Sign Up for News Updates
It’s hard to read any energy news these days without bumping into a story about how AI data centers and crypto mines are about to take down the national grid. The demand projections are “creating chaos in utility planning,” according to one source. Alarmists are worrying out loud about everything from grid capacity to grid fragility. On the latter note, there is no question that the grid is aging, much of it having been built more than 50 years ago. Some 30% of U.S. transmission and 46% of distribution assets are beyond, or close to, their useful life.
The situation has me reflecting on a big shift taking place in the energy transition conversation. The ink allocated by the press to climate change for a full decade or more is now flowing into a power shortage narrative. At the macro level, the competing forces of supply and demand frame the story. At the ground level, on the other hand, the situation sounds like this: As AI swallows up entire zip codes of supply all over the country, the average homeowner is watching their monthly electric bill go through the roof and wondering whether to pay for heating or eating.
EIA’s latest data confirms the trend. On average, residential retail prices for a kWH increased 5.2% over last July and power prices for the transportation sector have popped by 8.5% during the same period. Should we blame the price hikes all on data centers? The math, as the saying goes, “is mathing” in that direction. Bloomberg reports that wholesale electricity costs as much as 267% more than it did five years ago in areas near data centers, and that the increases are being passed directly onto ratepayers.
It’s easy to look into the future and see trainloads of angry consumers coming down the tracks when they realize they are subsidizing Amazon, Google, Meta, and OpenAI, among others.
Will Demand Weaken?
It would be nice if a way existed to quickly bring balance to the supply vs. demand curve in play right now, but Lawrence Berkeley National Laboratory said in a recent report that by 2030, data centers are forecast to consume between 325 TWh and 580 TWh of electricity per year, equivalent to 6.7% to 12% of all U.S. consumption.
Likewise, the Wall Street Journal recently reported on the oceans of money flowing into AI data center construction. One of the more striking quotes was from Meta’s CEO, Mark Zuckerberg, who said the company’s U.S. spending through 2028 was “probably going to be something like $600 billion.” Meta, along with Alphabet, Microsoft, and Amazon, are expected to spend about $400 billion in 2026 alone on data centers.
My view of power demand is that it’s big but won’t be as massive as is currently projected. The enormous projections for power simply cannot be met without an extraordinary alignment of capital investment, grid investment, regulatory streamlining, low energy prices, and an easing of supply constraints.
Will Supply Strengthen?
On the supply side of the power shortage challenge is the need for more generation. Small modular nuclear reactors (SMNRs) have many industry thinkers excited. The technology appears to be great but time-to-compute(the amount of time it takes to get a reactor permitted, built, and connected to a data center where revenue is being generated), is well over seven years.
TerraPower’s 345 MW small modular Natrium reactor in Wyoming has been under construction since 2024 and won’t be ready until 2030. Even when it’s done, it won’t satisfy a data center’s appetite. Allison Macfarlane, the former chair of the US Nuclear Regulatory Commission summed up the nuclear option for AI in an interview not long ago. She said, “Commercializing the [SMNR] technology will be difficult because a smaller reactor core also means a less efficient reactor – producing less energy from the same amount of fuel. You just can’t get around economies of scale. These are fun ideas, but the tech bros don’t seem to be grounded in reality.”
The reality at the moment is that hyperscalers need even more power than SMNRs can deliver alone, which means any nuclear plants capable of meeting their requirements take much longer, and the costs are mind-blowing. Units like Plant Vogtle near Waynesboro, Georgia, a 3.3 GW station that first started construction in 1987 and was just completed in 2023 – 36 years – came in at a price tag of well over $50 billion. Money aside, that time-to-compute doesn’t work.
Natural gas is a faster alternative, but time-to-compute isn’t much better than for an SMNR plant. New natural gas plants can take 5 to 7 years to come online and the day the first permit is filed for a new power station is the same day the gas-fired turbines for that station need to be ordered. Wait time for turbines is now over seven years and, of course, the costs, now up by 2.5X, are reflective of the demand. Natural gas fuel prices aren’t helping the matter either. Not long ago, prices at the Henry Hub jumped 57.6%, to $3.31 per million British thermal units. Pair all of this with the fact that natural gas pipelines are themselves a grid to which a data center must be connected, and you find yourself with a similar layer of problems associated with nuclear electrification.
What about solar? Yes, it’s intermittent power, but the Solar Energy Industry Association reports that more than 8,100 major solar projects currently in their database represent over 339 GW of capacity, which is a lot. EIA says that 63 gigawatts of new utility-scale solar capacity will be added in the U.S. this year. Subsidies for solar, however, are being clawed back by the government and by 2026, the ability to get a permit for a utility scale solar installation will be much tougher. Utility-scale solar today is more than 80% cheaper than it was a decade and a half ago, so the attractiveness of the cost does open an opportunity but what about time-to-compute? Permitting and construction aren’t the most difficult part of the equation; grid capacity is the challenge. The typical timeframe from connection request to commercial operation has now hit five years or more.
Time to Re-Compute
“The main thing is the plain thing,” I once heard said. And the plain thing these days is that supply cannot meet demand. Time-to-compute can only be shortened by working around grid capacity constraints, so assuming we do want the U.S. to win the AI race against competitor countries, we have to adopt new criteria for data centers’ build out.
Get Off-grid
Data centers have to go off-grid because the grid doesn’t have, and won’t have for many years to come, the capacity to manage the supply even if it were plentiful. When the power can’t get to the place it’s needed, it’s wasted. Further, if data centers are going to be formatted at hyperscale size, they must be moved away from population centers and not on top of productive agricultural land. Ratepayers, jolted into a panic by their rising electric bills, are going to become anti-data center activists and eventually, they’re going to build a list of grievances that elected officials and regulators won’t be able to ignore.
Get Clean
Local residents are beginning to challenge new data center installations for an array of reasons, including health concerns from gas turbine emission concentrations. Those same concerns will eventually eliminate diesel-based power generation from the option set as well. The electric grid, energized as it is today, certainly is not the cleanest source of power.
Get Compact
Are gigantic-sized data centers the only template available today? OpenAI’s Stargate facility in Abilene, Texas, is a 1.2 GW facility. Amazon’s data center outside of New Carlisle, Indiana, will consume 2.2 GW, enough to power 1 million homes, when it’s built out. Meta is building a 4 million square foot, 2 GW capacity data center in Richland Parish, Louisiana. The second phase of Elon Musk’s Tennessee-based data center is slated to devour 1.5 GW of power. While the answer to the question seems to be “yes,” the reality is, gigantic does not have to be the model.
Meta has already broken the mold. It’s using tent-like structures at its data center in New Albany, Ohio, to cut time-to-compute in half. There, the first five buildings took three years to build. Meta started building five ~125,000 square foot tents between April and June and expects them to be complete within a year. Aerial images taken a month ago show that all five structures have already been built.
On my Path to Zero podcast episodes, I often hand guests an imaginary magic wand and ask them to wave it to make a profound change to today’s energy-climate-power puzzles. The question often unlocks a fantastic idea from a brilliant guest and having experienced the power of the device first-hand, I’ll wave it now.
Getting off-grid, getting clean, and getting compact naturally changes the trajectory of data center design away from “mammothication.” When these three concepts frame the design, solutions come quickly into view. Imagine remote-located, grid-independent, human-scaled data centers powered by solar plus utility-scale batteries reinforced by propane or renewable propane generators to solve for intermittency, and voila! Many of the data center downsides we’re facing today are quickly managed and time-to-compute is accelerated.
There you have it. And the best news of all is that we have what we need today to solve the challenges of tomorrow.
