Friday, January 31st 2025

US Investigates Possible "Singapore" Loophole in China's Access to NVIDIA GPUs

Today, Bloomberg reported that the US government under Trump administration is probing whether Chinese AI company DeepSeek circumvented export restrictions to acquire advanced NVIDIA GPUs through Singaporean intermediaries. The investigation follows concerns that DeepSeek's AI model, R1—reportedly rivaling leading systems from OpenAI and Google—may have been trained using restricted hardware that is blocked from exporting to China. Singapore's role in NVIDIA's global sales has surged, with the nation accounting for 22% of the chipmaker's revenue in Q3 FY2025, up from 9% in Q3 FY2023. This spike coincides with tightened US export controls on AI chips to China, prompting speculation that Singapore serves as a pipe for Chinese firms to access high-end GPUs like the H100, which cannot be sold directly to China.

DeepSeek has not disclosed hardware details for R1 but revealed its earlier V3 model was trained using 2,048 H800 GPUs (2.8 million GPU hours), achieving efficiency surpassing Meta's Llama 3, which required 30.8 million GPU hours. Analysts suggest R1's performance implies even more powerful infrastructure, potentially involving restricted chips. US authorities, including the White House and FBI, are examining whether third parties in Singapore facilitated the transfer of controlled GPUs to DeepSeek. A well-known semiconductor analyst firm, SemiAnalysis, believes that DeepSeek acquired around 50,000 NVIDIA Hopper GPUs, which includes a mix of H100, H800, and H20. NVIDIA clarified that its reported Singapore revenue reflects "bill to" customer locations, not final destinations, stating most products are routed to the US or Western markets.
The company emphasized compliance with export laws but acknowledged Singapore's significant growth in trade. Meanwhile, Howard Lutnick, a Trump nominee to lead the Commerce Department, accused DeepSeek of evading US restrictions during his confirmation hearing, pledging rigorous enforcement of chip sales limits if appointed. Authorities have yet to conclude investigations, leaving questions about Singapore's role unresolved.
Source: via Tom's Hardware
Add your own comment

15 Comments on US Investigates Possible "Singapore" Loophole in China's Access to NVIDIA GPUs

#1
ZoneDymo
Look, I know people arnt stupid...but everyone and their mother predicted this when cards were going to be restricted for china.....so, what is up with this now?

EDIT: ah yeah, probably something to do with the Trump administration....
Posted on Reply
#2
TumbleGeorge
ZoneDymoEDIT: ah yeah, probably something to do with the Trump administration....
Rather, it is an inherited activity from the "previous" administration. Fortunately, or unfortunately, the replacement of the administration and the goals pursued is neither a quick process nor can it be comprehensive.
Posted on Reply
#3
Count von Schwalbe
Nocturnus Moderatus
Deepseek just unveiled their model. Not a coincidence they are now getting probed.
Posted on Reply
#5
MacZ
Restriction on exporting to a country while allowing export to other countries makes this just a game of Whack-a-mole.

When there is demand, there is a market (even black). It just makes things harder, not impossible. Are not hard drugs prohibited in the US ?

You need plutonium level of control if you want to completely restrict China, but that would be catastrophic to the economies of scale needed to produce such hardware.

This is a performative announcement as people should expect from politicians.

It was reported a while ago that renting H100 compute time was cheaper in China than the US.
Posted on Reply
#7
fec32a4de
MacZRestriction on exporting to a country while allowing export to other countries makes this just a game of Whack-a-mole.

When there is demand, there is a market (even black). It just makes things harder, not impossible. Are not hard drugs prohibited in the US ?

You need plutonium level of control if you want to completely restrict China, but that would be catastrophic to the economies of scale needed to produce such hardware.

This is a performative announcement as people should expect from politicians.

It was reported a while ago that renting H100 compute time was cheaper in China than the US.
Not to be that guy, but "Plutonium level of control" is weird in US and Canada (and France - separate rules).

You can obtain half an ounce of Plutonium, just need NRC form 30.37 I believe (I forget the form number) and $$$. Element collectors can do it and stuff.

But yes, you are correct. Hell I've met people shipping these GPUs to China and Russia. At an extreme markup. It did turn out to be far more profitable than selling drugs
Posted on Reply
#9
AnotherReader
axiumoneThat's not how any of this works. Running a model and training a model are two entirely different things. You can run the model on fairly modest specs from a decade ago.
Well, to be fair, for inference, you would run into limitations due to lower memory bandwidth from lower spec cards of that time. Even higher spec cards would struggle, considering none of them supported fp16 at double the rate of fp32.
Count von SchwalbeDeepseek just unveiled their model. Not a coincidence they are now getting probed.
They were fairly transparent that they used the H800 which is a H100 with halved NVLink bandwidth. It's the old story; restrictions lead to innovation.
Posted on Reply
#10
Count von Schwalbe
Nocturnus Moderatus
AnotherReaderThey were fairly transparent that they used the H800 which is a H100 with halved NVLink bandwidth. It's the old story; restrictions lead to innovation.
I am not arguing; just saying that it is just pretty common for publicity to attract regulatory attention. Not a political thing.
Posted on Reply
#11
igormp
AusWolfThey might not have needed to buy any Nvidia GPU.
www.techpowerup.com/331776/amd-details-deepseek-r1-performance-on-radeon-rx-7900-xtx-confirms-ryzen-ai-max-memory-sizes
Adding to what was said above, this article is talking about running the distilled models that are way smaller, and based on other existing models (llama and qwen), this has absolutely nothing to do with training the big MoE model.
Both the 4090 and 7900xtx have nothing to do with the GPUs that are actually used to train such big models (think A/100/H100/B200 from Nvidia and the Instinct lineup from AMD).
axiumoneYou can run the model on fairly modest specs from a decade ago.
You can't run the actual MoE model in a reasonable manner with any kind of consumer hardware tho.
Posted on Reply
#12
kondamin
next month American government has Malaysia probed over loophole.
Posted on Reply
#13
freeagent
kondaminnext month American government has Malaysia probed over loophole.
Now you're thinking like an apex predator :rolleyes:

:laugh:
Posted on Reply
#14
Count von Schwalbe
Nocturnus Moderatus
freeagentNow you're thinking like an apex predator :rolleyes:

:laugh:
He plays whack-a-mole
Posted on Reply
#15
TumbleGeorge
In fact, didn't I read an advertising article yesterday that Nvidia is providing support for DeepSeek? In other words, this model is the new flagship concept and, if it seems that maintaining nOpenAI is too much of a waste of resources, at least let your customers who want to work with DeepSeek know that they can work perfectly with it on Nvidia hardware and continue to buy it, even if it's in smaller quantities than they would need to play with the more performance-demanding models.
Posted on Reply
Add your own comment
Jan 31st, 2025 21:08 EST change timezone

New Forum Posts

Popular Reviews

Controversial News Posts