alt Hacker NewsI've done a bit of research into it, and I think it's a mix.
The energy thing is exaggerated. Want to compensate for a week of heavy AI coding usage? If you drive to work, telecommute for one day. Skip one long-distance DoorDash order. Cars use tons of energy, a lot more per-user than the fraction of data center capacity you're using with your AI stuff. A burrito on DoorDash delivered using an average MPG gas car uses something like a week's worth of heavy AI usage in energy.
The water thing is region-specific and where is a problem it's due to incompetent or corrupt goveranance. Bottom line is that if a data center is using too much water, it means they're not paying enough for that water. Make them pay more for their water and they'll magically discover closed cycle cooling.
The noise thing is in some cases real, and needs to be fixed with local regulation or a class action suit by neighbors to force them to deal with the noise problem.
I am a little puzzled though by where these things are being built. Phoenix? Other than cheap solar power I can't tell why you'd want a data center there given the energy needed for cooling.
Why are we not putting them up in the Arctic circle? Build run-of-the-river hydro in northern Canada and put them way up there. Free cooling, especially in winter. AI is not a very latency-sensitive load so an extra 50ms doesn't mean much.
I suspect state incentives are a reason. The giant Amazon data center proposal in Texas with the giant natural gas power plant sounds a lot like "taxpayer funded indirect fossil fuel industry subsidy." Offer some kind of big state incentive to get a huge natural gas customer with long-term locked-in demand.
Replies
> The energy thing is exaggerated. Want to compensate for a week of heavy AI coding usage? If you drive to work, telecommute for one day. Skip one long-distance DoorDash order. Cars use tons of energy, a lot more per-user than the fraction of data center capacity you're using with your AI stuff. A burrito on DoorDash delivered using an average MPG gas car uses something like a week's worth of heavy AI usage in energy.
It doesn't work like that because skipping your DoorDash burrito doesn't make the energy not used for that available for data center use.
The AI data center energy problem is that they use an order of magnitude more than regular data centers and they are being built at a massively higher rate than regular data centers and they are being built in places where many are being built and their owners are in a hurry.
The electric grid is designed to handle growth in demand, but this is growth way beyond anything that was anticipated.
> I am a little puzzled though by where these things are being built. Phoenix? Other than cheap solar power I can't tell why you'd want a data center there given the energy needed for cooling.
Not Phoenix, but I learned an interesting fact about some of the huge data centers in Las Vegas, which pre-date AI: since it's 72 degrees or cooler for more than 70% of the year, during those months they can just pipe in outside air without needing to cool it first.
On top of that, I think both Phoenix and Vegas have fewer natural disasters than other major metro areas.
> Make them pay more for their water and they'll magically discover closed cycle cooling.
A closed cycle cooling system doesn't solve the overall problem of water usage. Yes you have a cycle within the datacenter, but you need to then move that energy somewhere else, out of the system. Eventually that energy needs to find its way in the environment. A close loop system for a datacenter reduces the water use on site, but actually increase electro use. If that electricity comes from a power plant that is cooled by water you just moved the problem upstream
It's gotta be costs, right? Everyone wants this done fast and at large scale, so they're going for any sites with existing large electrical connections I assume.
> The energy thing is exaggerated.
https://www.wpr.org/news/wisconsin-energy-demand-to-rise-dri...
> Wisconsin energy demand to rise 40 percent by 2032, driven by data centers, draft report finds
> Draft report: Over 72% of peak electricity load is 'associated with three active data center developments'
https://www.wpr.org/news/we-energies-add-3gw-electric-grid-d...
https://www.cleanwisconsin.org/wisconsin-ai-data-center-powe...
> Two AI Data Centers Require More Energy Than All Wisconsin Homes Combined: Just two AI data center projects — the Microsoft data center in Mt. Pleasant and Vantage data center in Port Washington — will require a combined 3.9 gigawatts (GW) of electric power, which is enough energy to power 4.3 million Wisconsin homes. There are only 2.8 million housing units in the entire state.
Compare with the city of Los Angeles - https://www.ladwp.com/who-we-are/power-system
> In fiscal year 2023-24, we supplied more than 20,749 gigawatt-hours (GWh) of power for more than 1.6 million electric service customers, including nearly 6,000 in the Owens Valley.
That's an average of 2.3 GW over the year.
-----
> The water thing is region-specific and where is a problem it's due to incompetent or corrupt goveranance. Bottom line is that if a data center is using too much water, it means they're not paying enough for that water.
https://www.opb.org/article/2026/01/15/as-googles-water-dema...
> When Google purchased a defunct aluminum smelter in The Dalles, it also acquired the right to pump massive amounts of water from underground — up to 3.88 million gallons a day. That’s almost as much water as the entire city averaged per day in 2024.
> The corporation transferred those water rights to the city in 2021, Bentz noted, allowing The Dalles to pump more water from underground. The city only obtained part, not all, of the former smelter’s water rights, Mays said after OPB’s story published.
It's the rights to pump the water. Google can pump up to 3.88 million gallons per day and the city can't bill them a penny for that. Also, because it's an old claim, it's likely more senior to other water rights.
https://www.oregon.gov/owrd/programs/waterrights/pages/defau...