One could consider it fruitful to muse upon the question of - what can we learn from the nuclear arms race, in order to devise an effective framework of slowing down the development of AI capabilities at the frontier? There are several minimum prerequisites that the approach for slowing AI must share with nuclear -

1. all parties involved need to roughly agree on a protocol or mechanism which has some useful effect towards the goal of stopping / slowing down capabilities, such as export controls on technological and critical supply chains.
2. implementation of the methods adopted must be verifiable or made verifiable by other parties in the overall coalition
3. it should be hard, or made hard, to deceive another party of the effective implementation of the above measures

In practice, this is implemented via export controls on technology and materials, as well as mutual disarmament, and prevention of states without nuclear weapons from acquiring them. Thus a bottleneck or a chokepoint is created that obeys the minimum prerequisites listed above.

What if an existing bottleneck required for the buildout of large-scale AI training clusters is leveraged to effect something similar to slowing down AI? Say, all data centres that draw > X megawatts of power are monitored by a coalition of entities and the information regarding the training of AIs on such clusters is made public? It would be quite effective, as it would make the capex for hiding AI training prohibitive in comparison to doing it out in the open in accordance with the terms of the proposed treaty, as one would, for example, have to resort to the creation of a dedicated underground training cluster, or invent a technology that would allow you to position racks in a data centre much further away from each other than they would be in a normal data centre, to evade monitoring. One example of such a technology could be photonics-based interconnects that would do-away with the necessity of making a cluster fit in a small constrained area, at the cost of lower cost-per-unit reduction via economy of scale, as well as the increased upfront cost as well as risks associated compared to repurposing existing tech, thus effectively incentivising actors to cooperate and stay within the terms of the treaty i.e. build AI training clusters as per regulations, out in the open, where measures taken to slow progression of frontier capabilities can be more effectively implemented.