Where There's A Will, There's A Way To Get US Chips Into China

US trade restrictions have made it harder for Chinese companies and government agencies to get their hands on advanced semiconductor technologies, but apparently not impossible.

Over the past few years we've tracked a number of shady chip busts, some quite bizarre such as smuggling processors in prosthetic baby bumps, or inside trouser legs, and even packing GPUs alongside live lobsters. However, other efforts are more organized and involve seemingly legitimate businesses that funnel goods and intellectual property to the folks that Uncle Sam would rather did not have them.

The US government knows this is happening and even has a term it uses to describe such efforts: "Evasion routes." And where smuggling the real thing isn't possible or practical, corporate espionage is also on the table.

Shell games

These smuggling operations often follow a common pattern. Chips are first imported to a company located in a country where they're legal. Once beyond the reach of US authorities, the books are cooked and the semiconductors are then smuggled into China, Russia, or another country of concern.

This strategy is by no means new or unique to China or semiconductor tech. In fact, during the Cold War, the US Central Intelligence Agency allegedly used dummy companies to acquire titanium for military aerospace applications. This is also reportedly how the Chinese agency responsible for developing and maintaining the nation's nuclear weapons stockpile has been getting its hands on Nvidia and Intel processors, despite decades-old sanctions designed to prevent this from happening.

Two more recent examples include a scheme by two South Koreans who were busted last week after smuggling 96,000 US-derived chips into China between 2020 and 2023.

In another example, reported by The Wall Street Journal, autonomous trucking outfit TuSimple allegedly tried to ship Nvidia A100 GPUs to Australia and then on to China.

Introduced in 2020, Nvidia's A100 remains one of the most sought-after GPUs for AI training and inference applications. In late 2022, the US government moved to bar the sale of the accelerator, and the more powerful H100, in China. More recently, the US Commerce Department extended export restrictions to include the majority of modern datacenter GPUs and even a few consumer-grade graphics cards authorities felt China could use for AI applications.

In that case, authorities believe that TuSimple's executives intended to first ship the GPUs to Australia, where they were in the process of relocating, before passing them on to their Chinese division. However, the 24 GPUs – about enough for three GPU nodes – never made it out of the country. They were seized in transit by US Authorities, which were already wary of TuSimple's ties to China.

Despite the relatively small number of accelerators involved, the US Commerce Department's handling of the situation isn't all that surprising considering multiple embarrassing reports claiming export controls have done little to stop the flow of equipment to China, Russia, and other nations of concern.

However, because these transactions often appear legitimate, American companies have little incentive to ask too many questions about where their kit is going and who's using it.

The US, for its part, has attempted to stamp out these backchannels. If an organization is suspected of violating sanctions, it's only a matter of time before the business in question lands on the Entities List.

For example, last spring Inspur was accused of arming China's military with American equipment. However, in many cases these setups are little more than a front, making efforts to blacklist them a bit like a game of Whac-A-Mole.

Otherwise it appears that the number of advanced semiconductors smuggled into China remains fairly small. A report published late last year by the Center for New American Security, a Washington DC-based think tank, postulated that 12,500 of these chips reasonably could be smuggled into China each year. However, the authors admit that the estimate doesn't have the greatest foundation.

We suppose black market GPU dealers probably aren't in the habit of sharing how much product they're moving.

The group did, however, note that recent changes in trade policy could drive larger, potentially state-backed smuggling operations in the future.

• China puts homegrown GPUs and other AI infrastucture on its national to-do list
• The latest cold war is already being fought in the supply chain trenches
• US shorts China's Volt Typhoon crew targeting America's criticals
• Eyeing China, US may require clouds to report when foreign actors rent kit to build AI models

Cloudy complications

Further complicating the issue, widespread availability of GPU resources in the cloud means that, for the moment, Chinese companies can just rent time on accelerators that they're otherwise barred from buying.

US intelligence agencies have raised alarm bells about this loophole. Late last year, a New York Times report claimed US agencies suspected United Arab Emirates' AI darling G42 of secretly furnishing China with compute resources. The revelations ultimately drove the company to cut ties with Chinese equipment vendors.

~In the past week the US Commerce Department published a proposed regulation that would require American Infrastructure-as-a-Service (IaaS) providers to implement more stringent know-your-customer procedures to prevent persons or countries of concern from renting AI resources in violation of US export restrictions.

Under the proposed rules, IaaS providers would be required to report any foreign customers using their services to train AI models with potential dual use capabilities. In this context, dual use refers to models that could be employed for military or intelligence gathering purposes.

While this might make it harder for Chinese companies to openly train AI models in the cloud, it seems unlikely these measures will prevent them from using proxy companies in the US to avoid scrutiny.

Even if the US is successful in walling off access to cloud-based AI resources, the superpower's ability to police other nations only goes so far.

It's not hard to imagine a scenario in which a Chinese company secretly finances AI datacenters built in countries where American made GPUs are legal, but US influence is limited or non-existent. ®