I've enjoyed your hyperdimensional posts and am glad you're back at it. Re Thin Air, you mention fission, fusion and enhanced geothermal but are silent about solar, despite its low and declining costs (incl with batteries) Why silent? Is that simply because the US isn't good at making them right now?
Solar has already made an impact here—now, grid demand peaks kick in around 5-6PM if not later, because behind-the-meter solar pushed that peak back from its original 1-2PM.
Batteries can help with peak management, especially because we can still buy them at scale. For this reason, batteries the most comment asset in the ISO-NE interconnection queue (my neighborhood). But solar isn’t a capacity asset anymore, so if you’re going to attack demand peaks for system reliability, you’re better off just installing the battery instead of worrying about solar.
This sounds great - but we know that data center electricity consumption (especially for pre training) is extremely in flexible - the idea that Stargate would stop training machine God, even for two days and risk ruiningtheir training run seems like a tall ask, no? I'm sure I'm missing something...
They don’t have to stop training for two days—just 2-4 hour intervals, maybe 3-5 times per month. And if they don’t want to stop, they can discharge a battery during that time.
It’s a serious ask, but not as big as you’d think at first. And if the alternative is no compute, a hyperscaler might take the deal.
As someone who spends much of his time with data centers, technology and climate, I must admit I wasn't aware of this possibility. I will also admit that at first I was reluctant to believe that this is feasible and practical, after reading the paper you cited I am more inclined to think it's feasible and actually not that hard (especially after you read the table that marks the latest technical features that allow flexibility, installed by Google and others).
The only question here is whether they will be willing to do so. Investors and PEs are often, in my experience, emphasize maximal output, and even when the downtime is minimal. When it comes to the big tech players, I can totally see this happening as they are less sensitive to operational pressures from investors at this granularity, but they do face an impossible competition.
At any rate, excellent piece. The US needs all the help it can get with respect to power, while China has much overcapacity, and data centers are sometimes seen as a way to eat into the insane oversupply.
I've enjoyed your hyperdimensional posts and am glad you're back at it. Re Thin Air, you mention fission, fusion and enhanced geothermal but are silent about solar, despite its low and declining costs (incl with batteries) Why silent? Is that simply because the US isn't good at making them right now?
Solar has already made an impact here—now, grid demand peaks kick in around 5-6PM if not later, because behind-the-meter solar pushed that peak back from its original 1-2PM.
Batteries can help with peak management, especially because we can still buy them at scale. For this reason, batteries the most comment asset in the ISO-NE interconnection queue (my neighborhood). But solar isn’t a capacity asset anymore, so if you’re going to attack demand peaks for system reliability, you’re better off just installing the battery instead of worrying about solar.
This sounds great - but we know that data center electricity consumption (especially for pre training) is extremely in flexible - the idea that Stargate would stop training machine God, even for two days and risk ruiningtheir training run seems like a tall ask, no? I'm sure I'm missing something...
They don’t have to stop training for two days—just 2-4 hour intervals, maybe 3-5 times per month. And if they don’t want to stop, they can discharge a battery during that time.
It’s a serious ask, but not as big as you’d think at first. And if the alternative is no compute, a hyperscaler might take the deal.
As someone who spends much of his time with data centers, technology and climate, I must admit I wasn't aware of this possibility. I will also admit that at first I was reluctant to believe that this is feasible and practical, after reading the paper you cited I am more inclined to think it's feasible and actually not that hard (especially after you read the table that marks the latest technical features that allow flexibility, installed by Google and others).
The only question here is whether they will be willing to do so. Investors and PEs are often, in my experience, emphasize maximal output, and even when the downtime is minimal. When it comes to the big tech players, I can totally see this happening as they are less sensitive to operational pressures from investors at this granularity, but they do face an impossible competition.
At any rate, excellent piece. The US needs all the help it can get with respect to power, while China has much overcapacity, and data centers are sometimes seen as a way to eat into the insane oversupply.
Great to have you back, and a good point.
Points taken. The low-hanging fruit seems obvious. Except, perhaps, to the oblivious.