I just wanted to add one other thing on the hardware side.
These H200’s are power hogs, no doubt about it. But the next generation H300 or whatever it is, will be more efficient as the node process (or whatever its called) gets smaller and the hardware is optimized and can run things faster. I could still see NVIDIA coming out and charging more $/flop or whatever the comparison would be though even if it is more efficient power wise.
But that could mean that the electricity costs to run these models starts to drop if they truly are plateaued. We might not be following moores law on this anymore (I don’t actually know), but were not completely stagnant either.
So IF we are plateaued on this one aspect, then costs should start coming down in future years.
Edit: but they are locking in a lot of overhead costs at today’s prices which could ruin them.
I just wanted to add one other thing on the hardware side.
These H200’s are power hogs, no doubt about it. But the next generation H300 or whatever it is, will be more efficient as the node process (or whatever its called) gets smaller and the hardware is optimized and can run things faster. I could still see NVIDIA coming out and charging more $/flop or whatever the comparison would be though even if it is more efficient power wise.
But that could mean that the electricity costs to run these models starts to drop if they truly are plateaued. We might not be following moores law on this anymore (I don’t actually know), but were not completely stagnant either.
So IF we are plateaued on this one aspect, then costs should start coming down in future years.
Edit: but they are locking in a lot of overhead costs at today’s prices which could ruin them.