The Gigabucks Going Into Datacenter Gigawatts

The only thing that takes longer to bring online slower than a datacenter is a chip foundry, which is unfortunate for a lot of different reasons. But for those who want to invest in either, given the explosion of interest in GenAI and the enormous funds that companies of all sizes and stripes are willing to spend on it, the return on investment for AI datacenters is higher and quicker, and a whole lot less risky.

Which is why datacenter spending is absolutely booming this year and into the foreseeable future. This datacenter buildout is the best leading indicator of GenAI spending, or perhaps more accurately, the exuberance with GenAI spending is leading to the enthusiasm for building datacenters.

The two are interlinked for sure, and generally speaking, the construction of the datacenter plus its, powering, cooling, and cost of operation over four years is roughly half of the total cost of creating an at-scale AI system. The datacenter infrastructure – meaning the building and its power distribution and cooling systems – is probably on the order of 25 percent to 30 percent of the total cost, with the power and cooling being somewhere between 20 percent and 25 percent of the pie. So, for a shorthand, take the investments in datacenters – again, meaning the physical plant – and multiply it by 2X and you get the investment is the AI systems that will fill it; the same amount of the datacenter cost is four years of juice to power and cool it.

So, let’s look at the datacenter buildout underway in terms of total power capacity and money and see how that might translate into datacenter equipment sales.

The hyperscalers and large cloud builders are driving the majority of this capacity, but the so-called “neoclouds,” which are specialists who have done cryptocurrency mining and are now shifting to AI processing because there is more money in this, are on the rise, too.

According to Synergy Research Group, the hyperscalers and cloud builders – what Synergy calls simply “the hyperscalers” – doubled the number of their datacenters in the past five years, and now they are entering a phase of datacenter and region expansion as well as building out some new facilities. Synergy reckons that the hyperscalers and cloud builders had 1,103 major datacenters around the world as 2024 came to a close, and they had a known pipeline of another 497 facilities that were planned between 2025 and 2030.

Here is what the curve looks like for datacenter counts (left axis) and datacenter IT capacity (right axis, expressed in thousands of megawatts) from 2021 through 2030:

John Dinsdale, chief analyst at Synergy, says that there is a lot of flux between old and new datacenters, old and new regions, and owned versus leased facilities, but in general, over the next five years Synergy expects for the GenAI boom to drive a doubling of datacenter capacity at hyperscalers and cloud builders by 2030. GPU systems, of course, dominate those capacity investments.

The analysts at JLL Research have put out their own estimates of datacenter capacity out into the future, and here is what it looks like:

As expressed in gigawatts, the total datacenter capacity is growing at a compound annual growth rate of 15 percent between 2023 and 2027, inclusive. JLL says that the hyperscalers, large cloud builders, and co-location facilities will bring around 7 gigawatts of capacity online this year that was already under construction and break ground on another 10 gigawatts this year. All told, this represents $170 billion in asset value, says JLL, which will need financing from somewhere. There will be no shortage of private equity money to chase this, we think, as well as more traditional commercial real estate financing from banks that want to get in on the action.

Considering that 1 gigawatt can power around 800,000 homes in the United States, this additional datacenter capacity of 17 gigawatts is equivalent to 13.6 million homes. JLL estimates that the datacenters of the world burned around 50 gigawatts in 2024, double that of 2020, and that it will double again to 100 gigawatts by 2029. That said, datacenters will consume only about 2 percent of the electricity generated globally in 2025. Consumption varies by region, of course. In Ireland, datacenters consume 21 percent of the country’s power, for instance, according to JLL, and in Virginia, the datacenter capital of the world west of Washington DC, datacenters consume 26 percent of the state’s electrical power.

There’s Gold In Them There Electric Bills

Much has been made of the merger and acquisition activity for datacenter builders and operators in 2024, which added up to $73 billion. But if you look in the chart below, the spike in 2024 was just filling in the whole in 2023, which fell by 50 percent compared to 2022 to a mere $26 billion. We wonder if the lack of availability of GPUs from Nvidia actually stalled datacenter mergers and acquisitions a bit, but nonetheless, in 2021 and 2022 we saw four major datacenter operators acquired. They include:

  • Blackstone acquiring QTS Realty Trust for $10 billion in July 2021
  • KKR and Global Infrastructure Partners buying CyrusOne for $15 billion in November 2021.
  • American Tower acquiring CoreSite for $10.1 billion in November 2021
  • DataBridge Group buying Switch for $11 billion in May 2022

The big deal in 2024, which happened as the year was coming to an end, was the $15 billion acquisition of AirTrunk, a datacenter operator operating largely in the Asia/Pacific region and based in Sydney, Australia, by Blackstone. The other big deals last year were two equity investments in Vantage Data Centers for $9.2 billion for its US datacenters and another two deals for $3.1 billion combined for its European datacenters.

In that AirTrunk acquisition, Blackstone said that on the order of $1 trillion of datacenter capital expenditures would be spent in the United States over the next five years (presumably 2025 through 2029 inclusive), and that another $1 trillion would be spent outside of the United States. (By the way, Blackstone also owns stakes in traditional datacenter operator Digital Realty and GPU neocloud CoreWeave.)

The world spends $2.4 trillion a year on military defense, to put that into perspective, which is six times the rate that is being spent on datacenters and the gear that goes into them. One could argue that a big chunk of that datacenter is military in nature, either directly or indirectly.

Not at all surprisingly, private equity firms are driving more and more of the M&A activity for datacenter operators. Back in 2024, according to Synergy, private equity firms only accounted for 54 percent of the dough spent on acquiring datacenter operators; by 2021, the PE share rose to 65 percent, and since then it has ranged between 80 percent and 90 percent of the total money invested to acquire and expand these behemoth bit barn operators.

Looking more closely into this year, Synergy says that $7 billion in mergers and acquisitions deals have been inked so far in 2025, and another $15 billion of deals have the nod and the paperwork is progressing towards ink. Synergy has heard about another $20 billion in deals where operators are looking for funding to scale their datacenters.

Sign up to our Newsletter

Featuring highlights, analysis, and stories from the week directly from us to your inbox with nothing in between.
Subscribe now

8 Comments

  1. Hmmm it would be interesting to know what Doc Brown would say if Marty fired up the Delorean / Flux capacitor to head back to 1955 from 2025 (instead of 1985), and told Doc that modern data centers needed 1.21 GigaWatts in order to function.

    …Ah if only lightning could be convinced to fire 24×7 above a data center!

  2. I’m assuming that the pic at the top of the article represents the 2025 state-of-the-art data center BoPoD (Buildinglets on Pad on Dirt) packaging. And the buildinglets are perhaps at the current reticle limit.

    It appears that the the center “IO” buildinglet is connected to the outer 4 “compute” buildinglets with thin (but probably hit bit rate) serial interfaces.

    Obviously quite a few “capacitors” placed around the periphery of the package.

    Interesting.

  3. The picture at the top of the article is Meta’s $10B Richland Parish datacenter in Louisiana. This datacenter will consume 2 gigawatts. At 8 cents/kwatt-hour, which is the average industrial electricity price in the United States, the electricity to power this one datacenter is $1.4B per year. Microsoft’s Stargate datacenter will supposedly cost $100B, so imagine multiple sites with a total size about 10x bigger than what is in the picture at the top of the article and an even higher electricity bill.

    It would be great if some use could be found for the waste heat produced by these datacenters. Aquaculture or some manufacturing facility could be co-located with the datacenter so that the waste heat can be put to productive use. In a cold area like Canada or a Nordic country, there could be condos on top of or surrounding the datacenter. The waste heat from the datacenter could be used to heat the condos, provide warm water and melt snow.

    Instead of just optimizing the datacenter by itself, the bigger picture of local needs and resources (farm waste, livestock waste, …) should be examined so that a more holistic design can be made. For example, livestock waste can be use to grow cattails and cattails can be converted into ethanol using the waste heat from the datacenter. If the datacenter is built below ground in a basement, different types and sizes of facilities could be built above it and could change over time, as uses for the waste heat change. To get the datacenter online as quickly as possible, cooling towers could be put above the datacenter initially. The cooling towers could be swapped out as different uses for the waste heat are brought online.

  4. As usual , love the reporting and POV. Next perhaps is looking at how we deliver this generation capacity and move it thru the grid – if at all

Leave a Reply

Your email address will not be published.


*


This site uses Akismet to reduce spam. Learn how your comment data is processed.