If I'm doing the math right, you can get ~4000 queries per kWh. A quick Google search gave $0.04 per kWh when bought in bulk. So you can get around 100k queries per dollar of electricity.
That's.... a lot cheaper than I would have guessed. Obviously, the data centers cost quite a bit to build. But when you think of $20/mo for a typical subscription. That's not bad?
The fundamental flaw in AI energy/water doomerism has always been that energy costs money, water costs money, real estate costs money, but AI is being given away for free. There is obviously something wrong with the suggestion that AI is using all the energy and water.
It would be interesting to know how many queries they get per "base" model. I honestly have no idea what the scale is. In any case, that's the only way to know how to amortize training cost over inference.
That's.... a lot cheaper than I would have guessed. Obviously, the data centers cost quite a bit to build. But when you think of $20/mo for a typical subscription. That's not bad?
Did I do that right?