Data Centers Have Solved Their Speed-to-Power Problem — With Natural Gas

From the perspective of the stock market, you’re either in the AI business or you’re not. If you build the large language models pushing out the frontiers of artificial intelligence, investors love it. If you rent out the chips the large language models train on, investors love it. If you supply the servers that go in the data centers that power the large language models, investors love it. And, of course, if you design the chips themselves, investors love it.

But companies far from the software and semiconductor industry are profiting from this boom as well. One example that’s caught the market’s fancy is Caterpillar, better known for its scale-defying mining and construction equipment, which has become a “secular winner” in the AI boom, writes Bloomberg’s Joe Weisenthal.

Typically construction businesses do well when the overall economy is doing well — that is, they don’t typically take off with a major technological shift like AI. Now, however, Caterpillar has joined the ranks of the “picks and shovels” businesses capitalizing on the AI boom thanks to its gas turbine business, which is helping power OpenAI’s Stargate data center project in Abilene, Texas.

Just one link up the chain is another classic “old economy” business: Williams Companies, the natural gas infrastructure company that controls or has an interest in over 33,000 miles of pipeline and has been around in some form or another since the early 20th century.

Gas pipeline companies are not supposed to be particularly exciting, either. They build large-scale infrastructure. Their ratemaking is overseen by federal regulators. They pay dividends. The last gas pipeline company that got really into digital technology, well, uh, it was Enron.

But Williams’ shares are up around 28% in the past year — more than Caterpillar. That’s in part, due to its investing billions in powering data centers with behind the meter natural gas.

Last week, Williams announced that it would funnel over $3 billion into two data center projects, bringing its total investments in powering AI to $5 billion. This latest bet, the company said, is “to continue to deliver speed-to-market solutions in grid-constrained markets.”

If we stipulate that the turbines made by Caterpillar are powering the AI boom in a way analogous to the chips designed by Nvidia or AMD and fabricated by TSMC, then Williams, by developing behind the meter gas-fired power plants, is something more like a cloud computing provider or data center developer like CoreWeave, except that its facilities house gas turbines, not semiconductors.

The company has “seen the rapid emergence of the need for speed with respect to energy,” Williams Chief Executive Chad Zamarin said on an August earnings call.

And while Williams is not a traditional power plant developer or utility, it knows its way around natural gas. “We understand pipeline capacity,” Zamarin said on a May earnings call. “We obviously build a lot of pipeline and turbine facilities. And so, bringing all the different pieces together into a solution that is ready-made for a customer, I think, has been truly a differentiator.”

Williams is already behind the Socrates project for Meta in Ohio, described in a securities filing as a $1.6 billion project that will provide 400 megawatts of gas-fired power. That project has been “upsized” to $2 billion and 750 megawatts, according to Morgan Stanley analysts.

Meta CEO Mark Zuckerberg has said that “energy constraints” are a more pressing issue for artificial intelligence development than whether the marginal dollar invested is worth it. In other words, Zuckerberg expects to run out of energy before he runs out of projects that are worth pursuing.

That’s great news for anyone in the business of providing power to data centers quickly. The fact that developers seem to have found their answer in the Williamses and Caterpillars of the world, however, calls into question a key pillar of the renewable industry’s case for itself in a time of energy scarcity — that the fastest and cheapest way to get power for data centers is a mix of solar and batteries.

Just about every renewable developer or clean energy expert I’ve spoken to in the past year has pointed to renewables’ fast timeline and low cost to deploy compared to building new gas-fired, grid-scale generation as a reason why utilities and data centers should prefer them, even absent any concerns around greenhouse gas emissions.

“Renewables and battery storage are the lowest-cost form of power generation and capacity,” Next Era chief executive John Ketchum said on an April earnings call. “We can build these projects and get new electrons on the grid in 12 to 18 months.” Ketchum also said that the price of a gas-fired power plant had tripled, meanwhile lead times for turbines are stretching to the early 2030s.

The gas turbine shortage, however, is most severe for large turbines that are built into combined cycle systems for new power plants that serve the grid.

GE Vernova is discussing delivering turbines in 2029 and 2030. While one manufacturer of gas turbines, Mitsubishi Heavy Industries, has announced that it plans to expand its capacity, the industry overall remains capacity constrained.

But according to Morgan Stanley, Williams can set up behind the meter power plants in 18 months. xAI’s Colossus data center in Memphis, which was initially powered by on-site gas turbines, went from signing a lease to training a large language model in about six months.

These behind the meter plants often rely on cheaper, smaller, simple cycle turbines, which generate electricity just from the burning of natural gas, compared to combined cycle systems, which use the waste heat from the gas turbines to run steam turbines and generate more energy. The GE Vernova 7HA combined cycle turbines that utility Duke Energy buys, for instance, range in output from 290 to 430 megawatts. The simple cycle turbines being placed in Ohio for the Meta data center range in output from about 14 megawatts to 23 megawatts.

Simple cycle turbines also tend to be less efficient than the large combined cycle system used for grid-scale natural gas, according to energy analysts at BloombergNEF. The BNEF analysts put the emissions difference at almost 1,400 pounds of carbon per megawatt-hour for the single turbines, compared to just over 800 pounds for combined cycle.

Overall, Williams is under contract to install 6 gigawatts of behind-the-meter power, to be completed by the first half of 2027, Morgan Stanley analysts write. By comparison, a joint venture between GE Vernova, the independent power producer NRG, and the construction company Kiewit to develop combined cycle gas-fired power plants has a timeline that could stretch into 2032.

The Williams projects will pencil out on their own, the company says, but they have an obvious auxiliary benefit: more demand for natural gas.

Williams’ former chief executive, Alan Armstrong, told investors in a May earnings call that he was “encouraged” by the “indirect business we are seeing on our gas transmission systems,” i.e. how increased natural gas consumption benefits the company’s traditional pipeline business.

Wall Street has duly rewarded Williams for its aggressive moves.

Morgan Stanley analysts boosted their price target for the stock from $70 to $83 after last week’s $3 billion announcement, saying in a note to clients that the company has “shifted from an underappreciated value (impaired terminal value of existing assets) to underappreciated growth (accelerating project pipeline) story.” Mizuho Securities also boosted its price target from $67 to $72, with analyst Gabriel Moreen telling clients that Williams “continues to raise the bar on the scope and potential benefits.”

But at the same time, Moreen notes, “the announcement also likely enhances some investor skepticism around WMB pushing further into direct power generation and, to a lesser extent, prioritizing growth (and growth capex) at the expense of near-term free cash flow and balance sheet.”

In other words, the pipeline business is just like everyone else — torn between prudence in a time of vertiginous economic shifts and wanting to go all-in on the AI boom.

Williams seems to have decided on the latter. “We will be a big beneficiary of the fast rising data center power load,” Armstrong said.