AI-spurred surge in data-center development could have climate consequences

The artificial-intelligence boom could significantly hinder the fight against planet-altering climate change.

The large and growing demand for computing power to train and run the models that underlie chatbots such as OpenAI’s ChatGPT and Anthropic’s Claude is prompting a wave of data-center construction. Demand for electricity to power those data centers is in turn leading to new power-plant development. 

But while tech companies in the past have made pledges to reduce their carbon footprints and have touted the potential of nuclear power to provide needed data-center electricity, many of those developing data centers right now are turning to natural gas.

Just in the past year, the generating capacity of natural-gas plants in development in the United States has tripled. That amount is now equivalent to half the size of the current national fleet of such plants, according to Global Energy Monitor, an industry-research firm.  

“You’ve just seen an absolute surge in gas power … proposals in the U.S. in the last year,” said Jenny Martos, a research analyst for the firm who focuses on the natural-gas industry.

There’s a lot of uncertainty about whether those plants will be built at all, and — if they are — the extent to which they would be used. But assuming they are built out completely, they could boost national carbon emissions from power plants by more than 60% above current levels. 

Over the course of 50 years, the emissions from those new plants alone could result in the planet warming by about 0.024 degrees Celsius — a significant amount when climate scientists are pushing to limit worldwide warming to 1.5 degrees above pre-industrial levels. 

The plants could also be expected to generate significant amounts of other pollutants, which result in deaths and illnesses, according to Mark Jacobson, a professor of civil and environmental engineering at Stanford University who focuses on energy policies.

Building gas plants is “bad,” Jacobson said. “It’s bad for climate. It’s bad for air quality. It’s bad for the environment, in terms of land use.”

To a large extent, the artificial-intelligence boom is centered on San Francisco and the wider Bay Area. OpenAI and Anthropic, the two leading private AI-model developers, are based in The City, as are numerous smaller startups in the industry.

Other leading model developers — including Google, Meta and Elon Musk-owned xAI — are just down the Peninsula, as is Nvidia, which provides the high-powered chips the companies are largely using to train and run their systems. 

With the promise of transforming major portions of the economy, replacing older technologies and potentially even human workers, model developers have been in a race to build ever more capable AI models. That’s prompted the construction of new and more powerful data centers to train and run those systems. 

A gigawatt is 1 billion watts. A typical LED light bulb requires about 10 watts, so the power required by the current data-center fleet would power about 1.4 billion light bulbs. The power needed by those in development would power 18.7 billion.

Data-center use, particularly for AI, is already driving electricity demand. Last year, Lawrence Berkeley National Laboratory forecast that between 2023 and 2028, U.S. electricity usage consumed by data centers could jump from 176 terawatt hours — or 4.4% of total U.S. electricity usage — to 580 terawatt hours, or 12%.

Natural-gas plants with a collective generating capacity of 252 gigawatts are now in development — meaning they’ve been announced, are undergoing pre-construction work or are already being constructed, according to Global Energy Monitor. 

In comparison, all the gas plants in operation at the end of 2024 had a collective capacity of 506 gigawatts, according to the U.S. Energy Information Administration. And just a year earlier, the gas plants in development represented only 85 gigawatts of capacity, according to Global Energy Monitor.

Much of that surge in gas-plant development is related to data-center demand. In fact, plants that would provide on-site power to data centers represented 100 gigawatts of the in-development facilities, Global Energy Monitor reported. Other such plants would be built off-site by utilities, but their energy would still go to powering data centers.

Some of those plants might not be built. The spike in demand for such plants has resulted in a backlog of orders for the turbines that would power them. As of last May, that backlog had reached up to seven years, S&P Global reported.

Of the gas-plant projects in development in the U.S., 58% haven’t identified a turbine manufacturer, Martos said. What’s more, the projected completion dates for many of these projects are being pushed out over the next five years, she said.

“If you don’t have a turbine, you’re going to see some project delays, potential cancellations,” Martos said. “Some of these projects will not move forward.” 

Beyond supply-chain problems, the projects also face uncertain long-term demand, said Alex de Vries-Gao, a Ph.D. candidate at Vrije Universiteit Amsterdam. While AI is driving data-center construction and power usage now, that might not be the case in a few years, he said. 

Much of the investment in AI is being driven by hype and is resulting in huge losses at OpenAI and other companies. At some point, that hype bubble could burst, dramatically stemming such investments. That too could limit the number of gas plants that end up being built. 

“It’s a bit of a bit of a risky industry,” said de Vries, who is also the founder of Digiconomist, a research firm that focuses on the environmental consequences of new technologies. “There’s no guarantee of long-term demand at the moment.”

All that said, the AI companies, data-center operators and utilities are pushing forward with plans for new plants. Should those plants get built, they will have sizable environmental consequences. 

Were they to be run 24 hours a day every day of the year, plants with 252 gigawatts of capacity would generate 2.2 million gigawatt hours of electricity annually. On average, gas plants generate about 439 metric tons of carbon dioxide per gigawatt hour, according to data from the US EIA.

That means that over the course of the year, if those plants were running at full capacity, they’d generate about 969 million metric tons of carbon dioxide. Over the course of 50 years, they’d generate 48.5 billion metric tons of carbon dioxide.

All else being equal, each gigaton — a billion metric tons — of carbon dioxide emitted into the atmosphere can be expected to warm the planet by 0.0005 degrees, according to Michael Mann, a leading climate-change expert and the director of the Penn Center for Science, Sustainability and the Media and the University of Pennsylvania. 

So, over the course of 50 years, those gas plants that are now in development — if they ran at full capacity the entire time — would be expected to warm the planet by 0.024 degrees Celsius. 

To be sure, the plants almost certainly won’t run that much or potentially for that long, even if they are all built out. On average, gas plants of different types across the U.S. ran about 10 hours per day in 2024, according to U.S. EIA data. Meanwhile, the typical plant has an expected service life of about 30 years, according to Martos.

Factoring in that kind of utilization rate and service life, the plants would contribute about 0.006 degrees Celsius of warming over their lifespans. However, given that the plans in many cases would be used to directly power data centers — which could be expected to operate in many cases around the clock — that likely understates how much power they will generate, Martos said. 

Regardless, any additional warming is too much, Mann said. Scientists like himself have been urging world governments to limit total global warming to 1.5 degrees Celsius to avert the worst effects of climate change. 

“At a time when we need to be reducing emissions dramatically, building out natural gas infrastructure is taking us in the opposite direction,” Mann said in an email. 

It’s also unnecessary, he and Jacobson argue. Renewable-energy sources are less expensive to develop and can be faster to put in place, they said. 

If data-center providers and their utility partners decide it’s necessary to ramp up power on the grid to accommodate new data centers, the fastest and cheapest thing they could do would be to pay a bunch of homeowners or other building owners to deploy rooftop solar systems with battery backups, Jacobson said. That could be done in six months and could free up existing utility-scale power to handle new data-center loads, he said. 

The technology is proven, and it’s already worked at scale in California to reduce on-grid power demand, he said. 

Much of the jump in gas-plant development is happening in Texas, according to Global Energy Monitor. That’s no surprise, according to Jacobson — in recent decades, California has promoted energy efficiency in appliances and via building codes, and it has helped spur the adoption of rooftop solar and battery-storage systems. Texas is well behind in all of that, he said. 

The nation could accommodate new electricity demand in sustainable ways, Jacobson said. 

“You just have to have strong policies,” he said.

“Gas is only needed when you have weak policies,” he said.

If you have a tip about tech, startups or the venture industry, contact Troy Wolverton at twolverton@sfexaminer.com or via text or Signal at (415) 515-5594.