I thought the discussion was about "frontier labs" in the cloud versus open at home.

Others running open models in the cloud is a nice third alternative, but not a solution for those that need frontier models, which I believe continue to need training as well as inference. These are currently heavily subsidized and hardware constrained several years out.