I will try to answer later with actual math, though a commenter above did a rough estimate. I would think it pays back relatively quickly. Moreover, this approach cannot really scale as old hardware and parts is scarce, and using it at a larger scale would make it scarcer (due to more normal failures). That's if you look specifically at one kind of system, but I believe the real value is in making code portable across a range of low-power systems, including recent and old ones.
You're also looking at it from the perspective of not needing new hardware, but I already have both my phone and my Steam Deck.
Incidentally, I bought a Steam Deck in part because it was cheaper than a GPU upgrade for my desktop PC, while being more powerful and efficient than my current setup. My PC has been collecting dust as I boot the same disk on my Steam Deck, at 1/10-30 the power enveloppe. As an added benefit, I don't have to turn this high-power system on while it's 35C in the room.
You're also looking at it from the perspective of not needing new hardware, but I already have both my phone and my Steam Deck.
Incidentally, I bought a Steam Deck in part because it was cheaper than a GPU upgrade for my desktop PC, while being more powerful and efficient than my current setup. My PC has been collecting dust as I boot the same disk on my Steam Deck, at 1/10-30 the power enveloppe. As an added benefit, I don't have to turn this high-power system on while it's 35C in the room.