China doesn’t need to participate in Donald Trump’s pay-to-play deal with Nvidia, apparently.
Hot on the heels of news that the Trump administration will allow Jensen Huang to ship outdated, but still capable, H200 chips to approved customers in China, reports indicate that DeepSeek, the startup which briefly upended the entire global AI trade earlier this year, is already using the Blackwell.
How, you might ask, is that legal? Well, it’s not. But where there’s a will, there’s a way.
In this case, DeepSeek apparently colluded with data center operators in countries where Nvidia’s most advanced chips are permitted. According to The Information, the servers at those data centers were taken apart and shipped “in pieces” to China.
The article indicates DeepSeek used “several thousand” of the smuggled Blackwells to develop its next “major model.” The reporting, based on the accounts of half a dozen unnamed sources, described the scheme as “convoluted.”
There are a couple of ways to look at this. First, it suggests that for all the hoopla around domestic efforts to design and manufacture advanced chips, Chinese AI firms still very much want Nvidia’s technology and plainly still view it as superior to homegrown alternatives.
Second, it casts doubt on the (already dubious) notion that DeepSeek somehow possess the wherewithal to develop world-beating models for pennies on the dollar and without cutting-edge technology. That claim was the centerpiece of what the annals of market history will remember as “The DeepSeek scare.”
Third, it attests to the age-old adage that where there’s demand someone will supply. I realize the narcotics callbacks may feel gratuitous to those of you addicted to the Monthly Letters, but this is no different logistically from high-stakes, international drug smuggling: A for-profit scheme to export contraband through an elaborate ruse which surely involved paying off any number of intermediaries up to and including customs officials. Frank Lucas would be proud.
Fourth and finally, this underscores Huang’s contention that a strategy which tries to freeze Chinese AI developers out by denying to them the best chips is doomed to fail one way or another. That’s a flagrantly self-serving position on his part, but that doesn’t mean he’s wrong.
There’s now a very large black market for Nvidia’s chips, an unintended — but wholly predictable — consequence of US export restrictions. Arguably, the best outcome is Chinese companies procuring Nvidia chips illicitly, because the alternative is the development of homegrown technology which eventually becomes advanced enough that China achieves total AI self-sufficiency.
Although I realize it’s an exercise in futility, I nevertheless believe it’s necessary to impose the restrictions. Xi Jinping simply can’t be trusted. He’s impossible to read and as a psychotic ideologue, he’s not amenable to the sort of transactional relationships which make it “safe” (note the scare quotes) for the US to green-light sales of Nvidia’s most advanced chips to autocracies like Saudi Arabia.
In any event, and as alluded to above, the news at least suggests China isn’t anywhere near achieving the kind of domestic know-how that’d render export restrictions on advanced US technology redundant and moot.
Exactly.
If you ever tire of referencing drug trade analogies (which is a perfect analogy), you can also cite commodity trading (especially oil)- in both past and present situations where the traders trade for financial gain without regard for moral or legal consequences (other than not wanting to get caught). There are numerous examples, but the stories of oil trading in Libya, thereby funding both Gaddafi and the rebels through the back door, are fascinating. Currently, we’ve got Venezuelan and Russian oil trading. This is normal, not the exception.
The Nvida GPU CHIPS are superior for the initial training of LLMs. After that, the advantage is less compelling, as evidenced by efforts by Microsoft and others to bypass NVDA GPUs for inference.
But what was more significant about DeepSeekand other Chinese models is a different approach to analyzing.and using data in a fashion relevant to the end user.
If one is curious sbout this, I recall that in the past week the Medium published an indepth explanation of the more efficient approach the Chinese firms are using versus the brute force “big is better” US models employ.
Excellent and accurate analysis of NVIDIA and China situation Mr H. The official policy and posturing of both countries is a distraction and is being circumvented as demand for computing power is relentless. People are still stuck in the simplistic notion that AI is merely a new iteration of advanced software. Instead, it is a whole new computing paradigm requiring total overhaul of hardware infrastructure on a colossal scale.
You sound well informed so please tutor me on this. Are you speaking of the training step or inference?
I ask because I just enjoyed an extended visit from an old friend from the early 80s when I was a (very) junior coder on a database analysis program at an investment bank while he worked as a database writer at a marketing company. I moved over to trading while he had a long career designing and maintaining databases for private companies. So he has a far better feel for what is going on in AI.
He was impressed by and excited over the new Gemini model. But he was also really psyched about the new USB protocol that allows for ultra-fast data transfer. What caught his eye was that he read about how some gearheads had used the new USB technology to daisy chain five standard iPads which allowed them to run AI models. No doubt at the inference (not training) level? He was demanding that we drive over to the Apple store and buy five iPads so we could try it out then and there. Imagine how compute and energy efficient that array would be!