People are rightly freaking out about how their electric bills are on the rise at the same time as multi-gigawatt datacenters are popping up like mushrooms, particularly in the United States where the hyperscalers, cloud builders, and AI model builders are largely concentrated.
Even with all of the talk about what is the biggest bottleneck in the next wave of the GenAI revolution – some say it is the availability of electric power, others say it is the various HBM, DRAM, and flash memory supply shortages – the situation may not be as bad as it feels. At least not compared to the amount of power that is being consumed by the 8.3 people on Earth and their employers.
Apropos of the situation, the market researchers at Gartner have just cased the power consumption as well as the power allocation devoted to the world’s datacenters. Estimates vary widely, but there are somewhere between 11,500 and 12,000 very large datacenters in the world and then another 7 million or so enterprise and sovereign datacenters and dataclosets that, all together, host all of the servers, storage, and switching that has created this electronic Pangea that has transformed that world over the past six decades of commercial computing. The number of small datacenters is consolidating as the size of datacenters is exploding, which is a neat thing that we are not focused on at the moment. My best guess is that those big commercial datacenters run by hyperscalers, cloud builders, neoclouds, HPC centers, various sovereigns, and the largest enterprises account for about 75 percent of the juice consumed by datacenters. So all of those small facilities don’t add up to much, really.
Well, unless you have to pay for them. IT pain and budget scales with business size. No one can escape that in the 21st century.
Anyway, here is the Gartner power consumption forecast for datacenters for 2026 and 2027 with a baseline for 2025:
This is great in that it separates out traditional servers from AI servers from the rest of the gear burning juice in the datacenter. You can tease out the 2024 numbers from the growth rates, and some bits of the data are forecast for 2030, which means you can calculate a compound annual growth rate and fill in the gaps between past and future.
Which, as you expect, I did in the table below. I also added some context, gathering up worldwide overall electricity consumption and capacity as well as the same figures for the United States. Most of the data I added comes from the Energy Information Administration, the statistics and analytics division of the US Department of Energy, but some of the forecasts are from Ember, an energy consultancy.
Now * that* is a table. . . .
There is a lot of handwaving right now about what share of electricity consumption and what share of allocated electricity capacity is being sucked up by datacenters. It is true that it is a lot, but consumers burn a hell of a lot of juice and so does manufacturing.
I remember the same concerns after the Dot-Com Boom about how datacenters were consuming so much power, and how the projections into the future – made by Jonathan Koomey – at the time a staff scientist at the Lawrence Berkeley National Laboratory and a professor at Stanford University – were dire. Back in my days at The Register, I covered the original concerns Koomey rightfully raised in the mid-2000s, but I also covered the effect his raising of the issue had on IT infrastructure in 2011. And guess what? The economy slowed thanks to the Great Recession and server footprints didn’t grow as much, and moreover, mainframe-class server virtualization came to the X86 platform and servers got more efficient thanks to the prodding of Koomey and others. And we avoided the doomsday scenario of burning nearly 4 percent of the world’s power on datacenters, it ended up being less than 2 percent. Mind you, that was still higher than the 0.5 percent in 2000 and the 1 percent in 2010.
If we do the math now, the rate of power generation has been more than keeping pace with the rate of change in datacenter power consumption. But as was the case in the 2010s, as the hyperscalers and clouds grew exponentially, the consumption of power in the datacenter is growing faster than the rate of generation in the world, and so the slice the datacenter is taking is getting larger, as you can see in the table above. The power consumption of datacenters has a compound annual growth rate of 20.8 percent between 2024 and 2030, rising from 387 terawatt-hours (TWh) to a forecast of at least 1,200 TWh. That’s 3.1X growth in six years. Worldwide electricity consumption is expected to have a CAGR of only 3 percent over the same time span, growing from 30,856 TWh in 2024 to around 36,500 TWh in 2030. Oddly enough, worldwide generating capacity (express in gigawatts without the time factor) is expected to grow nearly twice as fast, at 5.9 percent CAGR. The capacity allocated to datacenters is anticipated to grow by 18.56 percent CAGR over those seven years, from 104 GW in 2024 to 290 GW in 2030. (Those are Gartner figures. I filled in the blanks with estimates, shown in bold red italics.
When you do the math, the power consumption of datacenters as a share of overall power consumption is on the rise, from 1.3 percent in 2024 to perhaps 3.3 percent in 2030. Ditto for power allocation for datacenters, which was 1.1 percent of global capacity in 2024 and which is expected to grow to 2.1 percent by 2030.
That is, unless we see this data and once again engineer ourselves away from inefficient architectures and towards ones that are perhaps more precisely tuned for AI workloads. We are seeing the beginnings of this shift happening right now, in fact, six years out. Or, maybe we will get nuclear fusion reactors and none of this will matter much. Personally, I like the idea of fusing three helium atoms and making carbon diamonds, but I don’t hear anyone talking about that. Probably because of the three body problem it represents. Maybe we can use AI to solve that? Maybe not.