Q&A with Climate Scientist David Victor, Part 2: AI Data Centers and Resources

In Part 2 of this three-part series, we chat with David Victor, a member of Dimensional’s network of ESG researchers and academics, about how the data centers used to power artificial intelligence are affecting energy demand, emissions, and water usage.¹ Victor is a professor of innovation and public policy at UC San Diego, director of the campus-wide Deep Decarbonization Initiative, and an adjunct professor of climate, atmospheric science, and physical oceanography at the Scripps Institution of Oceanography. He’s also part of the Washington, DC-based think tank the Brookings Institution.

Dimensional: Earlier this year we spoke with Michael Gillenwater, another member of our network of ESG researchers and academics, about AI data centers.²  We talked about how the current growth in energy demand due to data centers is a bit different from prior spikes in demand. Data centers are concentrated in specific locations, and demand is very choppy, going from “0 to 60” the minute a data center comes online. What have you observed about the increase in energy demand from data centers?

Victor: An interesting feature is that data centers want power fast. The business model in the data center world is first to power—these companies are in an arms race with each other to get their AI into a variety of applications. So what they care about is power and how to get that quickly. Typically the cheapest and fastest way to get electricity in most of the US is from natural-gas-fired generators, so many data centers are gas fired.

Dimensional: Is the increase in energy demand from data centers high enough to cause an increase in overall emissions? In the US, total emissions decreased in 2023 and 2024 but then grew a little in 2025.³

Victor: Data centers are certainly a big source of growth in electricity demand in the US; it’s approaching 5% of total power demand and rising very steeply. But the idea that data centers will cause all the growth in emissions is, to me, a degree of hyperventilation. In terms of total emissions, data centers currently make up a similar level to airlines: about 4% or 5% of total emissions in the US.

It is important to consider where the electricity for data centers is likely to come from. The cheapest and fastest way to get electricity in most of the US is from natural-gas-fired generators. But buying renewables plus batteries is also a fast way to get power. The companies driving data centers are paranoid about falling behind in the AI race, and they are willing to spend a lot for power—including clean power, if they can get it in a timely way. But speed is paramount, which is why so many data centers turn to natural gas. In the US, gas is widely available and relatively inexpensive. There have been some slowdowns in the natural gas turbine supply chains, but those problems are getting ironed out.

This demand for power comes at a time when the clean-tech revolution in the US is still alive and well. In 2024 and 2025, we saw continued investment in electrified heat, clean shipping, electrified transport, nuclear, carbon capture, renewable energy, etc.⁴ Investments have tended to be in technologies that are already mature, including technologies that could power data centers.

We’ve seen a number of data center companies sign memorandums of understanding for small modular reactors—these would be novel nuclear energy supply projects. But so far those are just press releases. Data centers are in a race for reliable power, so they can’t wait for first-of-a-kind energy supply projects to be built. When small reactors become truly available and reliable with known costs, data centers will be eager to buy.

I don’t see this as transforming the emissions picture quite yet, but it has started to transform the power sector, in particular the grid.

Dimensional: Can you expand on how it has started to transform the grid?

Victor: The moment a business—like a data center—thinks it might need more power, it files an interconnection request with the grid, and then the grid operator needs to figure out what upgrades and investments are needed for the grid to accommodate these new interconnections. The problem is that with data centers, almost all interconnection requests are totally speculative, and they may or may not end up needing the power they had requested. This dynamic makes it extremely hard to plan the grid. In the old-fashioned world—like five years ago—someone who filed an interconnection request intended to interconnect. In today’s world, we need a way to separate the wheat from the chaff by, for example, imposing higher filing fees and escrow requirements on entities that are filing interconnection requests. Those reforms are underway in many parts of the grid, but they are taking time to implement.

A key political question is who is going to pay for this grid infrastructure. Fulfilling these interconnection requests will raise the cost of electricity in a time when affordability is a key political concern and when electricity prices across the US have been generally rising.

In theory, data centers could be a boon to political forces worried about the rising cost of electricity. If we interconnect large loads—in other words, customers that demand a lot of electricity—to the grid to make those loads responsive to the conditions of the grid, the cost of electricity will actually go down. In effect, these new power loads are taking a very expensive capital asset, which is the grid, and amortizing that asset over a much larger throughput of electricity. That’s the hope and the promise, but it requires a very responsive grid. We haven’t yet seen that play out in reality.

Dimensional: Can you expand on your point about the responsiveness of the grid?

Victor: Responsiveness is how quickly the grid can adapt to sudden shifts in supply and demand. The grid has to match demand for electricity with supply in real time. Some sources of power are more responsive than others—natural gas can be turned on quickly, whereas sources like solar and wind are more variable in their output. Solar power is only produced when there is sunlight and wind power when there is wind.

Data centers are hugely interested in the responsiveness of the grid. They are aware that the cost of electricity is a core expense for the company and also that the question of who is going to pay for the increase in demand for electricity is a core political issue for the communities they’re located in. That has created very powerful incentives for data centers to adjust their demand for electricity to avoid causing spikes and strains on the grid during peak times. Data centers want to figure out which compute requirements are high priority and low priority and move those around during the day, if they can.

For example, if someone asks AI to draw them a picture of two cats dancing on a table, that’s a compute requirement that has to happen right then. In contrast with those on-demand compute needs, there are many compute requirements that can be staggered over the day or even over multiple-day periods. On top of this, we are starting to see more pressure from AI companies to make sure that other loads on the grid, such as power-hungry industrial loads, become more responsive. People have been talking about a highly responsive grid for a long time; the massive power needs of AI are creating a lot of pressure to deliver.

I should note that most AI-related compute loads are in the United States, which is, for the most part, power constrained. By contrast, new AI-related loads in China seem to be having many fewer problems getting access to power. Up to a point, access to power has become a strategic asset for the AI revolution.

Dimensional: Data centers have also been in the news for their water usage. Could you put the data center demand for water into perspective for us?

Victor: Water is one of those topics that once it’s in the news, it stays in the news. I look pretty closely at this and don’t see a significant issue related to water consumption at all. Water is used to cool data centers in open-loop evaporative cooling systems. But the big new data centers are typically closed-loop systems that continuously recirculate the water in various ways, use the heat for various purposes, and do not consume significant amounts of water.

Interestingly, water pricing and water allocation in the US have historically favored agriculture. For example, 79% of the water drawn from the Colorado River goes to agriculture. The result is that water is cheaper for agricultural operations compared to residential and industrial uses, which can lead to overconsumption. If we start getting serious about pricing water, the data centers won’t be the place where we see big changes.

Dimensional: Any final thoughts on the topic of AI that you’d like to leave us with?

Victor: It is still an open question as to whether AI is going to be biased in the direction of renewable energy or fossil fuels. A lot of the AI investments are going into cleaner technologies, which suggests that AI will fuel innovation in green technologies. A lot of the skills from an AI system allow you to run a much more complicated energy system—in particular, grids with lots of renewables and much lower emissions while keeping reliability high. At the same time, people forget that the same kinds of innovations in AI make it easier to run a complex refinery, make it easier to find and produce oil in complex oil fields, make it easier to produce coal, and so on. There is a great race underway in which clean has a slight advantage, but it’s not clear how that’s going to play out.

Stay tuned for Part 3, where we will explore how new technologies may have an impact on the pace of climate change.