back to article AI boom is great news for the nuclear power dreamers

Growing demand for power hungry AI datacenters has executives at Canadian uranium mine Cameco glowing with anticipation. Speaking on the resources firm's Q1 2024 earnings call Tuesday, CEO Timothy Gitzel indicated the computationally intensive nature of AI workloads, and the challenges faced obtaining adequate power for the …

  1. Filippo Silver badge

    Overhype LLMs.

    Build nukes to power LLMs.

    LLM bubble bursts.

    And all that's left is gigawatts of zero-carbon power, the main cost (build) already sunk, and desperate to be sold.

    Sounds good to me.

    1. Anonymous Coward
      Anonymous Coward

      The generators will be useful to support the inevitable take over of the robots...

    2. Anonymous Coward
      Anonymous Coward

      Let's talk again in 15 years when the first nuclear reactor will be be operational...

      Not sure AI will wait that long, though.

  2. Plest Silver badge
    Happy

    So Germany obviosuly won't be involved in this oncoming AI revolution and its demand for "3-Mile-Isle" style power then! Ha ha!!

    1. Yet Another Anonymous coward Silver badge

      Germany has the best solution for zero green energy.

      Mine lignite, export it to Poland (thus creating negative carbon), burn coal and import power.

      Unions happy, greens happy and plenty of reliable (if not quite cheap) power

    2. Anonymous Coward
      Anonymous Coward

      German power is already there. They put 1GWc of renewable on line per month on average. Your fantasy nuclear needs 10 year minimum to produce the first kWh. Whoever thinks they need nuclear power to fuel their AI revolution will have to wait quite a bit (or come to their sense).

  3. Anonymous Coward
    Anonymous Coward

    The Future's So Bright,

    I Gotta Wear Shades ... and no doubt a face mask too.

    1. MyffyW Silver badge

      Re: The Future's So Bright,

      This is why I come to El Reg

  4. Alf Garnett

    It sounds like a good deal to me. Nuclear plants don't produce any air pollution. The spent fuel is stored on site. With coal plants their spent fuel it vented to the atmosphere. Also building datacenters next to nuke plants means less energy lost in transmission vs putting the datacenter hundreds of miles from the nuke plant.

    1. sabroni Silver badge
      Thumb Up

      re: It sounds like a good deal to me.

      Me too! Me and mine will be long dead by the time that waste starts leaking.

    2. Anonymous Coward
      Anonymous Coward

      Sounds like a scam to me

      Planning, get all the permits to start and build an AI or other massive data centre as long as sufficient power and water for cooling is already available: one to a few years?

      Planning, get all the permits to start and build any power plant (coal, nuclear, gas) build and running in the West and democratic countries with plenty of permits and appeal processes: 5 to 15 years?

      So what is most likely going to happen:

      Announce building a massive data centre plus investing (directly or indirectly) in a big local nuclear power plant. Call the whole thing sustainable and net zero. Get the permissions for the data centre quickly, and even try and get government subsidies or tax exemptions for it. Have it up and running in one or a few years. This AI gold rush can't afford any delays. Then its mega rich Big Tech owners see it can run by the electricity provided by the already existing power grid. Prices at peak times soar super high for the small people on variable contracts. Prices for fixed contracts go up quickly too for the small people. Then all companies including the the mega rich Big Tech owners get the opportunity to earn very good money for not consuming electricity for a few peak hours and shift their workloads to other places in the world who have a large temporary excess production from solar and wind at noon in regions with simultaneous sun and windy weather. There they'll get paid high prices per kiloWatt consumed because the overproduction there risks to destabilize the local net. The AI data centre owners realize that sustaining the local power generation shortage is much more profitable then investing to expand power generation. They quickly dump the plan to invest in any power generation and find external reasons like stringent regulation or protest or material shortages to blame. The little people keep subsidizing this schema on their exploding power bills. It fits perfect in their so often used "promise all sorts of goods to local communities and most importantly their politicians, get out all you want now, take as much as you can, make sure you can bail out of whatever you promised at any time if it is not profitable" play book.

      So in order to not make it a pro or anti nuclear comment: let them (or their partners they invest in with strong binding and enforcible contracts), build and get up and running those nuclear (or renewables if they promise those) power plants up and running FIRST and THEN they can start building and operating their energy wasting warehouses.

