Substation Scout

September 24, 2026 · By Phil

Ahead of the curve and Behind-the-Meter

Data centers are "bringing their own" energy, and the trend is growing.

As I point out in my foreword on why I made this site in the first place, America is in a race for the future and we need to invest heavily in our grid to win it.

But that takes time. AI is moving so fast that we need stopgap solutions. In this context, it's no wonder so many AI data centers have decided to "bring their own power."

The approach typically takes one of two high-level forms:

1. On-site generation

  • Usually with natural gas
  • The gas either powers turbines from companies like GEV or is used in fuel cells from companies like Bloom Energy.

2. Co-location

  • Not the kind you are thinking, where we co-locate server hardware.
  • Energy co-location involves siting your data center near a big power plant, and getting a direct plug from the plant, bypassing the grid.

I find this topic to be interesting and monumentally important for us to keep pace as a nation in the AI race. Originally I had a Behind-the-Meter Tracker as one of my roadmap items on this site, but I found most of the info out there on these projects difficult to verify (thus undermining my "all sources verified" ethos).

But that doesn't mean we shouldn't talk about it... not least because the strategy has its own pitfalls, and we should track those and learn from them. In the two weeks leading up to this writing, we had two nice little learning moments:

In New Jersey, the DataOne facility being built for NBIS in Vineland was fined $1.07 million by the state DEP for installing and operating 62 natural gas generators — about 123 MW worth — without the required air permits. The wrinkle here: those generators are just the temporary bridge power. The permanent behind-the-meter plan is a three-phase Bloom Energy fuel cell deployment, and the developer says that one is already approved. So now our stopgaps need a stopgap. Wow.

And in New Mexico, Oracle sent a force majeure notice to Blue Owl Capital's Stack Infrastructure, the developer of the 2.45 GW Project Jupiter campus, preserving its right to defer payments if the site misses its 2028 opening. The pressure point is fuel. Jupiter is designed around Bloom fuel cells, and the Energy Transfer pipeline meant to feed them was denied right-of-way twice by the New Mexico State Land Office, pushing its in-service date from August 2026 to February 2027. Oracle says the project remains on schedule and that it stays committed to New Mexico. So this is a contractual hedge, not a cancellation. Still a bit dicey.

It would seem for the gas projects in particular that "behind-the-meter" may be a bit of a misnomer, as they've still got to stand in front of the gas meter — and the inspector.

These projects are nonetheless exciting and their developers really are ahead of the curve. I'll look to fold more Behind-the-Meter content into this site where appropriate. For now, just some fun further reading for data center and energy geeks:

On-site generation... gassing you up

Why it's cool

  • America is unique in its production and distribution of natural gas. It's super cool that we are leveraging this.
  • America is also unique in its precision manufacturing and innovation: the gas turbines from folks like GEV are themselves an engineering feat which other countries haven't been able to replicate; and the fuel cell technologies we pioneered are going mainstream at the perfect time.

Challenges

  • Supply Chain: GEV and its competitors have massive backlogs. You can't just 3D print one of these things, they are massive and incredibly precise jet engines. GE Vernova's gas equipment backlog and slot reservation agreements went from 100 GW to 116 GW in a single quarter this year, with at least 125 GW anticipated by year end.
  • Materials: Bloom Energy is emphatic that it does not expect the availability of rare earth elements from China to affect its 2026 production forecast, and its CEO has repeated that the company has no China supply chain. A short seller investigation published in July disputed that, focusing on scandium, the rare earth at the core of each fuel cell. Bloom filed an 8-K denying the allegations; a shareholder suit followed. Two parties, two sets of documents, no resolution yet... Worth watching.
  • Permitting: Per the Vineland story above, regulators aren't always on board with the speed at which these things are sited and brought online.

Selected links

Co-location... the coup d'tap

Why it's cool

  • What makes this one so cool is that it started as genuinely uncharted territory for energy deals... every deal was sort of a mini-coup against the energy establishment.
  • That's because the normal arrangement is that the power plant sells to the grid, and you deal with a franchise utility who has authority over the local portion of the grid. In these co-location arrangements, data centers are locating next to power plants and plugging more or less directly into them. Nobody was ever banned from buying the electrons — the fight was over who pays for the transmission system you're now sitting beside but not paying into. Utilities argued the structure shifted those costs onto everyone else.
  • This opens up a whole new category for site selection as new generation projects come online around the country and existing ones are scouted for nearby land opportunities. And it's a whole new reason to keep one's eye on the Queue Wire.

Challenges

  • The coup already got noticed and institutionalized. In December 2025 FERC found PJM's tariff unjust and unreasonable as to co-located load and ordered PJM to write real rules; by April 2026 FERC had issued a compliance order on PJM's Bring Your Own Generation tariff, approving four interconnection paths and rejecting two of PJM's proposals; by June it had added a broader category called Eligible Load, and PJM had set a 50 MW threshold for large load procedures. But what's still grey is rates and cost allocation, not legality.
  • The most famous co-location deal ended up going through the meter. FERC rejected the Talen-Amazon behind-the-meter interconnection agreement at Susquehanna and declined to rehear it. Talen instead signed a 1,920 MW front-of-meter PPA with AWS, which didn't require FERC approval at all, and the site made the behind-to-front transition in spring 2026. This is worth remembering when a site is pitched to you on the strength of a direct plug.
  • Big generation projects (the kind you'd want to co-locate next to) also take a long time to come to fruition if they're new (they're waiting in the interconnection queue themselves). If they're not new, they're an existing plant, and the real challenge is finding buildable land close enough to suit a mega campus.

Selected links

Other approaches... if you can't bring power, bend load

Why it's cool

  • The first two approaches are about getting electrons. This one is about being a better customer. If you can promise to back off for a couple of hours a year, the grid can say yes much sooner — and you never had to buy a turbine.
  • On-site batteries started life as UPS backup and have quietly become an interconnection tool. Aligned put a 31 MW / 62 MWh system next to a Pacific Northwest campus specifically to get connected years earlier than a conventional utility upgrade would have allowed. The battery isn't the power source. It's the concession that makes the utility comfortable.
  • Duke's Nicholas Institute put a number on the idea: across the 22 largest balancing authorities, 76 GW of new load could be integrated at an average curtailment rate of 0.25% of uptime, rising to 126 GW at 1%, with roughly 18 GW of that in PJM at 0.5%. Note carefully what that is and isn't. It's a system-level estimate of how much load fits under existing peaks. It is not a promise about any particular site.
  • And then there's nuclear, which is the opposite trade: nothing fast, everything durable. The hyperscalers have signed more nuclear offtake in the last two years than the country signed in the prior twenty.

Challenges

  • Curtailment is a contractual promise, not a physical property. A cluster running inference under an uptime SLA has a very different appetite for 44 hours of annual interruption than a training cluster does, and critics have argued that system-level headroom says nothing about whether the flexibility exists at the specific node where the load wants to sit.
  • Batteries buy hours, not months. A two-hour system covers a peak event. It does not cover a delayed interconnection.
  • The nuclear timeline is the whole problem. NuScale's cancellation of the Carbon Free Power Project in 2023 remains the reference point for the gap between a signed agreement and a running reactor. Most announced deals land in the early 2030s, and the ones arriving sooner are existing-plant restarts rather than new builds.
  • This category is also the hardest one to verify from the outside, for exactly the reason I cut the tracker: a flexibility commitment lives in a service agreement, and service agreements aren't public. What is public is the tariff proceeding that made the commitment possible. That's where I'd rather spend my time.

Selected links

Sources retrieved September 24, 2026.

More

All insights → · Queue Wire weekly deltas →