$9,000 per card was what I heard quoted in the office several months back and we had less than a dozen for a really small scale in-house operation. I can only imagine how much hardware something at scale requires.
Depends on how tolerant of half assing things you are. I grabbed a P40 which is enough for a hobbyist for 300$. It does a passible job if slightly slow. But it's plenty capable for a single user. Of course if you want to do anything cutting edge, you'll be spending tens of thousands. However a significant part of that is Nvidia being stingy with VRAM. It wouldn't be particularly difficult to make high VRAM cards that would be far more cost effective for AI, but it's more profitable for Nvidia to do it this way.
$9,000 per card was what I heard quoted in the office several months back and we had less than a dozen for a really small scale in-house operation. I can only imagine how much hardware something at scale requires.
Depends on how tolerant of half assing things you are. I grabbed a P40 which is enough for a hobbyist for 300$. It does a passible job if slightly slow. But it's plenty capable for a single user. Of course if you want to do anything cutting edge, you'll be spending tens of thousands. However a significant part of that is Nvidia being stingy with VRAM. It wouldn't be particularly difficult to make high VRAM cards that would be far more cost effective for AI, but it's more profitable for Nvidia to do it this way.