H200: A Perspective on Export Control Failures (Again)
Republishing in light of the H200 debate, which involves magical thinking, wishful thinking, and hubris. The lesson is capture the market to advance US interests
The Export Control Reform Act of 2018 (ECRA) states that national security “requires that the United States maintain its leadership in the science, technology, engineering, and manufacturing sectors, including foundational technology.” Unfortunately, the practice of export control since the 1980s provides a number of striking examples where a mechanical and heavy handed application of export controls has produced the opposite effect, usually as a result of underestimating both the damage to American industry from ill-advised regulation and the ability of other nations to develop substitute or replacement goods to circumvent US controls. AI and GPUs are no different. It is worth listing salient examples of export control failures to set the context for a reconsideration of the AI diffusion regulations.
In the 1980s, machine tools exported to the Soviet Union let it make quieter propellers for nuclear submarines. The United States tightened export controls on machine tools, with the result that American machine tool companies were driven out of business, leaving the United States dependent on foreign suppliers, including China (now a global leading supplier).
Until 1997, the U.S. allowed China to launch America-made satellites. The Chinese used this to improve their space launch capabilities but the US gained industrial and intelligence advantages (China allowed dozens of US experts to inspect and monitor launches and rockets. Concerned over Chinese access, the US blocked all exports of American satellites and their components to China, including European-made satellites with US components, and the US imposed onerous licensing conditions on European satellite makers. In response, the EU created its own component makers (using subsidies from the European Union) to replace U.S. suppliers and China simply moved on withe the Long March, now one of the world’s move successful launch vehicles. U.S. space companies lost market share and smaller U.S. component suppliers exited the satellite business.
The U.S. controlled and restricted encryption exports until 1999. The result was that despite an initial lead, U.S. software makers were being displaced by foreign software companies until the export restrictions were removed. After that, US firms outcompeted foreign suppliers, leading to a dominant position in internet software and services.
The United States tried repeatedly to restrict Chinese access to high performance computer chips (CPUs). The famous 1999 commercial of a “dangerous” Mac computer being surrounded by tanks dates from this period. The U.S. has hoped for decades to keep China two generations behind, but without success. Chip performance increased rapidly, making control thresholds obsolete, and older chips could be substituted for the most advanced ones to create militarily useful applications. The same is true today. One result was to inspire China to invest more to build its own chip industry and it now ranks 3rd overall in chip production.
The US liberalized rules for the export of commercial imagery satellites but then took an overly cautious approach in approving any sales, thus failing to capture the global market. The advent of “small-sats” allowed foreign producers to build competing satellites that provided adequate (if not “exquisite”) imagery. The US found it preferable to sell imagery as a service, which provided more control over access, rather than imagery satellites themselves. US export rules and risk averse implementation helped speed the growth of foreign capabilities and cost the US markets it could have dominated.
The US denied Turkish requests to buy American UAVs. In response, Turkey created its own drone industry and drones such as Bayraktar, that are cheaper, faster to build, and for many observers, better than US-made drones. The result of denying sales of US drones to global markets is that the US created powerful competitors in Turkey, Israel, and China (which has the largest share of the global drone market).
These were important technologies with strategic significance where the US sacrificed opportunities for global lead as a result of ill-conceived regulation and without affecting opponent capabilities. In the cases of both satellites and encryption, foreign manufactures even advertised their products as “ITAR-free,” meaning not subject to US export controls, successfully attracting foreign customers. Similar outcomes are likely from the AI diffusion rule. An overly risk-adverse approach not attuned to the global dispersion of technology and the increasing difficulty of denying access to it only damages the US advantage and build Chinese competitors.
On of the lessons from these experiences is that the US’s has much less ability now, after thirty years of global technology diffusion, to block access to significant dual-use technologies in ways that deny meaningful improvements to opponents’ military capabilities. There have been a few successes, such as designs for military jet engine components, but these have focused on the ‘know how” needed to build them as much as end products or services. Export controls can make opponent programs more expensive and their products less capable, but not to a degree that will provide advantage to the US or compensate for the damage these rules can do.
It is worth noting that the 1990s efforts to block access to high performance computers and chips were based on a performance threshold called MTOPS. This threshold experienced many difficulties. First, the controls could be circumvented by aggregating chips with lower performance thresholds not controlled for export. Second, the steady rate of improvement in chip performance meant that the controls had to be constantly updated to match foreign availability. US export regulations further incentivized China’s efforts to build its own chip industry. The precedent, like the other cases, is that the likely outcome of chip controls is a stronger Chinese AI and chip industry.
This is a pattern that stretches over decades and it is troubling that it must be told again: