Whatever else can be said about the second Trump administration, it is always teaching me about parts of the Constitution I had forgotten were even in there.
Case in point: Article I, Section 9, Clause 5 states that “No Tax or Duty shall be laid on Articles exported from any State.” This is known as the export clause, not to be confused with the import-export clause (Article I, Section 10, Clause 2). The Supreme Court has repeatedly held, most recently in 1996’s US v. IBM, that this clause bans Congress and the states from imposing taxes on goods exported from one state to another or from the US to foreign countries.
This story was first featured in the Future Perfect newsletter.
Sign up here to explore the big, complicated problems the world faces and the most efficient ways to solve them. Sent twice a week.
I found myself reading US v. IBM after President Donald Trump announced an innovative new deal with chipmakers Nvidia and AMD. They can now export certain previously restricted chips to China but have to pay a 15 percent tax to the federal government on the proceeds. Now, I’m not a lawyer, but several people who are lawyers, like former National Security Council official Peter Harrell, immediately interpreted this as a clearly unconstitutional export tax (and as illegal under the 2018 Export Control Reform Act, to boot).
At this point, there’s something kind of sad and impotent about complaining that something Trump is doing is illegal and unconstitutional. It feels like yelling at the refs that the Harlem Globetrotters aren’t playing fair; of course they aren’t, no one cares. The refs are unlikely to step in here, either. The parties with the standing to sue and block the export taxes are Nvidia and AMD, and they’ve already agreed to go along with it.
Maybe the best we can do is understand why this happened and what it means for the future of AI.
A brief history of the 2025 chip war
While AMD is included in the deal, for all practical purposes, the AI chips in question are being made by Nvidia — and the main one in question is the H20.
As I explained last month, the H20 is entirely the product of US export controls meant to limit export of excessively powerful chips to China. Nvidia took its flagship H100 chip, widely used for AI training, and dialed its processing power (as measured in floating point operations per second) way down, thus satisfying rules restricting advanced chips that the Biden administration put in place and Trump has maintained.
At the same time, it dialed up the memory bandwidth (or the rate at which data moves between the chip and system memory) past even H100’s levels. That makes the H20 better than the H100 at answering queries to AI models in action, even if it’s worse at training those models to start with.
Critics saw this all as an attempt to obey the letter of the export controls while violating their spirit. It still meant Nvidia was exporting very useful, powerful chips to Chinese AI firms, which could use those to catch up with or leap ahead of US firms — precisely what the Biden administration was trying to prevent. In April, the Trump administration seemed to agree when it sent Nvidia a letter informing it that it would not receive export licenses for shipping H20s to China.
Then, in July, reportedly after both some bargaining with China over rare earth metals and a personal entreaty from Nvidia founder and CEO Jensen Huang, Trump flip-flopped; the chips could go to China after all. The only thing new this month is that he wants to get a cut of the proceeds.
That, of course, is an important new element, not least because it seems bad that the president is asserting the power to unilaterally impose new taxes without Congress. (At least with tariffs, Trump has some laws Congress passed he can cite theoretically granting the authority.) But the big question about H20s remains the same: Does this help Chinese companies like DeepSeek catch up with US companies like OpenAI? And how bad is that, if it happens?
Talking through the pros and cons of H20s
The concerns here are such that maybe the best way to understand them is to imagine a debate between a pro-export and anti-export advocate. I’m taking some poetic license here, in part because people in the sector are often averse to plainly saying what they mean on the record. But I think it’s a fair reflection of the debate as I’ve heard it.
Anti-Export Guy: Trump says he wants the US to have “global dominance” in AI, and here he is, just letting China have very powerful chips. This obviously hurts the US’s edge.
Pro-Export Guy: Does it? Again, the H20 is powerful, but it’s no H100. In any case, Chinese firms are totally allowed to rent out advanced AI chips in US-based cloud servers. DeepSeek could even rent time on an H100 that way. So, why are we freaking out about exporting a weaker chip?
Anti-Export: You act like the cloud option is a loophole — it’s a feature! That way, they’re dependent on US servers and companies. If Chinese AI firms ever start making dangerous systems, the US can shut off their access, and they’ll be out of luck.
Pro-Export: Again, will they be out of luck? There’s a third option after Nvidia exports and US servers. Huawei is making its own AI-optimized chips. Chinese firms don’t want to depend on foreign servers forever, and if we deny them Nvidia chips, they’ll run right over to Huawei chips.
Anti-Export: Saying you don’t need Nvidia chips when you have Huawei chips is like if you told someone 20 years ago that they don’t need an iPod because they have a Zune. Yes, Huawei chips exist, but they’re so much worse. They’re lower bandwidth than H20s, Huawei’s software libraries are full of bugs, and the chips sometimes dangerously overheat.
Pro-Export: You’re exaggerating. By some metrics, Huawei’s latest systems (not just the chips, but the surrounding servers) outperform Nvidia’s top-end model — even though that model uses B200s that are faster than H100s and lightyears faster than you’d ever be allowed to export to China. Yes, programmers will have to learn Huawei’s libraries, and transitioning from Nvidia’s will take time, but it’s doable. Google, Anthropic, and OpenAI have all recently moved away from Nvidia chips toward things like Google’s own TPUs or Amazon’s Trainiums. That took effort, but they did it.
Anti-Export: Sure, but those companies still use Nvidia’s too. OpenAI wants 100,000 chips in one Norwegian facility alone. And while US companies may be trying out the competition, Chinese companies still vastly prefer Nvidia to Huawei. DeepSeek reportedly had to delay its latest model because it tried to train it on Huawei chips but couldn’t. Even if Huawei chips were popular, Huawei lacks the production capacity to meet demand. It relies on smuggled components to make its top-end chips and can make at most 200,000 this year, compared to the roughly 10 million Nvidia chips shipped annually. There’s no substitute for Nvidias.
What we’re fighting for
I suppose we’ll see, in the next few months and in the rollout of new chips from competitors like Huawei, who got the better of that argument. China is reportedly discouraging firms from using Nvidia chips in the wake of the export tax deal, largely to encourage them to use domestic chips like Huawei, though they are clearly not banning the firms from using Nvidias if they prove necessary. It’s also investigating whether the US is including spyware in them.
The bigger question this debate raises for me, and one I certainly can’t answer adequately here, is: To what degree is “beating China” on AI important for making the future of AI go well?
The answer for most US policymakers, and most people I know in the AI safety world, has been “very.” The Financial Times reports that some Trump officials are considering resigning in protest over allowing China to get H20s. As Leopold Aschenbrenner, the AI analyst turned hedge funder, put it bluntly in his influential 2024 essay “Situational Awareness”: “Superintelligence will give those who wield it the power to crush opposition, dissent, and lock in their grand plan for humanity.” If China “wins,” then, the result for humanity is permanent authoritarian repression.
No doubt, the Beijing regime is brutal, and I have no faith that they will use AI wisely. I’m very confident they’ll wield it to oppress Chinese citizens. But it feels as though “staying ahead of China” has become the sine qua non of US AI policy.
I worry less that this focus on China is directionally wrong and more that it is exaggerated. The bigger danger is that no one can control these systems, rather than that China can, and that the focus on staying ahead of China will cause the US to speed deployment of automated weapons systems that could prove deeply destabilizing and dangerous.
As with most aspects of AI, I feel like there’s a small island of things we’re all sure of and a vast ocean of unknowns. I think offering China H20s probably hurts AI safety a bit. I think.
Source: vox.com
