GINA RAIMONDO appeared visibly frustrated during her address at the Reagan National Defence Forum in California in December. Leading the Department of Commerce, she had just implemented stricter regulations on the sale of American semiconductors to China. However, Nvidia, the leading chip manufacturer globally, promptly initiated the development of a new, slightly less potent artificial intelligence (AI) chip tailored for the Chinese market to bypass these restrictions. Ms. Raimondo cautioned, “If you redesign a chip…that empowers [China] in AI, I will exert control the very next day.”
This statement carried a tone of bravado, considering that her department had taken a full year to revise the restrictions to counter Nvidia’s previous workaround. The ongoing five-year crusade against Chinese technology by the United States is escalating. Recent reports indicated that Jensen Huang, Nvidia’s CEO, and two other chip executives were summoned to testify before Congress regarding their business dealings in China. Additionally, on January 19th, ABB, a Swiss industrial conglomerate, disclosed that American legislators were scrutinizing its connections with China. ABB affirmed its cooperation with the investigation, while Nvidia stated its commitment to collaborating closely with the government to ensure adherence to export controls.
Both Democrats and Republicans are unlikely to ease their stance. In a presidential election year, Joe Biden, the Democratic president facing low popularity, cannot afford to appear weak on China. His Republican predecessor and primary rival, Donald Trump, has long positioned himself as a staunch critic of China. Advocates in Washington advocating a tough approach towards China aim to thwart Chinese attempts to circumvent regulations and rebuild essential technological capabilities domestically. Nonetheless, the checkered history of export controls underscores the challenges in designing harsher measures, which may not necessarily yield greater success.
China has managed to circumvent existing controls in some instances. For example, contrary to Ms. Raimondo’s concerns, it is feasible to train AI models using chips that are not cutting-edge, provided there is an adequate quantity available. If the intention is to prohibit the sale of any chip capable of “doing AI,” as implied by Ms. Raimondo, the U.S. must restrict the supply of a much broader range of chips to China.
The exact scope of such restrictions is challenging to ascertain. Trade data does not differentiate the graphics processing units (GPUs) utilized for AI model training and execution from the broader category of integrated circuits. However, an estimate of the potential impact of such a ban can be inferred from Nvidia’s financial reports. Nvidia, a provider of various GPUs, has derived between 21% and 26% of its revenues from China in recent years. In the nine months leading to October, the company generated $8.4 billion in revenue from the Chinese market. Virtually all of Nvidia’s products are AI-capable. Mr. Huang emphasized that the company is unprepared for a scenario where access to China is cut off.
Enforcement poses another challenge for the U.S. The Department of Commerce possesses the authority to penalize any violations it uncovers. Notably, it imposed a $300 million fine on Seagate, a hard-drive manufacturer, last year for allegedly violating export controls by supplying components to Huawei, a blacklisted Chinese tech giant. However, the onus of enforcement primarily falls on the chip companies themselves. This includes verifying that their customers are not acting as fronts for Chinese entities barred from trade. This task is intricate, as highlighted by Kevin Wolf, a U.S. attorney and former official, citing the widespread availability of small devices and technologies that closely resemble controlled technologies globally.
This scenario fosters conditions conducive to smuggling, which experts suggest is rampant yet challenging to quantify. It also spurs transshipment, where firms in countries not aligned with the American export-control framework, such as Singapore, can procure chips and forward them to Chinese entities without the knowledge of U.S. firms or the Department of Commerce. Nvidia’s latest quarterly earnings for 2023 reveal a fivefold increase in sales to Singapore compared to the same period in 2022, surpassing growth rates in other regions.
Among all customers in China, the People’s Liberation Army stands out as the entity best positioned to leverage such workarounds to acquire necessary chips. If one of America’s primary objectives is to impede China’s access to advanced technology for military AI development, it appears to be falling short. Instead, these controls are inflating the costs for Chinese buyers seeking American AI chips, aligning China’s tech sector with the government’s strategy of fostering indigenous technological capabilities. Previously, Chinese tech giants preferred purchasing superior American technology over investing in research and development. However, their incentives have shifted.
This shift is evident in Huawei’s case. Initially targeted by the U.S. in 2019 for allegedly violating sanctions on Iran, Huawei was subsequently cut off from chips produced using American technology through measures like the “foreign direct product rule” (FDPR). In 2022, the FDPR was extended to encompass the entire Chinese AI industry and expanded in October to cover a broader range of AI chips and chipmaking tools, necessitating licenses for product shipments to countries like the United Arab Emirates (excluding Singapore) suspected of serving as intermediaries for Chinese buyers.
Prior to the blacklist, Huawei sourced microprocessors from TSMC, a Taiwanese contract chip manufacturer, spending $5.4 billion on TSMC-made chips in 2020 before U.S. export controls extended to the Taiwanese firm. Subsequently, Huawei has increased collaboration with SMIC, China’s leading chipmaker. While SMIC’s capabilities were believed to lag behind TSMC’s, revelations emerged last year that the company was producing Huawei-designed AI chips like the Ascend and smartphone chips like the Kirin. This shift underscores China’s redirection towards domestic sources for chips due to restricted access to foreign suppliers.
Export controls were designed to anticipate such developments. They not only target chip trade but also tools essential for chip production. Collaboration with allies like the Netherlands and Japan, home to numerous toolmakers, has expanded the control measures to include restrictions on the sophistication of equipment sold to Chinese buyers. The classification of tools falling under these controls has been a subject of intense debate akin to chip regulations.
Of particular importance are the machines used for etching transistors onto silicon wafers. Cutting-edge lithography equipment, exclusively supplied by ASML, a Dutch firm, has been off-limits to China for years. However, older generations of lithography tools remain accessible. ASML’s sales to China surged in the past year, alongside other chipmaking tool manufacturers. Chinese sales accounted for nearly half of ASML’s total revenue in the most recent quarter. Similarly, other toolmakers witness substantial sales to China.
Export controls are incentivizing Chinese toolmakers to invest in technological catch-up with foreign counterparts. Domestic toolmakers are already experiencing revenue growth, with NAURA, a manufacturer of etching tools, projecting a nearly 50% revenue increase in 2023.
The U.S. campaign against Chinese technology appears to be both ineffective and counterproductive. Ineffectiveness stems from China’s adeptness at exploiting loopholes, while counterproductivity arises from the unintended consequence of fostering a more sophisticated Chinese industry. Additionally, the campaign’s premise, hinging on the significance of AI for future economic and military power balance, remains speculative. The assumption that AI’s strategic importance is contingent on computing power is debatable, as highlighted by Chris Miller, a technology historian at Tufts University. The costliness of computing power incentivizes AI developers to utilize it judiciously.
Despite these challenges, the U.S. is poised to strengthen export controls on AI chips, as hinted by Ms. Raimondo in December. Republican lawmakers are contemplating even more expansive controls. Some perceive a new threat stemming from China’s economic prowess rather than its techno-military capabilities. Chips are increasingly integral components in diverse applications, from electric vehicles to infrastructure systems. By 2027, TrendForce estimates that China could produce nearly 40% of such semiconductors. Existing export controls do little to curb China’s dominance in this sector, which relies on older American technology.
Three congressional Republicans—Mike Gallagher, Elise Stefanik, and Michael McCaul—are crafting legislation to sever China’s access to all American chip technology, not solely advanced variants. Garnering support for such stringent measures from allies may prove challenging, given the discontent among Japanese and Dutch businesses and governments towards the existing porous controls. However, should a return to power by Mr. Trump, known for his skepticism towards alliances, materialize, the lack of support is unlikely to hinder such policies.
