AI's appetite for power could double datacenter electricity bills by 2030
(2024/06/26)
- Reference: 1719432975
- News link: https://www.theregister.co.uk/2024/06/26/datacenter_electricity_demand_study/
- Source link:
A study predicts that US power consumption from datacenters will more than double by 2030.
The forecast comes from Rystad Energy in its latest [1]report on US power demand and supply, which includes figures for datacenters in particular. The analysts estimate that datacenter power demand will sit at 307 terawatt hours (TWh), more than double the 130 TWh the firm says was consumed in 2023.
The reasoning behind rapidly increasing power draw is the usual suspect. "The growth in electricity demand for datacenters will be heavily driven by those focused on artificial intelligence (AI)," Rystad says, "which consumes more electricity compared to traditional computing."
[2]
To meet this demand, Rystad says the best energy sources for datacenters will be renewables thanks to their scalability and increasing viability. Consequently, wind energy supply is expected to just about double by 2030, while solar will roughly triple, if the calculations are correct. Power generated by coal will decrease by about half, though won't be totally eliminated.
[3]
Although a doubling of datacenter power consumption in just over five years sounds extreme, Rystad's forecast is actually one of the more moderate predictions. In April, Goldman Sachs [4]published [PDF] its take on datacenter power usage and figured that consumption would nearly triple by 2030, with much of that demand being driven by AI.
[5]Supermicro plans to flood market with liquid-cooled datacenter tech
[6]Ohio power plants want special tariffs on datacenters to protect regional grid
[7]Datacenters looking to renewables, nuclear, and gas, in quest for more power
[8]Microsoft's carbon emissions up nearly 30% thanks to AI
The two firms, however, are clearly working with fairly different sets of numbers, since Goldman Sachs estimates current datacenter power consumption to be around 400 TWh while Rystad says it's 130 TWh.
Yet even Goldman Sach's prediction isn't the most extreme we've ever seen. Arm CEO Rene Haas has repeatedly shared his belief that a [9]full quarter of US power will go towards datacenters thanks to the AI revolution. For reference, Goldman Sachs claims a figure of eight percent based on its data. Rystad doesn't offer a similar metric at all.
The International Energy Agency (IEA) is on the same page as Haas and says [10]global electricity usage from datacenters will double by 2026 . Although the IEA doesn't provide data specifically for the US, its report does mention that a third of all datacenters are located in the states, which implies that the IEA expects datacenter power consumption to increase substantially in the US as well as the wider world. ®
Get our [11]Tech Resources
[1] https://www.rystadenergy.com/news/data-and-ev-create-300-twh-increase-us
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2ZnyPia4Rji5FoDfD-oyXAAAAAcE&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZnyPia4Rji5FoDfD-oyXAAAAAcE&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[4] https://www.goldmansachs.com/intelligence/pages/gs-research/generational-growth-ai-data-centers-and-the-coming-us-power-surge/report.pdf
[5] https://www.theregister.com/2024/06/19/supermicro_liquid_cooled_facilities/
[6] https://www.theregister.com/2024/05/21/ohio_datacenter_power/
[7] https://www.theregister.com/2024/05/16/datacenter_power_demands/
[8] https://www.theregister.com/2024/05/16/microsoft_co2_emissions/
[9] https://www.theregister.com/2024/04/09/ai_datacenters_unsustainable/
[10] https://www.theregister.com/2024/01/24/iea_report/
[11] https://whitepapers.theregister.com/
The forecast comes from Rystad Energy in its latest [1]report on US power demand and supply, which includes figures for datacenters in particular. The analysts estimate that datacenter power demand will sit at 307 terawatt hours (TWh), more than double the 130 TWh the firm says was consumed in 2023.
The reasoning behind rapidly increasing power draw is the usual suspect. "The growth in electricity demand for datacenters will be heavily driven by those focused on artificial intelligence (AI)," Rystad says, "which consumes more electricity compared to traditional computing."
[2]
To meet this demand, Rystad says the best energy sources for datacenters will be renewables thanks to their scalability and increasing viability. Consequently, wind energy supply is expected to just about double by 2030, while solar will roughly triple, if the calculations are correct. Power generated by coal will decrease by about half, though won't be totally eliminated.
