> Running a data center on unreliable energy would be shockingly stupid.
For the right kind of workloads and at sufficient scale, I wonder if this is actually true. (It probably is, but it's fun to hypothesize.) I'm assuming the workloads are mostly AI-related.
AI training presumably isn't super time-sensitive, so could you just pause it while it's cloudy?
AI inference, at least for language models, presumably isn't particularly network-intensive nor latency-sensitive (it's just text). So if one region is currently cloudy... spin it down and transfer load to a different region, where it's sunny? It's kind of like the "wide area grid" concept without actually needing to run power lines.
Yes, I know that in reality the capex of building and equipping a whole DC means you'll want to run it 24/7, but it is fun to think about ways you could take advantage of low cost energy. Maybe in a world where hardware somehow got way cheaper but energy usage remained high we'd see strategies like this get used.
> So if one region is currently cloudy... spin it down and transfer load to a different region, where it's sunny? It's kind of like the "wide area grid" concept without actually needing to run power lines.
> Yes, I know that in reality the capex of building and equipping a whole DC means you'll want to run it 24/7, but it is fun to think about ways you could take advantage of low cost energy.
There's some balance between maximizing capex, business continuity planning, room for growth, and natural peak and trough throughout the day.
You probably don't really want all your DCs maxxed out at the daily peak. Then you have no spare capacity for when you've lost N DCs on you biggest day of the year. N might typically be one, but if you have many DCs, you probably want to plan for two or three down.
Anyway, so on a normal day, when all your DCs are running, you do likely have some flexibility on where tasks run/where traffic lands. It makes sense to move traffic where it costs less to serve, within some reasonable bounds of service degradation. Even if electricity prices are the same, you might move traffic where the ambient temperature is lower, as that would reduce energy used for cooling and with it the energy bill.
You might have some non-interactive, non-time sensitive background jobs that could fill up spare DC capacity... but maybe it's worth putting a dollar amount on those --- if it's sunny and windy and energy is cheap, go ahead ... when it's cloudy and still and energy is expensive, some jobs may need to be descheduled.
> AI training presumably isn't super time-sensitive, so could you just pause it while it's cloudy?
or pause it when "organic traffic" has a peak demand, and resume in off-peak hours, so that the nuclear powerplant can operate efficiently without too much change in its output.
For the right kind of workloads and at sufficient scale, I wonder if this is actually true. (It probably is, but it's fun to hypothesize.) I'm assuming the workloads are mostly AI-related.
AI training presumably isn't super time-sensitive, so could you just pause it while it's cloudy?
AI inference, at least for language models, presumably isn't particularly network-intensive nor latency-sensitive (it's just text). So if one region is currently cloudy... spin it down and transfer load to a different region, where it's sunny? It's kind of like the "wide area grid" concept without actually needing to run power lines.
Yes, I know that in reality the capex of building and equipping a whole DC means you'll want to run it 24/7, but it is fun to think about ways you could take advantage of low cost energy. Maybe in a world where hardware somehow got way cheaper but energy usage remained high we'd see strategies like this get used.