On a more serious note, here's what NV probably should have done:

- Develop an SDK that trained a model approximating HQ render from LQ render
- Worked with a couple studios to take non real time path traced frames from the same art (+- extra detail maps/materials) as data set
- Target current PT quality at much lower costs, not dual 5090
- Show before-after with per-game models for a couple games as a preview
- Call it something else and not reuse DLSS brand