Want to read Slashdot from your mobile device? Point it at m.slashdot.org and keep reading!

 



Forgot your password?
typodupeerror
×
AI

Power Shortage To Hit 40% of AI Data Centres by 2027, Gartner Warns (gartner.com) 43

40% of AI data centers will face operational constraints due to power shortages by 2027 as AI drives unprecedented energy consumption, research firm Gartner said on Tuesday.

Data center power requirements for AI-optimized servers are projected to reach 500 terawatt-hours annually by 2027, more than double 2023 levels, as companies rapidly expand facilities to handle large language model training and implementation.

The surge in power demand will outpace utility providers' ability to expand capacity, Gartner analyst Bob Johnson said, leading to higher electricity costs that will cascade through the AI industry. Some operators are already seeking direct agreements with power producers to secure guaranteed supply.

Power Shortage To Hit 40% of AI Data Centres by 2027, Gartner Warns

Comments Filter:
  • by reynolds_john ( 242657 ) on Tuesday November 12, 2024 @01:09PM (#64940231)
    "... costs which will cascade to the average consumer..."
    • Virginia is the data center capital of the world because they don't have to pay for the power infrastructure after they bribe their way to a new data center. The costs are paid for by all the power customers - as in you and me.

      But power shortages shouldn't be a big problem for them. They'll have more than enough diesel generators to run 24/7 when they need to.

    • by dpille ( 547949 )
      Perhaps it's not as bad as "40%." The 2023 US electricity production was 4178 TWH, of which 250 was apparently "data center power requirements for AI-optimized servers." Adding another 250 across the next three years is only 6% of that 2023 production. That's not too far off of what we expect to add to the grid for renewable electricity generation over those years. While I understand that it's an issue to fail to reduce fossil fuel electricity generation, it's not the same issue as brownouts/blackouts fo
  • Not a shortage (Score:5, Insightful)

    by Atmchicago ( 555403 ) on Tuesday November 12, 2024 @01:13PM (#64940239)

    Econ 101 here: there won't be a shortage, there will be an increase in electricity costs. Maybe the datacenters will be too costly to run, to the dismay of the AI companies, or maybe poor people will be priced out of the market. But a shortage would happen if e.g. the government intervened and set low price caps on electricity.

    Yes, the summary basically says a much, but then what's with portraying it as a shortage? Doesn't everything face "operational constraints" based on real world costs? The particulars may vary, but competition for limited resources remains a constant.

    • To meet the basic demand of your civilization is a shortage.

      Looking at the very basic demands of civilization not being met and saying that's okay because shrug that's capitalism shrug is what we call end stage capitalism or techno feudalism.

      That's all well and good but I do so wish you wouldn't talk about it as if it's a) inevitable and b) not something that's going to affect you.
      • A shortage of power is different from, say, a shortage of milk. The per-household cost for a gallon of milk doesn't skyrocket even if there is a shortage .. there's just less of it to go around.

        Heavy electricity users can, however, be charged ever-increasing prices per kw-h to curb their use.

        • Unexpected consequences. If AI can generate better cash it will be better able to afford electricity. Lower income consumers will struggle and consider alternatives like gas. Meanwhile gets warmer and harder to breathe for many.
      • To meet the basic demand of your civilization is a shortage.

        Looking at the very basic demands of civilization not being met and saying that's okay because shrug that's capitalism shrug is what we call end stage capitalism or techno feudalism.

        That's all well and good but I do so wish you wouldn't talk about it as if it's a) inevitable and b) not something that's going to affect you.

        Old Luke, just a year younger than old Obi-Wan in the first movie, looks over and says, "Amazing. Every word you said there is wrong. [juliansimon.com]"

    • by PPH ( 736903 )

      if e.g. the government intervened and set low price caps on electricity.

      Which they do in many cases. Depending on your state's utility commission, they approve different price structures for different user types. If the residential rates are not increased in step with the going market rates, less and less power will be available for those users. The utilities magically deciding that residential grids need to be taken off line for "maintenance" [wikipedia.org]. While data centers, willing to pay higher market rates will keep running.

    • Dominion's estimates from 2021 for 2028 are 35% below actual growth projections now. No biggie, just a 35% increase....in 7 years. Shortages exist now, they're just manageable within the current system. Unplanned for breakneck growth being added isn't a recipe for stability.

      The same projection for 2038? 130-150% percent short.

      And of course, supply has to be able to *reach* demand and NIMBYism prevents a lot of needed transmission lines.

      https://protectpwc.org/2023/08... [protectpwc.org]

    • Semantics and speculation.

      Your comments are insightful and correct.
      But, so are the responses.

