Power Play | Mini-Series Issue #2 | September 2026
Hello friend,
Power Play is officially live!
When I started working on this series with Quartz, I assumed the big question would be whether America could produce enough electricity for AI. After spending months in the reporting, I’m not sure that’s even the most interesting question anymore.
Episodes 1 and 2 get at why. We’re trying to plan for a huge new source of electricity demand without really knowing how big it will be, while sitting on a grid that has more capacity than the headlines might suggest. Forecasting, infrastructure, incentives, timing, and access are all tangled together.
That’s what I want to unpack today.
A forecast is not a fact
One of the numbers that stuck with me from Episode 1 is the spread between credible forecasts for AI electricity demand in 2030: roughly 200 terawatt hours to 1,000 terawatt hours. That’s a fivefold difference for something utilities are already being asked to plan around.
Forecasting is always imperfect, but the stakes change when the response involves infrastructure that takes years to finance, permit, and build. An AI company can change its model or business plan relatively quickly. A substation or transmission upgrade can’t.
The interesting part is that uncertainty doesn’t give anyone permission to sit still. Utilities still have to make decisions. Investors still have to place bets. Communities still have to decide what gets built around them.
The lesson: The more permanent the decision, the more important it is to understand how temporary the assumptions underneath it might be.
When demand isn’t really demand
Then there’s the question of whether all the demand being forecast is even real.
Developers submit requests to utilities to reserve access to the grid, but those queues can include everything from serious hyperscale projects to speculative developers trying to increase the value of well-located land.
One example in Columbus makes the problem tangible. AEP reportedly had around 30 gigawatts of power requests. After introducing a rule requiring developers to put meaningful money behind them, that number dropped to 13 gigawatts.
Nothing changed about the grid overnight. What changed was the quality of the signal.
The lesson: Pipelines, waitlists, reservations, and expressions of interest can all look like demand. But interest is cheap until someone has to commit money, time, reputation, or a signed contract.
The grid paradox
Episode 2 flips the problem around.
If America is scrambling to find enough electricity for AI, why does the grid operate at roughly half its capacity on average?
The answer is that average capacity and usable capacity are not the same thing. The grid has to survive the hottest afternoon in Texas or the coldest stretch of winter, not just perform well on an ordinary Tuesday.
And before a new data centre can connect, utilities need to study what that load will do to the system. Equipment can take years to procure. Different utilities, states, and grid operators make decisions across a system that was never designed to respond to an industry asking for gigawatts in months.
That’s what makes “just build more power” feel incomplete. Sometimes the constraint is supply. Sometimes it’s access, equipment, incentives, or process.
The lesson: Capacity and access aren’t the same thing. Before adding more supply, it’s worth understanding what’s actually creating the bottleneck.
The question I’m carrying forward
What I find most interesting after these first two episodes is how quickly the simple version of this story falls apart.
“AI needs more power” is true, but it doesn’t tell you how much. “The grid is full” isn’t quite right either. And even the demand figures themselves are shaped by projects that may never happen.
That doesn’t mean the AI power problem isn’t real. It means the decisions are harder than the headlines make them sound.
That’s also what I’ve enjoyed about making the series. Jackie Snow, Cat Baab-Muguira, and Joseph Zeballos-Roig did the reporting; my job was often to ask the obvious question, hear the answer, and realize the obvious question wasn’t quite right in the first place.
Where we go next
Episodes 1 and 2 establish the basic tension: we’re making enormous infrastructure decisions under uncertainty, while the system we already have is more complicated, and in some ways more capable, than it first appears.
From here, we’ll look at what this buildout means for land and communities, whether we can use the existing grid better, and who actually writes the rules that determine what gets connected and who pays for it.
For now, AI wants your electricity and Gridlock are both live.
If you’ve been following the AI boom from the software side, I think they’ll make you look at the physical side of it a little differently.