I think you missed the forest for the trees. I did incorrectly cite GPT-4 as I was going from memory and that's suspect sometimes. I also didn't elaborate and maybe I should have given the snarky comments I'm seeing.
Actually the amount of power available matters because you are consuming energy in time. If I have a 1MW plant and a battery, I can generate 1GWh in about 3 weeks. This seems a little silly though. A Hyperscale DC campus is ~150MW to 200MW. If you plot the larger ones, they are almost all near power stations with >1GW capacity (not all).
The industry trend is towards building 1GW datacenters. Last I checked these would consume ~8.7TWh (assuming PUE of 1). However, the 8.7TWh while relevant is meaningless unless the power to the DC can be 1GW. Since the plant itself has to generate more than 1GW (the plant has a cap ratio so more than this, plus other demand, etc..) for such a site, then it follows that there are limited number of sites in the US (this is public info see EIA.gov or Wikipedia).
Grok3 is already at 140MW (100 days of training ==> 336GWh) at ~10^26 FLOP. Model FLOP is increasing at ~5x per year so by 2030, we are expecting to be ~10^28 and that would take ~10GW (24PWh). If I am optimistic and say that the efficiency can improve by 1.3x per year, then we still need a very large generating station to meet the demand or we need to distribute among many smaller sites.
You can push the numbers around however you like but the conclusion is the same, the timing may be different.
There's a reason why all the hyperscalers are investing in nuclear, large generating capacity and the highest cap factor of any form of energy.
My 2nd comment still stands, and you left unaddressed (remember the forest?)..
Actually the amount of power available matters because you are consuming energy in time. If I have a 1MW plant and a battery, I can generate 1GWh in about 3 weeks. This seems a little silly though. A Hyperscale DC campus is ~150MW to 200MW. If you plot the larger ones, they are almost all near power stations with >1GW capacity (not all).
The industry trend is towards building 1GW datacenters. Last I checked these would consume ~8.7TWh (assuming PUE of 1). However, the 8.7TWh while relevant is meaningless unless the power to the DC can be 1GW. Since the plant itself has to generate more than 1GW (the plant has a cap ratio so more than this, plus other demand, etc..) for such a site, then it follows that there are limited number of sites in the US (this is public info see EIA.gov or Wikipedia).
Grok3 is already at 140MW (100 days of training ==> 336GWh) at ~10^26 FLOP. Model FLOP is increasing at ~5x per year so by 2030, we are expecting to be ~10^28 and that would take ~10GW (24PWh). If I am optimistic and say that the efficiency can improve by 1.3x per year, then we still need a very large generating station to meet the demand or we need to distribute among many smaller sites.
You can push the numbers around however you like but the conclusion is the same, the timing may be different.
There's a reason why all the hyperscalers are investing in nuclear, large generating capacity and the highest cap factor of any form of energy.
My 2nd comment still stands, and you left unaddressed (remember the forest?)..