Hacker News Reader

Comment by labrador | Hacker News Reader

For scale: The 1960's era US Navy submarine I served on had a 78MW reactor, so 10GW is 128 nuclear submarines

For a better sense of scale: it's about 2% of the average US electricity consumption, and about the same as the average electricity consumption of the Netherlands (18 million people)

Wtf, and this is from 1 company

How many atrophic,xAi,google,Microsoft would be????

having around 5% entire country infrastructure on AI hardware seems excessive no???

Around 5% in the next 5 years for AI alone sounds pretty in-line with projections I've seen.

Isn't this pretty bad for the climate? I don't dare to ask ChatGPT now /S

No this is from just one partnership, my sense is that OpenAI alone wants more than that.

5% to 10% of US electricity going to AI in 10 years is consistent with the current valuations of AI companies.

For another sense of scale: A 500MW AI-centric datacenter could cost $10 billion or more to build. So 10GW is $200 billion!

Some more context, Nuclear power stations can be up to 2GW, offshore windfarms are seemingly hitting a plateau at ~1.5GW, individual turbines in operations now are 15MW. Grids are already strained, 525kV DC systems can transmit ~2GW of power per cable bundle…

Adding 10GW of offtake to any grid is going to cause significant problems and likely require CAPEX intensive upgrades (try buy 525kV dc cable from an established player and you are waiting until 2030+), as well as new generation for the power!

Yeah the path forward here is going to be Apple-like vertical supply chain integration. There is absolutely no spare capacity in the infra side of electrical right now, at least in the US.

And there is great cost saving potential in vertical integration. Distribution and transmission are huge costs. If you can build a data center right next to a power plant and just take all their power you get much better prices. Not trivial to do with the kinds of bursty loads that seem typical of AI data centers, but if you can engineer your way to a steady load (or at least steady enough that traditional grid smoothing techniques work) you can get a substantial advantage

> bursty loads that seem typical of AI data centers

Don’t datacenters want to run at their rated capacity 24/7?

I don’t think that’s possible with large scale power infrastructure, and specifically grid infrastructure is so tightly regulated. Closest that I’m aware of was TSMC buying the output of an entire offshore windfarm for 25yrs (largest power purchase contract ever - TSMC / Ørsted)… maybe Microsoft re starting nuclear power plants, or Google reporting offshore wind sites come out of contract (but nothing at the 10GW scale).

In the long run, perhaps this will give us a better power grid, just like the dotcom bubble gave rise to broadband?

But that's assuming they actually have to transport power over long distances right? If they colocate these massive AI datacenters right next to the power generation plants, it should be cheap to transport the power. You don't need to upgrade massive sections of the grid and build long-distance power lines.

The xAI Colossus 2 1GW data centers seem to be located about ~20 miles from the power generation utility (https://semianalysis.com/2025/09/16/xais-colossus-2-first-gi...)

20 miles is a long way to move power, on land you have huge issues over getting permits for construction as it’s so disruptive, offshore specialist vessels that serve a global existing supply chain.

This still blows my mind.

If each human brain consumes ~20W then 10 GW is like 500 M people, that sounds like a lot of thinking. Maybe LLMs are moving in the complete opposite direction and at some point something else will appear that vaporizes this inefficiency making all of this worthless.

I don’t know, just looking at insects like flies and all the information they manage to process with what I assume is a ridiculous amount of energy suggests to me there must be a more efficient way to ‘think’, lol.

We know for a fact that current LLMs are massively inefficient, this is not a new thing. But every optimization you make will allow you to run more inference with this hardware, there's not a reason for it to make it meaningless any more than more efficient cars didn't obsolete roads.

> But every optimization you make will allow you to run more inference with this hardware

Unless the optimization relies in part on a different hardware architecture, and is no more efficient than current techniques on existing hardware.

> there's not a reason for it to make it meaningless any more than more efficient cars didn't obsolete roads

Rail cars are pretty darned efficient, but they don’t really work on roads made for the other kind.

Or just ten very safe РБМК reactors rated 1GW each (they can't explode).

You almost got me. RBMKs had this problem with large positive void coefficients that was buried by the Soviet Union, which lead to Chernobyl.

The control rods with the graphite on the tip was the cherry on top...

A big power station of any type is ~1GW. Nuclear is slow to build, so I'd have to guess natural gas.

It should be illegal to build that much fossil fuel powerplants to just train LLMs.

The platant disregard of global warming by AI investors is truly repulsive.

Keep in mind that the industrial processes that consume fossil fuel also contribute to quality of life in various ways. Improvements in emergency response and early detection infrastructure alone have resulted in deaths from extreme weather events reaching record low levels. Poverty as a whole has seen record-breaking decreases over the last 30 years.

So there are other factors to weigh besides how much contributes to CO2 emissions.

The US is adding significantly more solar, and slightly more wind, than natural gas every year. This doesn't have to be placed where people already are, but can be placed where energy is the cheapest, which favours solar and wind substantially more than gas (or nuclear).

The reasonable (cost effective, can be done quickly) thing to do is put this wherever you can generate solar + wind the most reliably, build out a giant battery bank, and use the grid as a backup generator. Over time build a better and better connection to the grid to sell excess energy.

Trump is personally and vindictively against green energy.

He wants coal and gas.

When the price of gas is so much higher than solar, that hardly matters. No reason the data centre have to be in the US.

A typical reactor core is 1 GW, so it's also one rather big nuclear power plant.

More like two (and a half )

Reply on news.ycombinator.com