Right, this is the classic VC playbook and people should really know better. They use capital to undercut competitors on cost to gain market share, then slowly crank up the costs until they are (hopefully) profitable. It makes 0 sense long-term to spend tens of millions on training a frontier model and releasing it for free for someone else to host on their GPUs. It is naive to believe that when the capital starts to dry up and the VCs are looking for a profit that the models will continue to remain open-weight.

Chinese labs are a bit different since they are somewhat state-funded, so I'd expect them to shift to a model where Chinese models are hosted on Chinese infra.