The U.S. Government is considering new export controls on semiconductors made by U.S. companies, especially chips from NVIDIA used in training AI models and running data centers. I was recently interviewed on CNBC—not once, but twice—about the potential damaging impacts of these restrictions, if indeed the Biden administration goes through with them.
This issue is complex and deserves a more complete discussion rather than some knee-jerk reaction. The focus has inevitably turned to NVIDIA, a company that is absolutely killing it on every front right now. There’s also been more heat than light created by China’s recent moves to pressure U.S.–based Micron Technology and restrict the export of germanium and gallium, vital elements for microchips, EVs, military gear and more.
Technology moves so fast that we sometimes have short memories. But it’s important to put these developments into historical context. These tensions go back at least twenty years. I clearly remember when China essentially banned IBM, Cisco and even Microsoft from their critical infrastructure. And on the consumer side, they still block or censor Google, Twitter and Meta (Facebook, Instagram, Threads). To be fair, the U.S. has also enacted bans on telecom equipment from Huawei and ZTE, citing national security reasons. And the TikTok circus is ongoing.
I hope the current flare-up over semiconductors will pass, but we need to be sure we understand its implications for AI, where China’s ambitions are sky-high.
China will keep developing AI capabilities with or without U.S. chips
Last October, the U.S. government imposed some limits on exports of the hottest AI chips. These limits particularly affected NVIDIA, whose A100 and H100 processors dominate the field for training AI models. Led by Nvidia but also including Intel, AMD, Qualcomm and others, U.S. chip makers produce the best, most energy-efficient and most cost-efficient chips for AI training and data centers, just as they do in many other areas.
The October restrictions have led Nvidia to release A800 and H800 models, which throttle down performance in line with the regulations. But restricting chips further won’t stop China from developing AI—and would in fact be counterproductive.
China’s chip industry isn’t as advanced as its U.S. counterpart, but Chinese chip makers like Huawei and Biren are capable of producing good designs that get the job done at lower performance levels. Of course, the Chinese companies working in AI would rather use Nvidia’s products, which are the best right now in training in terms of silicon, software and reliability, but if you aggregate enough of the lower-performing chips, you can accomplish the same AI tasks eventually. This approach would impose marginally higher costs for new data centers but hardly cripple them. Also, tighter restrictions wouldn’t impact the large installed base of Nvidia chips already in Chinese data centers, or China’s access to domestic and foreign cloud-based compute options, or China’s homegrown supercomputer capacity.
To think of it another way, all the foundational work to train models for the generative AI that’s been wowing us this year was done before the H100 even existed. Sure, Nvidia’s chips really are the state of the art for AI, but it’s not like they’re the only chips that can help a company working in that area.
In short, putting export restrictions on Nvidia chips assuming it will choke off the Chinese AI industry doesn’t make sense. It would certainly inconvenience Chinese AI companies, but it would hardly prevent them from developing the AI they want to pursue.
Beware the unintended consequence of promoting the Chinese chip industry
The bigger risk here—one pointed out by Nvidia CFO Colette Kress at a conference a couple of weeks ago—is that draconian export controls might cost the U.S. some of its leadership in the chip industry over the long run. The October restrictions were balanced, allowing the continued export of U.S. chips from Nvidia (and others) but at a moderate performance threshold. Following that course would keep Chinese AI companies dependent on U.S. technology, and it would keep AI data centers in the U.S. ahead of their Chinese counterparts.
By contrast, going down the slippery slope of tightening performance thresholds would ultimately result in spurring the growth of China’s domestic chip industry. The Chinese chip makers don’t have a big footprint yet, but there are a lot of them that already have a solid level of competence. When I was an AMD executive 20 years ago, the Chinese chip companies we dealt with weren’t great at design—but it’s not like that anymore. If Chinese AI companies need to turn to them for chips previously sourced from the U.S., it only makes sense that these suppliers will ramp up their design and production capabilities to meet the demand. While these companies are not yet on the level of Nvidia or Intel or Qualcomm, we don’t need to give them any more reasons to accelerate their evolution.
Limitations in U.S. regulatory thinking
One of the underpinnings of U.S. trade policy is the “high wall around a small garden” approach. This means that when you’re working in a technology niche where the U.S. has leadership, you put a high wall around that technology so other countries can’t easily catch up. It’s an approach that makes perfect sense if you’re talking about, say, the high-end avionics that control stealth fighters. You really don’t want tech like that to leak out. But it doesn’t make sense in commercial AI.
In the AI garden, the seeds are the AI software frameworks—which China already has access to. The plants in the garden are the AI models in use, which again are already available to Chinese AI companies. Nvidia provides the best shovels and pruning shears to tend the garden, but not the only means to tend it. So it doesn’t make sense to try to build a high wall around it.
I believe this is another example where the U.S. Government wants to apply a regulatory philosophy that doesn’t fit the real world. We’ve seen this recently with proposed know-your-customer (KYC) regulations for IaaS providers that won’t actually do anything to stop bad actors. (I went into great detail about that in a recent article.) We’ve seen it in the FTC’s skepticism toward sensible acquisition deals like Adobe-Figma, Broadcom-VMware or Microsoft-Activision (in which a federal judge just rightly ruled against the FTC). I get it that we need responsible regulatory oversight to make sure that companies don’t unfairly suppress competition, don’t gouge consumers and so on. But we also need regulatory oversight that makes logical sense.
In this case, it does not follow logically that Nvidia’s success in AI chips makes it absolutely indispensable for pursuing AI at all. And to over-regulate these chips creates the risk that the U.S. could fumble away its technology leadership. Would you rather have Chinese AI customers continue to fuel Nvidia’s growth and success? Or would you rather they spend their yuan to fuel the growth and success of Chinese suppliers?
I think the proposed tighter export controls speak to generalized anxiety about AI mixed with generalized anxiety about China as the major rival for the U.S. These are weighty topics that deserve well-reasoned consideration. But restricting Nvidia from China isn’t going to make the U.S. safer or more competitive in the long run.