Earlier this year, we argued that artificial intelligence would shape both economics and politics for decades. The current debate over controlling AI is proving that point. But the conversation has reached a stage where vague calls for restraint are no longer enough. The question now is concrete: if the world truly wanted to cap AI's capabilities, what exactly would we have to stop?
The answer, if we follow the logic to its end, is almost everything. A lasting global ceiling would require control not just over models and their use, but over the entire infrastructure that produces them: processors, accelerators, memory, interconnects, and the software that ties them together. It would even extend to the people allowed to work on these systems. Restrictions limited to models or today's deployment arrangements would leave the rest of the system free to advance, and the ceiling would quietly erode.
The Proposals on the Table
Dario Amodei, CEO of Anthropic, has called for slowing the pace of AI capability improvement. His proposal includes external evaluators working inside labs, safety requirements tied to capability levels, and international coordination. He also supports chip export controls and acknowledges how difficult comprehensive global restraint would be. His goal is to buy time for safety research to catch up.
Others go further. In 2023, Eliezer Yudkowsky proposed an indefinite worldwide moratorium on large AI training runs. Between these positions lie a spectrum of ideas: development restrictions, deployment conditions, hardware controls, and ongoing supervision of specific applications or users. The range is wide, and agreement that AI needs control does not settle which of these ambitions people actually mean.
Some of the loudest voices now come from late converts—those who once dismissed AI as hype. They are using the control debate to reinforce their negative view of innovation industries. This article is not about them. The issues at stake are far more serious than a stock market argument.
The Hard Part: From Agreement to Enforcement
It is encouraging that capable people are working on these questions in detail. A measure intended to buy time should be judged on whether it actually does. But the broader challenge is containing AI's expanding capabilities worldwide. Here is a simple test: suppose everyone agrees, everyone complies, and models stop improving. What can still happen?
Moving from agreement to legal language, from language to measurement, and from measurement to enforcement, the same difficulty recurs. Capability can grow through changes outside the specific activity being restricted. Researchers recognize these effects, but the implications for a global agreement are rarely followed through.
Consider the supply chain. Chip export controls, for example, are a favored tool. But chips are not the only bottleneck. Memory, interconnects, and the software that orchestrates them all matter. Restricting one element while leaving others free may simply shift the constraint. The same applies to people: if certain researchers are barred from working on frontier models, others with similar skills may emerge elsewhere.
The political obstacles are just as formidable. What one government wants slowed may be another's chance to catch up. Restricting another country's access signals distrust and gives those excluded a stronger incentive to develop their own alternatives. Within countries, a firm stance by one party gives its opponents an opening to champion a different course. Companies will lobby for rules that constrain competitors more than themselves. The debate is likely to lose its primary purpose in a world of competitive nations, corporations, and political landscapes.
Politicians answer to their own electorates. Companies answer to customers, owners, and employees. The benefits surrendered may be immediate and local; the dangers avoided may be uncertain and shared. The same person can sincerely fear uncontrolled AI and sincerely fear what falling behind would mean for their country.
Extend the proposed controls to memory, processors, and communications, and more industries enter the negotiation. Equipment that makes AI agents more effective may also improve medical imaging or reduce electricity consumption. The scope of the bargain grows with the scope of the promised control.
These are formidable problems. For the sake of argument, let us assume they are solved. Now write the rule. "Do not develop AI beyond the agreed frontier." Eventually, someone must measure compliance. How do you measure a capability that can emerge from a combination of hardware, software, and data that no single actor fully controls? The measurement problem alone could sink any agreement.
The debate over AI control is not going away. But it is time to move from slogans to specifics. The question is not whether we want control, but what we are willing to stop—and who will enforce it. For the Indo-Pacific, where chip supply chains and AI ambitions are deeply intertwined, the answer will shape the region's economic and strategic future. As US and Japan race to secure chip minerals and China develops its own counter-strategies, the practical meaning of AI control is being tested in real time.
