There’s one story that’s dominated the headlines and created a huge surge of client interest lately, and that’s the conflux of AI and data centers. Everyone is salivating: The AI evangelists are ready to spend hundreds of billions of dollars on infrastructure, much of which will flow to power generation, and everyone from oil and gas majors to small-time renewables providers wants a piece of the action. The hype is fueled from the top, with President Donald Trump ripping up the Biden administration’s provisional rules on AI safety. This hype has led to a commensurate level of misinformation mythmaking and bad strategy. We’re in an AI bubble — something I’ve written about in the past and will be discussing at length at our Houston Forum — but that doesn’t mean energy companies are doomed to lose money. Data centers can be a very attractive business as long as you don’t fall for these three most common myths:
- Myth #1: AI will be the biggest driver of data center power demand. Energy demand for data centers is growing quickly, but traditional data center use-cases — streaming, cloud services, data management — not only make up the largest share but also will account for the largest portion of growth through 2026. AI energy use isn’t exactly small, and its percentage growth is high, but dedicated AI data centers will still consume less energy than cryptocurrency mining in 2026. There’s a huge push for AI energy efficiency right now, and leading models are dramatically cutting costs in both training and usage energy, whereas something like bitcoin mining (given a huge boost by Trump) gets algorithmically less efficient over time. The challenge with AI data center demand is more concentration: In specific regions like Northern Virginia, projected data center demand dominates all other forms of growth. But for energy providers, focusing exclusively on AI demand misses the bigger picture.
- Myth #2: Small nuclear reactors (SMRs) will power data centers. Data center energy demand has caused some existing nuclear reactors to be brought back online; this, combined with the hype around SMRs, has stoked a lot of excitement about the potential of SMRs for data center power. On paper, there’s a plausible argument: Data centers need reliable power and are willing to pay a premium; nuclear is the most reliable baseload energy around but needs to charge a premium to make economic sense. Despite the narrative, the numbers just don’t add up. My colleagues Anirudh Bhoopalam and Karthik Subramanian put together a brilliant cost-modeling exercise on low-carbon power for data centers. First-of-a-kind SMRs would cost almost triple what natural gas turbines would (a minimum of USD 331/MWh vs. a maximum of USD 124/MWh) in the best-case scenario. A more realistic estimate, factoring in cost overruns and delays, has SMRs costing over 10 times more than natural gas. Series production of SMRs is necessary for the economics to pencil out, and we don’t expect first-of-a-kind SMRs to come online before 2035 at the earliest. Cheap nuclear just isn’t in the cards in the next two decades.
- Myth #3: Renewables are the only route to low-carbon data centers. Tech companies have been quick to sign power purchase agreements with renewables providers to lock in low-carbon power for their data centers. But natural gas with carbon capture is actually cost competitive with renewables around the world. Even in gas-starved Germany, gas turbines with carbon capture deliver electricity at a levelized cost of USD 147/MWh, competitive with the top end of our cost estimate for wind and solar (USD 152/MWh). Renewables represent the lowest possible cost in Germany, but they may not be available to every developer. In most other regions, gas is cheaper. It’s the lowest-cost option in Virginia, where data center construction is booming, and it (perhaps obviously) dominates in the Middle East, where countries like the United Arab Emirates have been making a major push into AI. Renewables have been more successful at this point than carbon capture, but we have a very good test case coming up with ExxonMobil’s 1.5-GW natural gas power plant and combined carbon capture facility. If this approach proves profitable and scalable (and the carbon capture portion actually works), it is expected to be replicated across the U.S. and globally.
I’ve written many times that we’re in an AI bubble. I do expect an AI crash and a lot of big companies like OpenAI to hit rough seas. But that doesn’t mean that data center demand isn’t going to grow or that data centers are necessarily a bad investment. In the late 1990s and early 2000s, the dot-com bubble led to a huge amount of investment in internet infrastructure: the fiber-optic backbone that AI is now being built on. Many of the companies in that bubble crashed out, but the infrastructure proved crucial to the expansion of social media and the modern internet in the 2000s. In a similar vein, there’s real money to be made by providing power to data centers, especially if you take care to insulate yourself from a potential downturn in the AI market. It’s not just about providing power either — the carbon that gets captured has to go somewhere. Storage is one option, but utilization is generally preferable: Cement companies could have a real opportunity to leverage these CO2 streams for mineral carbonation. The energy of AI will be front and center at our April Forum in Houston — so, if you want to make money from the data center boom, you’ll want to be there!
