Muddled AI Chip Curbs Will Backfire on Alliances

The rules read like a foreign-policy brief, but the impact will land in cubicles and on factory floors. The Council on Foreign Relations piece calling the new AI chip export policy to China “strategically incoherent and unenforceable” isn’t just tossing around think-tank jargon. Listen to the language: that phrasing is doing double duty as legal critique and political side-eye. The deeper problem is that the policy performs a status move—showing toughness—while doing little to reshape the technological choices that actually matter.

Supporters of the policy have an easy talking point: better to act now, even imperfectly, to raise costs and slow an adversary’s trajectory in AI. That line works on cable news and in donor briefings. But the CFR critique points to the less telegenic reality: imperfect controls aren’t harmless; they’re distortions. They create incentives for concealment, accelerate investment outside U.S. reach, and load allies with compliance headaches that quietly drain political goodwill. That’s a management tell—when leadership reaches for symbolic action first and hopes the operational details magically line up later.

On paper, the intent is crisp enough: slow access to high-end chips that could power advanced AI systems. In practice, the actual controls look more like a sieve. The language around tiers, thresholds, and carve-outs reads like the minutes of a bruising interagency meeting, a bureaucratic compromise trying to split the difference between economic pain and security posturing. The spreadsheet is clean; the world isn’t.

Inside companies, this kind of policy doesn’t arrive as “grand strategy.” It arrives as memos. Contract reviews. Procurement freezes. Engineers told they can’t ship something they finished six months ago because a rule changed while they slept. People feel these changes before they can name them. The CFR piece hints at this: firms are being told, implicitly, that Washington wants two contradictory things at once—to constrain hardware flows and keep supply chains running smoothly. Compliance teams then have to invent a story where both are true.

Compliance itself is not just a neat list of controlled items. It’s emails that go unanswered because everyone’s scared of putting the wrong thing in writing. It’s delayed shipments, strained vendor relationships, and lawyers quietly monetizing ambiguity. The article is right about the enforcement problem: global supply chains are elastic, manufacturing footprints shift, and design tools can be decoupled from the fabs that make physical chips. That’s not a technical footnote; it’s a behavioral forecast. Firms will route around friction because that is what firms do.

So you get the predictable workarounds. Components get reclassified. Orders are split into bite-sized pieces. Third-party brokers suddenly look more attractive. Some production inches toward jurisdictions that aren’t fully lined up with the policy. Allies are dragged into the same maze—asked to mirror controls, they weigh their economic exposure against political solidarity and wonder how long the solidarity will actually be rewarded. The practical result is leakage: fewer traceable exports, more creative intermediaries.

A policy that is unenforceable can be worse than no policy if it simply offloads risk onto industry and partners while leaving the underlying security problem roughly where it started. The CFR critique stops just short of saying this outright, but it’s embedded in the argument: you do not “win” by making private firms the front line in a geopolitical blockade unless you also change incentives and provide real support. Right now, the burden flows down; the benefits are abstract and upstream.

Where I diverge a bit from the CFR tone is on that word “incoherent.” The label is accurate, but it can make coherence sound like a philosophical virtue instead of what it really is here: a political outcome that has to be built on purpose. Coherence comes from aligning tools to the actual risk. If the core fear is dual-use acceleration—AI applications that materially reshape military decision-making—then the policy should be obsessed with end-use pathways and know-your-customer regimes, not just raw silicon. If the deeper concern is long-term talent and software ecosystems, then choking off chips is like regulating gasoline to control where people choose to drive.

The CFR article sketches this misalignment; it could push harder. Policymakers need to say out loud what they think the definitive risk is and then design controls around observable, provable transfers linked to that risk. That means fewer abstract categories and more focus on traceable use cases. It also means admitting that some “losses” in commercial access might be an acceptable price for controls that actually work, instead of spreading the pain thinly and pretending it’s precision.

There is a quieter cost here that doesn’t show up in official impact analyses: internal trust. Every time a company reorganizes around a new, hazy rule set, people update their private assessment of how predictable their work really is. Engineers and operations managers learn that strategic priorities can flip on the basis of policies that feel like they were written for a different world than the one they inhabit. The spreadsheet misses the human part, and that human part pushes back in the only way it can—by nudging decisions away from scrutiny, toward places and structures where Washington’s reach is weaker.

If the goal is to reduce genuine risk, the next draft of this policy needs three things: tighter alignment between restrictions and concrete end uses, serious investment in allied enforcement capacity so private firms aren’t the de facto border patrol, and exemptions or licenses that are explicitly tied to transparency and monitoring instead of buried in fine print.

Give this policy a few years as written and the most advanced knowledge about how to dodge it will sit exactly where Washington least wants it: in the workflows of the companies and countries it was meant to constrain.