Mega-tech operators are in pursuit of land and power generation to supply the stampede of data center facilities set to be built across the United States, which may translate into higher natural gas consumption.
The forecasts for new power generation to fuel hyperscale data centers are all over the map. Many operators are eyeing underutilized capacity already on the grid. Some are clinching power purchase agreements to rely on renewables and alternative energies. And natural gas operators are sure to gain a piece of the market, according to Enverus energy transition analyst Carson Kearl.
Kearl, who recently completed an in-depth forecast on data center power demand, discussed his findings during an interview on NGI’s Hub & Flow podcast.
“The easy way to think about a data center is that it's just a really really big computer,” Kearl said. “When you put all of those computers together, it enables infrastructure for the internet and all of the applications we use to operate on a day-to-day basis.”
There are “roughly 2,000 facilities in the Lower 48.” Combined, the facilities today have “rated capacity of roughly 20 GW of power.”
Data generated and consumed has grown at a 33% compound annual growth rate since 2010. Enverus used that figure as a proxy to extrapolate the forecast increase in power demand.
“Our expectation is that trend breaking to the upside, with our low case representing the highest growth rate of 48% a year,” Kearl said.
The “Magnificent Seven,” or the biggest tech companies according to Wall Street, are driving the surge in artificial intelligence (AI) growth. Apple, Alphabet, Amazon, Meta, Microsoft, Nvidia and Tesla join a plethora of large and small operators, including Netflix, which would use smaller-scale centers to store data and for training, Kearl noted.
Open 24-7
Every data center would need generation to power up 24 hours a day, seven days a week. What could that mean for natural gas? Around 7 GW of load is equivalent to an incremental 1 Bcf of natural gas, Kearl noted.
In the Enverus base case, installed data center capacity grows by 14 GW from 2023 to 2030. That could equate to a gain of up to 2 Bcf/d of gas-fired generation – if every facility were completely powered by natural gas. By 2035, the base case may nearly double, and gas use could increase to 4.2 Bcf/d.
The base case “is essentially setting the stage for what you could imagine would happen in a world before AI,” Kearl told NGI. “So that's kind of the trajectory that we were on in terms of the growth out of the facilities in this space.” The base case “is what we think is the most reasonable given the infrastructure constraints.”
In the Enverus low case, there would be “limited growth in computing” based on the historical five-year average. That reduces the load growth estimate to 8 GW, or a 0.9 Bcf/d increase in implied gas demand. That scenario is “unlikely and a conservative floor value,” according to Kearl and his team.
As a caution, though, power demand may not translate one-to-one into net new load growth.
“There are a lot of reasons that incremental load might not be served by gas-fired generation,” Kearl said. “We like to highlight the idea of substitution.
“There are a lot of underutilized industrial facilities across the Lower 48 that have the site, and a lot of the characteristics that a new data center builder would like in a site, as well as the existing interconnect. So we do foresee a future where some of those facilities,” fired by gas, coal, renewables or other sources, already would be available.
And renewables generation would continue to take a bigger share. For the big tech companies geared to sharply reducing emissions, that could reduce how much natural gas is used.
“We expect an average of 50 GW of annual wind and solar additions by the end of the decade that will compete with gas for existing load from marginal power consumers,” Kearl said.
Smaller Gas Opportunities?
While many are tripping over the coming advent of hyperscale projects, the smaller ventures could be advantageous for natural gas consumption, Kearl noted.
“It's really important to think about these facilities in sort of two different frames,” he said. “You have an ‘inference’ data center, which is your consumer-facing application. That's going to be a relatively flat load profile and that's going to be the majority – probably 80% to 85% of the facilities built…”
The hyperscale facilities need 18 to 24 months to enter the load interconnection queue online, he added.
“You also have training data centers” composed of graphics processing units, or GPU clusters. They are “essentially small data centers for training or other purposes,” Kearl noted. “They're able to be built faster…
“From that point of view, there is probably a lot more opportunity for natural gas behind the meter sitting on the training facilities” than for the hyperscale projects.
Expected infrastructure constraints are likely to cause “pain for a lot of these providers,” Kearl said. “The utilities aren’t necessarily the most welcoming to those large companies with these large loads that have high reliability requirements.”
As most of the new load is expected to be “grid connected,” that makes it “really hard to say where the electrons are coming from.”
As an example, Amazon Web Services (AWS) in March acquired Talen Energy Corp.’s Cumulus data center campus at the Susquehanna nuclear power station in Pennsylvania. Talen in 2023 had completed the first 48 MW, 300,000 square-foot hyperscale facility at the site. AWS is aiming to develop a 960 MW campus.
The AWS-Talen facility is “an example of the ideal situation for one of these data center builders,” Kearl noted. “They partner with this nuclear facility and they get guaranteed power. They're grid connected. And it's fair enough to say that that's where the electricity is coming from.”
Financing? No Problem
The capital cost to build data centers that would add 70 GW of capacity would be “close to $700 billion…between now and 2028,” Kearl said. “From a market cap perspective, you could buy pretty much every energy company on the S&P 500.”
Kearl said “there is going to be a relatively positive buildout” of capacity in Georgia, Texas, Virginia and potentially Florida. California and Arizona also are being eyed to the extent that the grid would allow.
As technology options expand at a remarkable pace, the “million dollar question” is how efficiencies would reduce the need for power generation.
“I spoke with a rep from Nvidia a few weeks ago and it seems extremely apparent to me that they don't believe that their ability to improve the efficiency of these chips is declining,” Kearl said. “They will be able to continue to pump out better and more efficient designs through time, which leads to a view that the growth doesn't necessarily need to happen as fast as some are saying.”
Additionally, the existing data centers are being upgraded “on roughly a three-year cycle.” That would allow the providers to “squeeze more energy out of the existing facilities…Then you may only be applying for a marginal increase in the power you're allowed to consume rather than trying to structure an entire new facility.”
It’s “equally dependent on the commercial viability of a lot of these AI technologies,” Kearl said. ChatGPT and LlamaGPT are “very cool and they're very fun, but the commercial use cases in a lot of situations, and for a lot of organizations, aren't necessarily apparent yet. There needs to be an uptick in demand from the market to support the growth that a lot of those organizations are calling for.”
