That said, a lot of data centres will be constructed over the next few years, and they will use a lot of electricity. That's true even if the maximalist position advanced by Varun and others is fantasyland stuff. And if those data centres can be made even a little bit flexible, they can be mostly powered by cheaper, faster to build renewables and batteries rather than gas, which will be a huge win.
For what it's worth as somebody who actually uses LLMs for coding every working day, Silicon Valley is high on its own supply when it comes to generative AI. This is in large part because because the thing it turns out to be genuinely really good at is a) coding, and b) more specifically, small greenfield prototype software projects (it's useful, but much less so, with large existing systems that actually have consequences if they fail). Silicon Valley systematically overestimates the utility of greenfield prototype software projects.
This is such an important shift in the conversation. We keep asking how much electricity AI will consume, but flexibility may matter just as much as consumption. If data centres can move or pause workloads when the grid is constrained, they stop being just another huge load and start becoming part of the grid solution.
I keep coming back this question: whether this flexibility will actually be enforceable at scale, especially when reliability and AI workloads are competing priorities. That’s where the idea gets really interesting.
Efficient use of the grid is important, but using the existing grid more efficiently currently means running the fossil plants that make up ~60% of the grid 24/7 instead of part-time. Efficiency gains also only get you so far until you need more capacity, and of course, AI demand is not going to neatly stick to this slack capacity slice of the pie, especially if it booms as much as the bulls think. Having not read the Nature paper, I somehow suspect they have the right idea.
Interesting, but the Joe & Jenny Paycheck family energy needs to be generated and stored not in the industrial AI arteries, but within the distribution capillaries of the grid: Large workplace, shopping center, and multi-family residential parking lots with solar canopies, integrated non-flammable BESS, and Vehicle-2-Grid chargers,....everywhere, ASAP.
The quote "AI customers are doing the most economically important work in America" made my vision go white with rage.
Varun may be smart but he has a serious case of engineers/physicists disease.
Now excuse me while I go write "Adventurism is not the solution" on the white board fifty times.
...Not all engineers....
That said, a lot of data centres will be constructed over the next few years, and they will use a lot of electricity. That's true even if the maximalist position advanced by Varun and others is fantasyland stuff. And if those data centres can be made even a little bit flexible, they can be mostly powered by cheaper, faster to build renewables and batteries rather than gas, which will be a huge win.
For what it's worth as somebody who actually uses LLMs for coding every working day, Silicon Valley is high on its own supply when it comes to generative AI. This is in large part because because the thing it turns out to be genuinely really good at is a) coding, and b) more specifically, small greenfield prototype software projects (it's useful, but much less so, with large existing systems that actually have consequences if they fail). Silicon Valley systematically overestimates the utility of greenfield prototype software projects.
This is such an important shift in the conversation. We keep asking how much electricity AI will consume, but flexibility may matter just as much as consumption. If data centres can move or pause workloads when the grid is constrained, they stop being just another huge load and start becoming part of the grid solution.
I keep coming back this question: whether this flexibility will actually be enforceable at scale, especially when reliability and AI workloads are competing priorities. That’s where the idea gets really interesting.
Efficient use of the grid is important, but using the existing grid more efficiently currently means running the fossil plants that make up ~60% of the grid 24/7 instead of part-time. Efficiency gains also only get you so far until you need more capacity, and of course, AI demand is not going to neatly stick to this slack capacity slice of the pie, especially if it booms as much as the bulls think. Having not read the Nature paper, I somehow suspect they have the right idea.
Interesting, but the Joe & Jenny Paycheck family energy needs to be generated and stored not in the industrial AI arteries, but within the distribution capillaries of the grid: Large workplace, shopping center, and multi-family residential parking lots with solar canopies, integrated non-flammable BESS, and Vehicle-2-Grid chargers,....everywhere, ASAP.