Three Time-to-Power Strategies that Failed in 2025

We’ve learned a lot in the last year about what won’t work for accelerating time-to-power for large data center projects.

Building behind-the-meter load at existing power plants

Amazon and Talen Energy had the idea to build a large data center at an existing nuclear power plant, Susquehana, and to connect the data center directly to the power plant, not to the grid. By avoiding a grid connection, the data center wouldn’t have to go through PJM’s interconnection queue process.

But FERC rejected the plan for complicated cost allocation reasons. Basically, when Susquehanna was built, the local utility spent a lot of money building transmission upgrades to support the plant. All PJM customers are now paying back the utility for those investments. If the Susquehanna now disconnects from the grid to supply Amazon directly, PJM grid users would be stuck paying off Susquehanna’s upgrades, which doesn’t seem fair now that they won’t be getting the benefits of having Susquehanna’s capacity on-grid.

So it seems that this strategy is dead. Once a power plant is on-grid, you can’t take it off-grid. Amazon has pivoted that data center project to being on-grid.

Restarting mothballed power plants

Microsoft and Constellation had the idea to restart the mothballed Three Mile Island nuclear power plant (yea, that one). This is an on-grid strategy, so Microsoft’s data center would have to go through the interconnection queue. Slow. But it seemed like a relatively fast way to bring online nuclear capacity to provide clean, firm power to Microsoft’s data center.

Unfortunately, it’s turned out that Three Mile Island’s grid interconnection is effectively gone and will have to start from scratch. It still has its physical connection to the grid, but in the time that it’s been offline, the “backbone” transmission capacity that it had has been filled by other generators. New backbone capacity will have to be built before the power plant can come online, which means it has to go to the back of the interconnection queue like any other project.

Taking over interconnections from shuttered loads, like closed down factories, is unlikely to work, for the same reason.

There’s still an advantage to restarts. It’s always faster to restart than to build from scratch. But it’s not a strategy for jumping the queue, like people hoped.

Off-grid data centers

Fermi had the idea to build off-grid power generation that data centers could connect to behind-the-meter. That way, both the load and the generation skip the interconnection queue. And it avoids the cost allocation problem of Susquehana because the grid is mostly irrelevant to a mostly off-grid project.

Off-grid has inherent disadvantages. Grid power, which is produced by an optimal mix of thousands of plants, is always going to be cheaper and more reliable, on average, than off-grid power. The trade is that, theoretically, off-grid power can come online faster because it doesn’t have to go through the interconnection queue and wait for transmission upgrades to be built.

The problem Fermi ran into is that the wait time to procure generating units - nuclear AP1000s and natural gas turbines - stretched into the 2030s. At that rate, you might as well get into the interconnection queue to get the benefits of grid power. Their customer backed out.

What strategies remain viable?

Traditional development. Maybe slow and steady wins the race. Just get in the queue and manage the project as best you can.

Large flexible loads. If grid operators can rely on data centers to turn off during the tightest 1% of hours on the grid, then data centers could come online before their transmission upgrades have been built. This depends on policy changes from the various grid operators, and it brings up complicated operational, market, and regulated cost allocation questions. But I’m optimistic that this will work.

Data centers could take over the interconnection of other large loads. A developer buys an aluminum or a cement plant, shuts it down, and builds a data center in its place, plugging the data center into the old plant’s grid interconnection. Presumably there’d be no issue from the grid operator’s perspective, as long as the data center’s peak load is less than that of the plant it’s replacing. But the local politics would be brutal. You’d be shutting down a plant that presumably employs hundreds of people in a community to replace it with a data center that would employ dozens of the people. You’d need the community’s permission to build the data center, and I don’t think they’d be too happy with you. I wouldn’t bet on this.

Distributed energy resources - like residential/commercial solar and battery storage - can be deployed quickly to take existing residential load off of the system and make room for new large data center loads. This will tend to be more expensive than utility-scale generation because the fixed install costs are spread over a much smaller capacity at every individual site. There’s also a big operational overhead to acquiring customers, installing the resource, and maintaining it. But it’s the fastest way to get capacity on the grid. In order for data centers to benefit from it, grid operators would need to account for this increased distributed generation in their planning models in order for it to “make room” for new large loads. There would also need to be rules by which a large load gets credit for supporting distributed generation. At the moment, a data center developer can’t move ahead in the interconnection queue by supporting distributed generation.

Space-based compute is a very complicated and expensive way to build a 1 GW data center, but it’s probably the only way to build a 100 GW data center. It’s the most scalable approach. So if scaling laws are still holding in, say, 2040 and AI labs want to build truly ludicrously large clusters, whoever is ahead on space-based compute will run away with the game. The question is whether SpaceX / xAI can stay solvent until then. Unfortunately, this project, which should be the coolest thing ever, is likely to be ruined with a stupid name that Elon Musk thinks is funny. My money is on ‘Skynet’ or a pun based on that.