Even before the AI revolution (or bubble) took off in 2023, datacenters already used a decent amount of electricity. And the industry's increasing appetite for machine learning will require even more power.

The estimates for the amount of power AI-infused datacenters might use are incredible. "By taking a middle figure of 300 TWh for the annual global datacenter power consumption, Uptime Intelligence puts generative AI annualized energy at around 2.3 percent of the total grid power consumption by datacenters in the first quarter of 2024," the Uptime Institute predicted in conversation with The Register.

"However, this could reach 7.3 percent by the first quarter of 2025," Uptime told us.

On a similar note, Arm CEO Rene Haas has said a full quarter of the US's electricity might go towards datacenters (including those with AI) by 2030. Some countries might even see higher usage sooner - Ireland is estimated to be looking at giving a third of all its power to datacenters by 2026, according to the International Energy Agency (IEA).

These predictions might be a bit pessimistic (or optimistic, depending on what industry you're in), because they are predicated on the future success of big industry players like Amazon, Google, and Microsoft. And it's not because all of the hype surrounding AI could peter out - being able to consume that much electricity on a national and global scale runs up against the problem of actually finding places to generate it.

No stone left unturned for datacenter power sources

Thankfully, the path for increasing the share of power used by datacenters probably won't be achieved by rerouting. In order to operate, bit barns will have to bring their own power to the grid - offsetting their usage, and maybe even resulting in a net gain of available electricity.

All this energy won't be coming from a single source. Datacenters will need to thrive on a "buffet of choices," Uptime Institute analyst Jay Dietrich opined. Those choices will principally include solar, wind, nuclear, and natural gas - each of which has its unique pros and cons.

Solar and wind energy are the fastest growing sources of energy, and have been for 19 years in a row, according to climate think tank Ember in its 2023 review of global energy trends. However, power created by solar panels and wind turbines hinges on the weather - and in 2023 Ember reported these factors caused lower energy production than expected.

Additionally, the time of day is also an important factor in both solar and wind-derived electricity. Obviously, solar panels aren't producing much at night and wind strength isn't consistent throughout the day.

To rely solely on solar and wind energy - the latter of which averages to 50 percent of theoretical max power - a high amount of redundancy would need to be built, Dietrich explained. This would result in tons of extra power being produced at peak production times, but storing all of it for later use isn't feasible - the technology just isn't there yet.

Energy sources that provide a baseline level of power will have to support solar and wind. Nuclear is an appealing candidate, given the move to low-carbon sources of energy. We're already seeing this pan out, with Amazon moving into a datacenter right next to a Susquehanna nuclear power plant in Pennsylvania.

"The traditional gigawatt scale nuclear power plants (NPPs), commonly 2-4 gigawatts, are providing critical baseload power to the grid," Omdia analyst Alan Howard explained. "In fact, there is a reasonable amount of recommissioning of shuttered NPPs."

The alliance between datacenters and nuclear power is actually quite beneficial for both parties. Operators can reserve capacity by using a purchasing power agreement (PPA), which provides nuclear power plants with much needed liquidity. This kind of agreement enabled the Palisades, a Michigan-based nuclear power plant to be recommissioned last year when Wolverine Power Cooperative signed a PPA with Holtec, the owner of the plant.

The future of green tech is murky

But just investing in existing technologies isn't going to cut it, so big tech is looking into cutting-edge solutions for delivering future needs. Investments are going towards lots of solutions which may or may not pan out - also known as throwing money at the problem until it's solved.

Take for instance Exowatt, a solar panel startup that broke cover last month with the financial support of OpenAI CEO Sam Altman and Andreessen Horowitz. Exowatt's pitch is that it claims to be able to put solar panels and heat batteries into a single unit. With tens of millions of dollars behind it, Exowatt may succeed. Or it may not. We won't know for some time.

Nuclear is also an incredibly important frontier for innovation - particularly in respect to small module reactors (SMR), which are basically envisioned as mini nuclear power plants that are easier to produce and run. While the traditional nuclear power plant can provide power on a gigawatt scale, SMRs should deliver only around 50 to 100 megawatts each. An even smaller variant - termed the microreactor - is theorized to produce 10 to 50 megawatts.

"In time, SMRs will be the pathway for smaller yet critical power generation for many markets that are in need of baseload power for the community, and those deployments to support industries like datacenters," Howard predicted.

However, "in time" doesn't mean SMRs will be powering datacenters in just a few years. "The R&D on SMRs and microreactors has been going on a long time, but developers are seeing light at the end of the tunnel," he explained.