[3]
Although a doubling of datacenter power consumption in just over five years sounds extreme, Rystad's forecast is actually one of the more moderate predictions. In April, Goldman Sachs [4]published [PDF] its take on datacenter power usage and figured that consumption would nearly triple by 2030, with much of that demand being driven by AI.
[5]Supermicro plans to flood market with liquid-cooled datacenter tech
[6]Ohio power plants want special tariffs on datacenters to protect regional grid
[7]Datacenters looking to renewables, nuclear, and gas, in quest for more power
[8]Microsoft's carbon emissions up nearly 30% thanks to AI
The two firms, however, are clearly working with fairly different sets of numbers, since Goldman Sachs estimates current datacenter power consumption to be around 400 TWh while Rystad says it's 130 TWh.
Yet even Goldman Sach's prediction isn't the most extreme we've ever seen. Arm CEO Rene Haas has repeatedly shared his belief that a [9]full quarter of US power will go towards datacenters thanks to the AI revolution. For reference, Goldman Sachs claims a figure of eight percent based on its data. Rystad doesn't offer a similar metric at all.
The International Energy Agency (IEA) is on the same page as Haas and says [10]global electricity usage from datacenters will double by 2026 . Although the IEA doesn't provide data specifically for the US, its report does mention that a third of all datacenters are located in the states, which implies that the IEA expects datacenter power consumption to increase substantially in the US as well as the wider world. ®
Get our [11]Tech Resources
[1] https://www.rystadenergy.com/news/data-and-ev-create-300-twh-increase-us
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2ZnyPia4Rji5FoDfD-oyXAAAAAcE&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZnyPia4Rji5FoDfD-oyXAAAAAcE&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[4] https://www.goldmansachs.com/intelligence/pages/gs-research/generational-growth-ai-data-centers-and-the-coming-us-power-surge/report.pdf
[5] https://www.theregister.com/2024/06/19/supermicro_liquid_cooled_facilities/
[6] https://www.theregister.com/2024/05/21/ohio_datacenter_power/
[7] https://www.theregister.com/2024/05/16/datacenter_power_demands/
[8] https://www.theregister.com/2024/05/16/microsoft_co2_emissions/
[9] https://www.theregister.com/2024/04/09/ai_datacenters_unsustainable/
[10] https://www.theregister.com/2024/01/24/iea_report/
[11] https://whitepapers.theregister.com/
Once AI consumers start having to pay for the energy they use the curve will drop off
Anonymous Coward
There is 1 ~ 2 year bubble lead time before AI consumers are going to be asked to actually - at a minimum - pay for the energy costs. The logic for the big guys expanding at huge losses now is to ensure a they have a chair when the music stops later.
Could?
ecofeco
Texas has entered the chat.
https://www.houstonchronicle.com/business/energy/article/ai-data-centers-power-demand-19393986.php
chips equals kilowatts
An H100 with infrastructure is probably 2000 watts.
NVIDIA sold "half a million H100s in Q3 2023". Or about 2 million units a year, and rising.
2M x 2kW = 4GW. Jiggawatts Marty, Jiggawatts. And that's just for last year. About half of that will be in the US.
It takes about 10 or 20 years for a baseline-capable power system like a nuke or hydro-dam or coal-plant to be completed after initial start. Call it a 2GWe nuke which is a really big one.
And it costs maybe $20G to build. Of public money mind you, as very few big power plants get built with solely private capital.
Talk about externalizing your costs! OpenAI, Google, Microsoft and AWS are leveraging other people's power systems for their own uses. Yes they pay for the electric but they are also displacing a lot of other entirely valid uses merely because they are bigger and more influential than the masses of other users, like electric car owners, small businesses, even some big businesses.
Some jurisdictions are running out of electric capacity, like Quebec. Some jurisdictions are building extremely expensive base-load systems that won't be online for years. Meanwhile the hyper-intensive datacenters are adding megawatts of demand every week.
When the AI chips have the all the electricity, will the rest of us be stuck with rolling blackouts?
Sources: Toms Hardware (half million). The Logic (Quebec).