They should consider buying or renting their own GPUs to run one of the cheaper open models instead of paying the labs.

They also should have a smart model supervise the cheaper and less capable ones.