Thanks, that seems to sum it up.

Going up to twenty year old level, the next question is why would investors be so confident that "new free/libre model anyone can run on their GPU compute" would on balance mean a reduction in demand for GPU compute versus "training new model doesn't need as much compute as we thought"? Perhaps they are right, but perhaps the opposite could be true as well. For example, new free/libre model expands market for GPU compute to wider audience by lowering entry cost.