From https://metaculus.com/questions/16639/non-democratic-brontier-ai-lab-before-2026/

Currently the largest model currently known is GPT-4 and while information about the compute needed to train the model is not public, GPT-4 is estimated by Epoch to have been trained with 2e25 FLOP.

We define a "Frontier AI Lab" as one who has trained models within one order of magnitude of GPT-4 (>2e24 FLOP). As of 2023-04-28, these are OpenAI, Anthropic, Microsoft, and Google Deepmind, [source]. Labs that would join this with a ~10x growth in training compute include Meta, Yandex, Amazon, and Hugging Face.

The fact that all frontier AI labs are currently in democracies which may mean it is easier to coordinate these labs to produce AI more safely either via public pressure, lab agreement, and/or regulation than if we had to coordinate frontier AI labs in non-democracies. Will this convenient status quo hold?