  5. Splod

    We had better find some new sources of Uranium then. We seem to be having problems with many of the sources. A lot of them may end up on the BRICS side of the world. https://world-nuclear.org/information-library/nuclear-fuel-cycle/mining-of-uranium/world-uranium-mining-production

    Small, local reactors would seem a good answer to generation problems (without any idea of cost) as they could provide a more localised supply which could be huge benefit in unstable times. And we need a lot more generation and distribution if we seriously want electric vehicles - even though I don't think they are the panacea they are claimed to be.

  6. Anonymous Coward
    Anonymous Coward

    We're going about this the wrong way...

    A human brain requires about 25 watts of energy. A gigawatt would be enough to power 40 million human brains.

    Wouldn't our energy (pun intended), be better spent on working out how we can do so much, so efficiently. 200 gigawatts would be enough to give us the brainpower of the entire (human) planet.

    Sorry... a load of nonsense really. Just a random thought that popped into my head.

    1. Filippo Silver badge

      Re: We're going about this the wrong way...

      It's not a load of nonsense. It's a good question, with non-obvious answers.

      AFAIK, we could build artificial neural networks in hardware, and they would be several orders of magnitude more energy-efficient. There have been some experiments in this direction, see neuromorphic chips.

      I believe that the reason we don't do that at scale is that if you build such a device, many features of the network design would then be baked in the hardware and difficult or impossible to change. Whereas big banks of GPUs can be easily reprogrammed if you come up with some new idea. In the current context, designs get done and redone very quickly, making the idea unfeasible. Can't do an entire hardware-level redesign every few months.

      If we get to the point where we can say "this, this is how it's done for the foreseeable future", then yes, devices could be made that mimic the brain much more closely and that run off watts instead of kilowatts.

      1. Anonymous Coward
        Anonymous Coward

        Prometheus Complex

        > we could build artificial neural networks in hardware, and they would be several orders of magnitude more energy-efficient

        man, open a biology book and look for action potential. Then come back here.

        > If we get to the point where we can say "this, this is how it's done for the foreseeable future", then yes, devices could be made that mimic the brain much more closely and that run off watts instead of kilowatts.

        1. Filippo Silver badge

          Re: Prometheus Complex

          Uh, I would, but I doubt I would find anything about computers, which is what we're discussing here. I was comparing NNs implemented in GPUs to NNs implemented in neuromorphic circuits. I was not comparing to NNs implemented in biology.

          I don't claim to be a specialist in any of those fields, though. If I said anything wrong, I'd welcome correction, but please be a bit more clear?

          1. Anonymous Coward
            Anonymous Coward

            Re: Prometheus Complex

            You could learn about ion concentration gradients and how little energy they need. How fast they commute.

            That would possibly lead you to reconsider your initial statement "we could build artificial neural networks in hardware, and they would be several orders of magnitude more energy-efficient."

            Also learn about how biological and flexible brains are. There is no segregation between learning and inference for instance.

            1. Filippo Silver badge

              Re: Prometheus Complex

              Okay. There is a reading comprehension problem here. I apologize for not being sufficiently clear, and I'll try to explain myself again.

              The OP was wondering whether we could mimic biology in order to improve performance. I answered that some research is being done in that direction, specifically neuromorphic chips. Neuromorphic chips are not biological neurons, they are just bits of silicon that get arranged in a somewhat similar fashion.

              Therefore, I was comparing neuromorphic chips to banks of GPUs.

              Neuromorphic chips are several orders of magnitude more efficient than banks of GPU.

              I have never claimed they are more efficient than actual biological neurons. I never attempted to compare actual biological neurons to anything. The energy efficiency of biological neurons does not feature anywhere in what I was talking about.

              Is that clearer now?

              1. Anonymous Coward
                Anonymous Coward

                Re: Prometheus Complex

                Not sure where the "reading comprehension issue" is coming from... Let me remind you what the OP was posting.

                >> A human brain requires about 25 watts of energy. A gigawatt would be enough to power 40 million human brains.

                Wouldn't our energy (pun intended), be better spent on working out how we can do so much, so efficiently. 200 gigawatts would be enough to give us the brainpower of the entire (human) planet. Sorry... a load of nonsense really. Just a random thought that popped into my head.

                To which your reponse is:

                >It's not a load of nonsense. It's a good question, with non-obvious answers. AFAIK, we could build artificial neural networks in hardware, and they would be several orders of magnitude more energy-efficient [than what? human brain was cited twice already by OP!]. There have been some experiments in this direction, see neuromorphic chips.

                > Therefore, I was comparing neuromorphic chips to banks of GPUs.

                It looked a lot like you were comparing neuromorphic chips to human brains. JUST AS THE OP WAS comparing human brains to hardware. Banks of GPUs were not mentioned anywhere in the upstream posts.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like