      It is just semantics and verbiage in many ways. Operational constraints, shortages, perceived or relative shortages - it is partly just "spin" on what happens when demand outpaces production. But, there are differences depending on the forces jiggering the situation. Is demand outpacing actual production, "easy" to fix by ramping up existing infrastructure, or is demand outpacing production capacity, which requir

  • I anticipate the growth of data centers outside of the USA, powered by cheap solar, geothermal, nuclear, wind, etc. Intermittent availability will be buffered by batteries/standby storage (gravity sleds, air storage, etc.), and possibly by standby generators.

    Things will shift such that the constraint will then be bandwidth, but data transmission infrastructure is a lot cheaper than power transmission infrastructure.

    Or... they could do what Elon has been doing, and just run craptons of portable generators u

    • by rossdee ( 243626 )

      Offshore sounds like a good idea, not only can you use offshore solar and wind, but you can use the ocean for cooling, especially in deep water.

    • They put these datacenters on internet backbones...like Northern VA. They need the proximity so they aren't going to build masses of them far away from major traffic hubs.

    • Hence my suggestion that they should put the data centers in Iceland, of all places. Reason: plentiful available geothermal power.

  • by shilly ( 142940 ) on Tuesday November 12, 2024 @01:31PM (#64940287)

    Wonder if anyone on here is able to answer these:
    1. What is it about AI DCs that is so power-hungry? Because I wonder if there’s room for meaningful reductions in power consumption through optimisation, but the answer to that depends on what is driving the power requirements, hence my question
    2. Someone has already made a joke about solar + storage. But in seriousness, I wonder if anyone knows what percentage of a typical DC’s power draw over the course of a year can be met through on-site solar & storage? Is it 2%, 20%, 99%? I couldn’t hazard a guess.

    • It looks like it's a matter of compute density. The hardware tackling LLM training tasks doesn't take up much space compared to an EPYC or Xeon. Physics still rears it's ugly head in the form of power draw.

    • It's just like CPU's, memory, and most other semiconductor development over time. First they push for performance and capability, and then it gets refined for power efficiency. These days a key feature for laptops is battery life, so low power draw gets a lot of attention. I think we are still seeing early development for AI components, and if data centers can't get enough juice they will preferentially buy components that don't draw so much.

      Nvidia for example is now shipping their new Blackwell systems;
      "e

      • by GoTeam ( 5042081 )

        Some day I expect an AI will be given a big hunk of money to figure out what it needs to do to power itself. Have power plants built, improve transmissions lines, whatever.

        Do you want to be a battery like in the Matrix? Because that's how you get the Matrix!!! /s

        • People as batteries was jarringly unbelievable for me in that movie. What would we be storing up and discharging? Is there all that much extra energy in our excretions?

          What would concern me is if we could be organically harnessed as augmentations to the AI. A form of slavery perhaps, your brain would be used for certain kinds of inferences. And in a sense things are moving in that direction, right? Humans feed the AI's the training data that make them seem smart. Oh, and there have already been successful a

    • AI currently uses The same technology that video cards use to push pixels in games, basically massive parallelized processing of matrix math. And there's no getting around it That's just a very power-hungry thing. Modern video cards usually need anywhere from 150 to 300 watts and up. You just have a whole lot of circuitry that runs really really hot so it needs a whole lot of fans to move the heat around.

      As for solar and wind they've been capable of providing full base load power for some time The probl
  • by eepok ( 545733 ) on Tuesday November 12, 2024 @02:19PM (#64940447) Homepage

    The power supplied to any location is a known quantity (or at least "knowable quantity") at any point in time. There's only a "shortage" if you have been guaranteed the power and you're not receiving that power.

    Planning a self-imposed excess of demand is a YOU problem. Don't try to externalize the responsibility on your local grid because you're attempting to capitalize on the next craze.

  • I will wager the bubble will pop sooner or later...just like it did with bitcoin. I don't care what the Winkelvoss grifters say, bitcoin is a pyramid scheme. I think Generative AI will have to DRASTICALLY improve if it will be thought of as much more than that in 2027. We've been through this before....vague promises, no demonstration of clear value...immense hype...followed by a disappearance into the fringes of the technology community.

    Most of the tech fads lately have been kinda desperate and add lit
  • Since they are causing the shortage and resulting utility price hikes, make the Corporate Overlords pay for fucking once. In fact they should also be paying enough to cover a significant discount to residential rates. Corporations should be last and residential customers first.

    Oh who am I kidding? Soon there will an AI Recovery Surcharge on residential bills, so corporations can also fuck you in the ARS.
  • the ordinary domestic consumer - will pay for any upgrades - through the nose of course.

Surprise your boss. Get to work on time.

Working...