The video argues that rising compute costs will not simply make all models more expensive at the same rate. Instead, they will push the market toward the most powerful, efficient, and valuable applications, while further increasing industry concentration.

What I don’t understand is that the video was published on August 4, while DeepSeek V4 Flash 0731 had already been released on July 31. Why didn’t the video take open-source models into account at all?

Why would open models change this argument?

Which model is "best" changes too fast to be worth caring which particular model is the best, even on the timescale of making youtube videos.

If anything, my initial expectation would be that "compute is too expensive to leave money on the table" causes "therefore we shouldn't give away these somewhat expensive to train models at zero cost" in most cases.