I worked for a company making GPU clouds. The biggest problem we had for deployments was not getting GPUs -- we had plenty of those sitting in warehouses. The biggest issue was finding data center space with sufficient power and cooling. There was plenty of square footage, just not enough power for it all.
They're now building gigawatt datacenters to handle all the GPUs.
The big question is were to build them. There are only a few places with cheap and plentiful power. One of those is Quebec (but it's not that big and there is a lot of regulation). Another is Texas (except their grid isn't very stable). And the last is China. And you can't build a datacenter in China unless you're Chinese.
It'll be interesting to see how this pans out. Maybe the current admin (which is big on deregulation) will make it easier to build power plants, especially nuclear ones.
Edit: I wrote this comment four days ago. I couldn't figure out why I was suddenly getting a bunch of replies to it. Apparently when HN does a second chance, they just reset the time on all the comments. Odd, but I guess it makes sense knowing what I know about how the sorting is calculated. It's probably the easiest way.
In my part of PA there are 3 in the process of going in nearby. I think the largest of the 3 is "only" 828 megawatts though. One of the others is supposed to be 300MW, and I'm not sure about the 3rd. There is another group talking about 3 more campuses with a combined power budget of 1.3GW about 55 miles from here. But then while we don't have cheap land, we do have nuclear and hydroelectric in the area, so I guess the makes it attractive.
I think Texas (ERCOT) is a terrible option these days. Meta recently made a fantastic choice by picking Louisiana for their new monster. The MISO grid tends to be cheaper and less volatile than ERCOT.
Energy efficient computing is a very exciting field. I hope it will get more attention driven by these economic constraints.
As a short teaser:
Landauers principle suggests that the energy required to erase one bit off information is bounded from below by k_BTln(2). This could lead us down a path towards reversible computing, to avoid energy costs for deleting information.
The work started around 10-15 years ago and is now largely done. Many people confuse large absolute numbers like 1 kW with inefficiency but today's GPUs/TPUs are close to the practical efficiency limit with today's 3 nm technology.
My hunch: we’ll see three things happen in parallel
- AI backend providers vertically integrating into energy production (like xAI’s gas plants, or Meta’s local generation experiments),
- renewed interest in genuinely efficient computing paradigms (e.g. reversible/approximate computing, analog accelerators),
- a political battle over whether AI workloads deserve priority access to power vs. EVs, homes, or manufacturing, alongside an increase in energy prices.
You need cheap, reliable power + political/regulatory willingness + cooling. That’s a very short list of geographies. And even then, power buildout timelines (whether nuclear, gas, or grid-scale solar+batteries) move at "utility speed", which is decades, not quarters. That doesn’t match the cadence of GPU product launches.
The timing mismatch is crucial - data centers can be built in 12-18 months, but new power generation takes 5-10 years minimum. We're essentially trying to scale AI demand faster than energy infrastructure can physically respond. This creates interesting arbitrage opportunities in power-rich but compute-poor regions.
Remove restrictions on solar import from China. 62 GW may sound a large number, but China added 277GW solar in just 2024. They have the surplus capacity and hence cheapest price.
I got down voted when I said China is likely to win the AI race as they are also targeting the other big cost of computing power/energy on another thread. Today solar + BESS is cheaper than coal where as costs for both keep decreasing each year.
One really interesting strategy the US could pursue here would be to heavily tariff solar[1] and just randomly attack wind projects[2]. Just completely self-own itself on the two cheapest energy sources.
It wouldn't make any sense, but it would be provocative, really drive engagement.
Ah, finally an acknowledgement that the melting 12VHPWR connectors short-circuiting Nvidia's top-of-the-line hardware, the RTX 4090 and 5090, may finally have economic ramifications. Heh.
20 comments
[ 5.5 ms ] story [ 43.8 ms ] threadThey're now building gigawatt datacenters to handle all the GPUs.
The big question is were to build them. There are only a few places with cheap and plentiful power. One of those is Quebec (but it's not that big and there is a lot of regulation). Another is Texas (except their grid isn't very stable). And the last is China. And you can't build a datacenter in China unless you're Chinese.
It'll be interesting to see how this pans out. Maybe the current admin (which is big on deregulation) will make it easier to build power plants, especially nuclear ones.
Edit: I wrote this comment four days ago. I couldn't figure out why I was suddenly getting a bunch of replies to it. Apparently when HN does a second chance, they just reset the time on all the comments. Odd, but I guess it makes sense knowing what I know about how the sorting is calculated. It's probably the easiest way.
Sort of a baffling statement. Quebec is gigantic. It would be a top-20 nation, by extent, if it were a nation.
They aren’t big on deregulation at all. They’re big on selective regulation. They’re also big on killing any power project that isn’t oil or coal.
https://apnews.com/article/trump-offshore-wind-renewable-ene...
We are in for a painful lesson on why China’s investment in renewables wasn’t just good for their ecology.
Has a lot of hydroelectric and the nights get super-cold, so you could open the roof for free ventilation
As a short teaser: Landauers principle suggests that the energy required to erase one bit off information is bounded from below by k_BTln(2). This could lead us down a path towards reversible computing, to avoid energy costs for deleting information.
https://en.wikipedia.org/wiki/Landauer%27s_principle
https://en.wikipedia.org/wiki/Reversible_computing
- AI backend providers vertically integrating into energy production (like xAI’s gas plants, or Meta’s local generation experiments),
- renewed interest in genuinely efficient computing paradigms (e.g. reversible/approximate computing, analog accelerators),
- a political battle over whether AI workloads deserve priority access to power vs. EVs, homes, or manufacturing, alongside an increase in energy prices.
You need cheap, reliable power + political/regulatory willingness + cooling. That’s a very short list of geographies. And even then, power buildout timelines (whether nuclear, gas, or grid-scale solar+batteries) move at "utility speed", which is decades, not quarters. That doesn’t match the cadence of GPU product launches.
It wouldn't make any sense, but it would be provocative, really drive engagement.
[1] https://seia.org/news/solar-tariff-impacts/
[2] https://www.npr.org/2025/08/31/nx-s1-5522943/trump-offshore-...
I would have thought so.
If so building data centres near hydro or geothermal plants (I'm from New Zealand where we have a lot of both) would make sense