Comment by LaurensBER
1 hour ago
> 1GW of power datacenters, more efficient GPUs, more efficient training, Xiaomi/Deepseek disk based caching + other efficiency gains.
I haven't done the math but if I see how many tokens I can burn per month for 100 USD it's pretty mind boggling to imagine how much capacity is being added by this build out.
This is not the place to discuss it but even in my wildest predictions I find it hard to imagine that we're really going to need all this capacity anytime soon (although I guess anything spare can go into training).
It’s not ‘spare goes to training’. These data centres are training first. Inference is really rather efficient, so not why datacenters are so big.