With the 5090 you need to buy the rest of the computer though, and the Dgx spark will run 1/4th as slow but use 1/5th the electricity. And the spark would be able to run things the 5090 just couldn’t, like the Qwen3.5 122b. Which is all just to say that for llm workflows there is no easy answer. And if you media generation it gets even more complicated.