As the markets reel and tumble from the first jabs in the opening rounds of the global fight on trade, it might seem safe to assume that any forward movement on chip export rules will be lost in the fray. Think again. The President has signaled a willingness to make deals. Those deals will likely include agreements to secure access to the advanced chipsets that AI relies on. They may even prove to be the deal clincher for both sides.
With advanced chipsets, the United States controls a good that the rest of the world (including China) wants and that we want much of the rest of the world (excluding China) to have. We want data centers around the world to be built with chips made by NVIDIA and AMD (and possibly maybe Intel) not Huawei or SMIC. To achieve that outcome, we need to make it easy for our friends and allies and hard for our adversaries to acquire these chips.
In my last post, I argued that the China Hawks in the Trump Administration are likely to win out over the AI bulls when it comes to chip export controls. While the Administration will no doubt revoke the Biden era “diffusion rules”, it’s highly likely that the new rules will respond to industry feedback but carry the same intent – to limit Chinese access to the most advanced chipsets.
Comments on the current interim final rule are due on May 15th. Industry partners would be wise to submit comments to shape it. But what should those new rules look like? Here is my stab at it.
First, the new regime needs to make sure that US companies have access to the chips they need before any chips head overseas. Once market demand is met in the United States, then, and only then, should chip exports be allowed.
Second, there is little reason to set chip allocations on a country-by-country basis. Under the new rules, there should be two binary questions to determine whether an export is allowed: 1) is the country trusted? and 2) is the company or joint venture that is buying, controlling, and installing the chips trusted?
If the answer to both those questions is yes, the export should be allowed. If the answer to either of those questions is no, it shouldn’t.
Third, if all parties are trusted, the controls necessary to keep the chips from being diverted to China can be minimal. Shipments can be tracked and installations verified remotely. Penalties that punish both the host country and private parties by blocking any future exports should result in sufficient compliance.
Fourth, the rule can drop the cybersecurity requirements to protect model weights in the current rule as they are both not implementable and unnecessary. I am a big proponent of fixing FedRAMP but there is little evidence that FedRAMP compliance keeps advanced persistent threat actors out of cloud systems. If the companies designing frontier AI models aren’t sufficiently incentivized to protect their intellectual property, there is no control regime in the world that can do so.
Finally, there is little point in building a regime to prevent China from acquiring chips if Chinese companies can simply rent the same chips at cloud service providers. I’ll tackle the thorny issue of know-your-customer rules in my next post…