"In the short term, say two to four years, what we'll see is operating test reactors. The reality, however, is that broad commercial deployment and adoption is likely in the ten-plus year range."

Likewise, Dietrich expects it will take at least five to ten years for SMRs to hit the scene, owing to hurdles with permits and manufacturing. He notes that there are around 70 firms touting plans for SMRs only a few will be able to produce systems in commercial volumes.

"I certainly hope we improve the timeline," Howard urged. "One of the biggest challenges to the SMR and microreactor timeline is regulatory processes. The US Department of Energy and the federal government have been working to improve this while maintaining the necessary due diligence."

Of course, there's also nuclear power's classic problem: convincing locals. Just like with any large project, building a nuclear power plant will require the consent of those in the neighborhood, who can sometimes be skeptical of having a nuclear reactor nearby.

"Acceptance of nuclear power from a public perception perspective has made a lot of progress, but there is work to do," Howard conceded.

"Just like datacenters, there will always be some that will fight nuclear. Saying we need more education is true, but it's a bit cliché. The right way to do it is community engagement that spans subject areas that are of concern to community stakeholders and demonstrates how a datacenter, or SMRs, or a Walmart for that matter, will benefit them."

Also, we're probably going to have to use more natural gas

But even after expanding the use of existing technologies and introducing new ones, it still might not be enough to solve the power needs of server centers. Although operators like Amazon and Google market themselves on their commitment to green energy, it's not entirely clear if they'll be able to expand as much as they'd like without relying on natural gas.

In the hierarchy of carbon-based energy sources, natural gas is certainly one of the least offensive - but it still does create pollutants like carbon dioxide, methane, and nitrous oxide. Natural gas as a power source for datacenters has been proposed for years now, stretching back at least as far as 2020 per reports [PDF] from companies like construction biz Black & Veatch.

We asked Amazon about its use of natural gas and other fossil fuels, but the cloud provider did not answer our queries - though it did recount its progress in bringing up non-carbon sources of energy. It's hard to say how much Amazon in particular is relying on natural gas, but The Register reported last year that it planned to transition some Oregon datacenters to natural gas fuel cells.

For its part, the natural gas industry predicts the AI boom and the datacenters supporting it will be a growth opportunity. "Natural gas demand growth is continuing, empowering the US as electricity demand grows," declared TC Energy President and CEO Francois Poirier during the company's quarterly earnings call on May 3. "New growth drivers like datacenters will help continue that positive growth momentum."

In fact, the AI revolution might be mostly powered by natural gas, according to a Goldman Sachs report [PDF], which forecast that through 2030, 60 percent of the power required for new datacenters will come from natural gas, while 40 percent will come from renewables. This could be especially true if cutting-edge renewable tech doesn't work out the way the industry hopes - which would make datacenters a net loss for green energy.

25 percent or more power used by datacenters by 2030? Try 10 percent or less

It might seem like the whole field is in agreement that datacenter power consumption is bound to grow and be the single largest consumer of electricity across the world, but if anything these predictions are probably at the most extreme end of the spectrum. It might be possible for datacenters to consume that much power if AI lives up to all of the hype - but reality will likely be somewhat different.

The main issue is that building out all these new sources of power on this kind of scale is going to be extremely challenging. Datacenters are only going to power on if the power is there to accommodate them, according to Dietrich - who estimates even 20 percent of US power usage going to datacenters by 2030 is probably not going to happen. Goldman Sachs similarly estimated a mere 8 percent, give or take three, would be consumed by datacenters in the same time frame.

If the datacenter industry won't be able to consume more power for its aims, then it will have to ration its power more carefully. One such solution is offloading AI work from the cloud and into local devices. Obviously, the typical phone or laptop won't be able to run GPT-4 - but there are smaller AI models that might still be powerful enough, such as LLaMa-7B, which has just seven billion parameters instead of the hundred billion in GPT-4.

"AI development is in its early, brute force stage.," Dietrich predicts.

"As computer scientists and programmers grapple with the energy, computer hardware, and software demands of these new systems, innovation will likely deliver some general and many targeted AI applications on standard, lower power compute platforms with more of the work performed by well designed algorithms inside the software. This should enable the delivery of AI applications more efficiently and with smaller resource demands."

Arm CEO Rene Haas has recommended that chip designers pursue more efficient designs in order to use the same amount of power for more work - and it's easy to guess what architecture Haas suggests for doing that.

Ultimately, Dietrich argued, the datacenter industry's power problem will "have to work itself out" - and that could certainly mean abandoning lofty goals of green energy-powered datacenters. ®

https://www.theregister.com//2024/05/16/datacenter_power_demands/