Even AI-specialized data centers won’t operate anywhere close to 100% capacity utilization. If they’re wholly dedicated to training they might get to 70-80%, but we’re not aware of any data centers planned solely for training. www.powerpolicy.net/p/the-puzzle...
It's the training I primarily meant, because I HAVE seen that proposed near the footprint where I work, but maybe they don't materialize. I think at the utility we tend to have two mindsets: 1. We do not trust that customers will do what they say 2. Customers should be able to run at rating
When the loads are varied and distributed it's easy to say they won't all max out in coincident manner, but these gigascale single loads are *spooky*
Seems very unlikely there will be gigawatt+ scale DCs committed solely to training, since generally the only way to monetize the GPUs is via inference… could be some limited exceptions though
Out of curiosity have you heard anything about training duty cycle? About the GPUs pulling max power for something like 2s, then off for 2, etc etc?
Here’s the perspective of @astridatkinson.bsky.social, who used to work on compute resource optimization at Google: www.linkedin.com/posts/tylerh...
I’m struggling to work out why folks are concentrating on “100% utilization”. It don’t matter as such because ultimates a DC consumes a bunch of GWh annually. 🤷🏾 And every DC operator I have worked with absolutely *wants* to run close to full utilization and never does..
..because the smart/larger ones have always got expansion phases underway every time they pass a trigger/milestone on utilization. 🤷🏾 60% utilization in phase 3 is 120% utilization in phase 1. 🤷🏾
The primary driver of mass US gas capacity expansion right now is the assumption that new loads are 100% inflexible, and therefore trigger the need for new firm capacity… www.linkedin.com/posts/tylerh...
This is very strange because it implies the generators aren't talking to the retailers who aren't talking to the DC operators? If they're *really* assuming DC loads are "inflexible" then clearly they have no idea what it is DCs do and how they work.. (Which seems impossible at this level).
By default, power system planners generally assume all new load is 100% inflexible for the purpose of system planning (i.e. for grid planning and reserve margin planning) open.substack.com/pub/powerpol...
This is a carefully written statement and I have no fault with it.. I’d note that actual operations though are guided by contracts written with energy intensive customers. I’ve seen some of the relevant parts of those contracts and there is quite a lot of “flexibility” in them.
The “flexibility” is contractually identified (along with penalties) and it would be a strange simplification for system planners to just ignore this data? 🤔🤷🏾 At the end of the day, DC operators aren’t really any different to other energy intensive industries. They “resell” energy. As compute.
I'll throw something out here that's worth considering too - DCs are actually *too large* to be made part of under frequency load shed schemes and the like because dumping one via the methods at the utility's disposal (open breakers) creates too large a step change on the system
Do DCs connect to the medium voltage distribution system like a large industrial load would?
Perhaps smaller DCs? Everything in the MW class very quickly starts needing pretty large transmission (with appropriate step down infrastructure) and as the OP alluded to.. the other things needed to deal with potentially large step changes in supply/loss.
Most of them are directly transmission connected these days
Can you describe exactly how that works and what it means in more detail? What voltage do they get the connection at? What kind of poles and wires?
I don’t exactly know what this means. Yes, they are (for large commercial facilities) in the ones I’ve seen but the utility still builds and owns the infrastructure that steps the voltage down from HV to LV (33/11kV). Even when this is “embedded” into the DC structure. If you have examples?
That isn’t quite (completely) true. DCs are often much more flexible because of the nature of their construction (multiple data halls is just one example) and in some scenarios can drop some of their load.. (When given an economic reason to do so).
There is not to my knowledge at present a mechanism to force a DC to ramp down in controlled fashion at utility command quickly enough without creating that massive step, but I think there's going to need to be
Oh see here I am usually thinking of *transmission* capacity but he's on the gen side I see your point more