DeepSeek's newly earned reputation, their deep understanding of GPU algorithm optimization, and the support from the Chinese central government, could they have been invited to the 910D development process? if so, we might see a new type of hardware that provides a magnitude of performance gain for AI tasks, even on the older nodes.

    and even Nvidia is doing this kind of optimization since the demand for R1 deployment is just enormous...
    1894172956726890623View: https://x.com/NVIDIAAIDev/status/1894172956726890623
    Reply
  • rambo919
    JTWrenn said:
    The biggest issue I have with this is from everything I can see...it just doesn't stop the chips from getting there.
    That is not the primary goal though, the goal is long term national and economy security.

    The US tech industry has basically been farmed off and they are now too dependent on foreign factories.... this is like trying to steer a massive ship.... turning takes a long time.

    The AI part is just the newest concern... and it's even more difficult because no one actually knows the capability of anyone else... they just know what gets released to the public for PR and social programming.
    Reply