The Center for Security and Emerging Technology has published an essay asking a disarmingly simple question: should we track AI chips? The answer, as CSET frames it, is less obvious than the question.

The piece, titled "Should We Track AI Chips?", examines whether advanced accelerators — the hardware that trains and runs large AI models — can or should be individually monitored as they move through the global supply chain. It is a question with direct consequences for export controls, since a control regime that cannot see its subject is mostly theater.

Why does the question matter now?

Washington has built its China AI policy on the assumption that chips can be identified, counted, and stopped at borders. Successive export control packages have restricted the sale of high-end accelerators to Chinese buyers. A tracking regime would extend that logic from the point of sale to the point of use, and beyond.

CSET's contribution is to slow the conversation down. Tracking sounds like an administrative detail. In practice it raises hard questions about cost, technical feasibility, and who bears the burden — chipmakers, cloud providers, distributors, or governments.

What would tracking actually involve?

The essay does not pretend the problem is trivial. A workable regime would need to answer several questions:

• Which chips count as "AI chips," given that the same silicon increasingly serves training, inference, and ordinary datacenter work?
• Where in the supply chain does tracking begin — the fab, the package, the board, the rack?
• Who enforces compliance, and against whom, when a device changes hands three times in three jurisdictions?
• What does a tracking requirement do to the commercial position of the US firms that dominate accelerator design?

None of these questions has a settled answer. That is roughly the point of the essay.

The commercial dimension

For semiconductor firms, the stakes are not abstract. Every additional compliance layer touches the same companies that already carry the weight of license requirements, entity list screenings, and customer due diligence. A tracking mandate would add cost to products whose margins depend on volume and rapid iteration.

There is also a competitive irony worth noting. The tighter the monitoring regime, the stronger the incentive for customers to design around it — through domestic accelerators, through chiplets, through software workarounds. Controls tend to manufacture their own competition.

The verdict, provisionally

CSET does not settle the matter, and a research essay is not a regulation. What the piece does is put a marker down: before governments build a tracking apparatus, someone should establish that the apparatus can work, and that its benefits exceed its costs.

That is a modest ambition. It is also the correct one. The export control debate has spent several years arguing about which chips to restrict; the tracking question asks whether anyone can find the restricted chips after they ship.

Expect the answer to shape the next round of policy design, and with it, the compliance burden on every accelerator vendor selling across borders.