The Yulin Naval Base, China’s largest, lies on the southern coast of Hainan, an island province the size of Taiwan in the northernmost reaches of the South China Sea. Yulin’s deepwater bays house two large-ship destroyer piers, six nuclear submarine piers and a pier for degaussing, a process that reduces the magnetic field of China’s warships, making them harder to detect. To the United States and its allies in the Indo-Pacific, Yulin is a $50-billion-dollar reminder not just of China’s increasingly aggressive claims in the South China Sea, but its ambitions to reach farther into the Pacific, by way of Taiwan if necessary.
Those ambitions have fueled the United States’ increased vigilance over the transfer of cutting-edge technology to China. In 2023, when the Biden administration announced its second round of restrictions on Nvidia’s most advanced artificial intelligence chips, it insisted that the move was not meant to hurt China economically. According to then-Commerce Department Secretary Gina Raimondo, the rules were aimed at “sophisticated computers that are critical to Chinese military applications.”
Last month, President Donald Trump reversed some of these restrictions, allowing Nvidia to export its H20 chip to China. He later suggested that he could go further, allowing Nvidia to sell China a pared down version of its newer and more advanced Blackwell architecture. To understand the backlash against Trump’s increasingly soft stance on chips, one must look not to Yulin Naval Base, but roughly 50 kilometers northwest, where the Ledong Air Base stands as a testament to what can happen when civilian computing is repurposed to military ends.
Ledong is home to the 809th Brigade of the People’s Liberation Army Air Force, or PLAAF. The 908th includes a group of jets called the Xi’an JH-7, a two-seat fighter-bomber that entered service in 2004. JH-7s are far from China’s most advanced plane, but there are still more than 200 of them in service today, and they specialize in precision air-to-surface operations, just what China would need in an aerial bombing campaign of Taiwan. Their development also represented a landmark in the PLA’s adoption of high-tech aircraft design.
In the 1980s, the PLAAF began a development program for replacing the outdated Nanchang Q-5 and Harbin H-5 bombers that had been its workhorses during the Sino-Vietnamese war. Up until that time, China had carried out its modification of Soviet aircraft designs by hand, using compasses, french curves and a technique called lofting, which allowed designers to transform two-dimensional drawings into three-dimensional surfaces.
Designing aircraft by hand was incredibly time consuming. By the time China began using its own computer-aided design, or CAD, software in the 1990s, it cut the design and drawing process in half. But CAD, of which the JH-7 is thought to be the first Chinese fighter jet to make use, is only half the story. Being able to quickly generate and modify an aircraft design is one thing, but determining what exactly makes a good design is another.
The Chinese approach to evaluating design quality was similarly analogue. To estimate an aircraft’s aerodynamic efficiency, engineers had to fly a prototype in a wind tunnel, or make inferences from the performance of past models. Measuring the stealth of a fighter jet, or its ability to evade detection from radar systems, also relied on real-world prototype testing. To speed up the evaluation process, China would need to adopt a mathematical technique called computational fluid dynamics, or CFD, a notoriously complex field of mathematics that allows one to calculate forces like lift and drag.
Though CAD and CFD both started to be adopted in military design around the same time, from the perspective of computational load, they were two completely different beasts. In CAD software, the variables are the preferences of the designer — a bit sharper on the curve here, a bit shallower on the dent there. But in CFD, the variables are the movements of free molecules, all colliding with one another and reacting to the shape and movement of the aircraft over time. Calculations can quickly spiral out of control. To integrate CAD, the Chinese military needed to retrain designers used to working with pencil and paper. But to efficiently evaluate and modify digital designs, they would need a supercomputer.
In November of 1997, President Bill Clinton signed into law the National Defense Authorization Act for Fiscal Year 1998. Among its provisions were the requirement that all companies involved in the Department of Energy’s Accelerated Strategic Computing Initiative must report to the secretary of defense any sale to a Tier III (or third world) country of a computer capable of exceeding two billion theoretical operations per seconds. Nvidia’s H20 chip can handle 296 trillion theoretical operations per second, but in the late 1990s, two billion was near the cutting edge.
Earlier that year, an investigation by the House of Representatives found that supercomputers originating in the U.S. had been used by “countries of proliferation concern” for use in weapons programs. “Recent reports indicate,” the findings read, “that China has purchased hundreds of supercomputers for us in its weapons programs and that the United States is unsure of the location of those supercomputers or the purposes for which they are being used.”
To this day, publicly available evidence that the PLAAF designed the JH-7 using a U.S. supercomputer remains circumstantial. (Though one academic paper claims that Chinese engineers later bragged about having done just that.) But in some ways, that uncertainty is exactly the point. If a country acquires the design for a missile silo, then it can build a missile silo. But if it acquires the computing power to calculate how efficiently weapons fly through the air, and shoot through the water, there is no telling what it will create.
Leave a Reply