Bas_

The jump from 5*10^26 to 10^27 is very small (requiring only a doubling of compute used) given the massive buildout of data centers by AI companies they clearly expect to need a lot more compute for their next models, and will have the capability to do so. I therefore consider it extremely likely that a US entity will do this.

Non US entities are trickier. the highest current models from outside the US are in the low 10^25 range, from Chinese companies, arab companies (UAE / Saudi, with chinese help) and Mistral (french/european). Given that the best models among those are already nearing the mid 10^25s it seems that they would need to increase their compute used by about 20X ( ~four doublings) over the next ~1.5 years. This is roughly equal to the increase in computing power for Chinese models over the last 2 years (from 5*10^23 to ~10^25) so it might be possible for one of these organizations to achieve this. It would require very fast growth and sufficient financing, at a time when the non US companies seem more focused on adoption than creating top of the line models. I therefore give it a good but not extremely high chance of happening. It will be possible to forecast this more accurately in early 2026 when we know if the non US are scaling at the rate that would be required to meet this goal or if they are falling behind the curve. 

US export restrictions are unlikely to make a large impact on this, since they can be evaded with transshipment through third countries, and substituted with Chinese made semiconductors (which have become advanced enough to offer competitive performance (though not at competitive prices))

The time window is too short for it to realistically happen. None of the three are on the brink of default

There is too little time left, the chance that this will happen within the next 3 days is substanially less than 1