|
29 January 2025 (Washington, DC) - - As I have noted in numerous posts, the world’s major economies — the U.S., China, the European Union, and others — have been instituting new policies and export controls to protect national security and economic advantages.
In the U.S., the primary industries with supply chains that involve technologies that can be used for military as well as commercial purposes — such as advanced semiconductors, artificial intelligence, and quantum computing — are the ones facing government policies aimed at protecting national security and economic competitive advantages. It is here that sanctions are most prevalent.
So it is in these supply chains that the U.S. has exercised the most sanctions against China to prevent that countries access to technologies with national-security implications — especially producers of semiconductors and artificial intelligence.
Ah, would that it be so simple.
Since the beginning of the Ukraine war, the ineffectiveness of Western sanctions on Russia has become increasingly clear. And so the same is true of US sanctions against China — especially the Biden’s administration’s ban on high-performance semiconductors and, more recently, on graphics processing units (GPUs). The foreign policy community, illiterate in matters of economics and finance, massively overestimated the effect such sanctions would have, as the measures rely on the idea that China’s rise is based entirely on U.S. technology.
The response to Chinese AI company DeepSeek and its destabilization of U.S. tech stocks in recent days demonstrates how this strategy has backfired. Advocates for sanctions still argue that they will have an effect ... if only we wait long enough.
Yet the very opposite is true. The longer that sanctions are applied, the more likely it is that countries will find ways to circumvent them — or in the case of DeepSeek, simply outsmart them.
Since at least the turn of the millennium, Beijing has been crowding in on the so-called Stem subjects (science, technology, engineering and mathematics), with American and British universities particularly popular among Chinese students. China has its own highly competitive university sector that is also heavily focused on these subjects (media studies seems to be of lesser interest).
The West maintains a few inherited advantages. Patented proprietary technologies exist to protect. But as the car industry has now found out, this works until it doesn’t. China was previously ineffective at producing semiconductors, as recounted by Chris Miller in his book "Chip Wars". Then it got very, very good at it.
Note to readers: Miller's book masterfully unfolds the intricate story of how semiconductors evolved to become the linchpin of our modern existence. Pause for a moment, and redirect your gaze from this screen to the device that holds it (assuming you're reading this on your mobile). Concealed beneath its sleek exterior are intricate chips, diverse and sophisticated, enabling your machine to seamlessly execute a myriad of functions. If you want a good overview on how that whole world is put together and works, Miller's book is the place to start.
But this was only a snapshot, rather than a detail from which to extrapolate. Since the various restrictions imposed by the U.S. on ASML, Europe’s most important tech company, China has been improving its production capacity. With talent, money and natural resources, why wouldn’t the country succeed eventually?
Sanctions are not only failing to accomplish their task: they are actively counterproductive. DeepSeek, as a product of a sanctions regime, needed to develop its own system — and choose a route that was more efficient and less costly than the one taken by the U.S.
But it is the unintended side effects of sanctions which should concern us. They caused economic shifts in Western countries for which those governments were unprepared, as well as giving China the opportunity to revamp its technology production systems.
The Russia sanctions were a factor in the deindustrialization of Germany and, by undermining Olaf Scholz’s governing coalition, in the rise of the AfD. That has been covered in detail by my boss, Greg Bufithis, so I will leave it to you to thread through his posts.
As for China and DeepSeek, the conventional thinking held that developing leading AI required loads of expensive, cutting-edge computer chips — and that Chinese companies would have trouble competing because they couldn’t get those chips.
So DeepSeek decided it would rebuild, and "re-innovate" its systems. In an interview with a Chinese publication in 2023 that somebody recently found, the founder of DeepSeek said almost all of the technical positions at the company were filled by fresh graduates or people with one or two years of experience. He hired 1000s of them.
Experience, he said, was a potential obstacle:
"When doing something, experienced people will tell you without hesitation that you should do it this way. But inexperienced people will have to repeatedly explore and think seriously about how to do it, and then find a solution that suits the current actual situation".
And these "inexperienced" engineers and computer scientists discovered something.
Until recently, the pioneering AI models that lie behind programs such as OpenAI’s ChatGPT were trained on a vast compilation of text, images and other data. They employed specialized algorithms to find patterns that a chatbot could use to hold a conversation.
But the DeepSeek team said "Hold on ..."
The team's tactic was to cut down on the data processing needed to train the models, using some very innovative inventions and techniques of its own, some adopted by similarly constrained Chinese AI companies.
Imagine the earlier versions of ChatGPT as a librarian who has read all the books in the library. When asked a question, it gives an answer based on the many books it has read.
This process is time-consuming and expensive. It takes electricity-hungry computer chips to read those books.
DeepSeek took another approach. Its librarian hasn’t read all the books but is trained to hunt out the right book for the answer after it is asked a question.
Layered on top of that is another technique, called “mixture of experts.” Rather than trying to find a librarian who can master questions on any topic, DeepSeek and some other AI developers do something akin to delegating questions to a roster of experts in specific fields, such as fiction, periodicals and cooking. Each expert needs less training, easing the demand on chips to do everything at once.
DeepSeek’s approach requires less time and power before the question is asked, but uses more time and power while answering. All things considered, DeepSeek’s shortcuts help it train AI at a fraction of the cost of competing models.
DeepSeek "gets it". Engineering is about constraints. The Chinese engineers had limited resources, and they had to find creative solutions.
Ok, to be fair, ingenuity explains only part of DeepSeek’s success. There is a dark side, the "sneaky" part to which I referred in my subtitle above. It is about U.S. export controls. DeepSeek got a "window" to buy powerful American chips.
The Biden administration in 2022 put in place controls on chips exported to China. U.S. companies that wanted to sell to China first needed to throttle a chip function called interconnect bandwidth, which refers to the speed at which data is transferred.
In response, Nvidia, the world’s leading designer of AI chips, came up with a new product for China that complied with this parameter — but compensated for it by maintaining high performance in other ways. That resulted in a chip that some analysts said was almost as powerful as Nvidia’s best chip at the time.
U.S. officials vented publicly and privately that while Nvidia didn’t break the law, it broke the spirit of it. The government had hoped that industry leaders would be collaborative in designing effective export controls on fast-changing technology.
A year after the initial controls, the government tightened the rules. Still, that left an opening of about a year for DeepSeek to buy Nvidia’s powerful China-market chip, called the H800. In a research paper published last month, DeepSeek acknowledged it used 2,048 of these Nvidia chips to train one of its AI models.
Since the rules were revised in 2023, Nvidia designed a new export-control-compliant chip for China that is significantly less powerful than the H800.
And many American AI industry leaders are skeptical that DeepSeek has revealed all of its secrets over the last week. They said Chinese researchers could have stockpiled leading-edge Nvidia chips before the U.S. restrictions, or used workarounds such as accessing Nvidia-enabled computing power from countries outside the U.S. and China.
But going further down the road, perhaps the most disastrous geopolitical effect is that what the U.S. actually did was to pull China and Russia into a new anti-Western strategic alliance.
And so America and Europe must now reckon with a coming multipolar order they are ill prepared for, in which global influence is spread more evenly among and yet beyond the old powers.
The game has only just begun.
| 
