Name the constraint first
Make the utility name the constraint before designing anything. Queue position, network upgrade, and capacity shortfall are three different problems.
IN-01Insight · Grid interconnection
Behind-the-meter AI compute: when it beats the queue
On-site generation does not make you invisible to the grid, and it rarely wins on the price of electricity. It wins — or fails — on how many months of delay it actually removes. Here is the framework and the arithmetic, with every input stated.
PUBLISHED LAST VERIFIED BY JOSEF ELIMELECHREVIEWED PODOS AI ENGINEERING
- 5+ yrs
- Median US request to commercial operation, 2025 builds
- 13%
- Of capacity requested 2000–2020 operating by end 2025
- 100 h/yr
- Federal ceiling on emergency-classified engines
The short answer
Behind-the-meter generation pays only when it pulls energization forward by more than the delay needed to repay its capital and its energy premium — a threshold near two years at plausible AI-compute margins, not the five-plus years operators assume they are escaping. The queue is real: median request-to-operation exceeded five years for US projects built in 2025, and only 13% of the capacity requested between 2000 and 2020 was operating by the end of 2025.[1]But the deciding questions are regulatory and protection-engineering ones — the prime mover's permit class, whether the site can island deliberately, and how firm its grid-facing service must be.[2][5]
First principles
What you are actually escaping — and what you are not
The queue you are avoiding may not be the queue you are in
As of the end of 2025, roughly 8,200 projects were actively seeking grid interconnection in the US — 1,312 GW of generation and about 749 GW of storage. Median duration from request to commercial operation was over five years for projects built in 2025, and of the capacity requested between 2000 and 2020, 13% was operating by the end of 2025, 75% had been withdrawn, and 10% was still waiting.[1]
The number that should change how an operator reads this is quieter: 549 GW already holds a draft or executed interconnection agreement and still has not reached commercial operation, including 45 GW of gas.[1] For that capacity, the interconnection right is not the binding constraint — equipment, labour, and capital are. A private generator buys you out of the study queue, not the turbine order book, and active gas capacity in the queues grew 86% in 2025, to 253 GW[1]— the signature of many developers reaching the same conclusion at once. FERC's Order No. 2023 queue reforms are being implemented,[3] but Berkeley Lab judges it too early to measure their effect.[1]
Behind-the-meter is not a way to be invisible
On 18 December 2025, FERC ruled on the co-location proceeding it had opened against PJM the previous February. It found PJM's tariff unjust and unreasonable for lacking clarity on service to co-located load, found the tariff's existing Behind-the-Meter Generation rules no longer just and reasonable, and directed PJM to make the customer serving a co-located load choose among four transmission services: network integration service, a new interim non-firm service available while upgrades complete, and new firm and non-firm contract demand services.[2]
Read structurally rather than as news, that says a load sitting beside its own generator is still a transmission customer. The variable is not whether you take service but how firm it is — FERC framed the gap as a missing provision for customers "able to limit their energy withdrawals from the transmission system."[2] That turns islanding capability into a commercial instrument. A site that can verifiably curtail can argue for non-firm service; a site that cannot buys firm service, and firm service at campus scale is most of the money co-location was supposed to save.
Table 1 · Original framework
The decision framework: six gates, in order
The gates are sequential and disqualifying. A plan that fails an early gate does not get to argue the economics of a later one — which is the usual failure mode, because the economics are the easiest part to model and the least likely to be the constraint.
| Gate | Question | What to evaluate | What disqualifies behind-the-meter |
|---|---|---|---|
| BTM-01 | What is actually blocking energization? | Study-queue position, network upgrade, or delivered capacity — make the utility name it. | On-site generation does not clear an upstream network upgrade that the load itself triggers. |
| BTM-02 | Is the prime mover permitted to run? | Air-permit class, expected annual run hours, site emissions limits. | Engines classified as emergency are capped at 100 hours a year and barred from peak shaving or paid supply. |
| BTM-03 | Can the site island deliberately? | Protection coordination, intentional-island design, reconnection ramp rate. | A site that cannot resynchronize under control is a liability to the grid and to itself. |
| BTM-04 | How firm must the grid-facing service be? | The service taken when not islanded; whether withdrawals can be verifiably limited. | A load that cannot curtail pays for firm service — which is most of the saving. |
| BTM-05 | What is the residual reliability plan? | On-site plant redundancy, fuel supply, N+1 prime movers, standby behind them. | A standby classification sized for a grid outage does not cover the on-site plant's own forced outages. |
| BTM-06 | What is the exit? | Date grid capacity arrives, cost of stranding, second life as backup or peaking. | Bridge plant with no second life must repay its whole cost inside the bridge window. |
The existing plant
The plant you already have is not the plant you need
Most data centers already island — accidentally
NERC's review of a July 2024 event in the Eastern Interconnection is the most useful document in this debate, and it is rarely read as one. A lightning arrestor failed on a 230 kV line; auto-reclosing configured for three attempts at each terminal produced six successive faults in 82 seconds, each cleared correctly in 42 to 66 milliseconds, with voltage in the affected area dipping to 0.25–0.40 per unit. About 1,500 MW of load came off — all of it data-center type, none of it disconnected by utility equipment. Customer-side protection transferred the load to backup, and roughly 1,260 MW did not return for hours.[4]
The mechanism matters more than the magnitude. Beyond ordinary UPS transfer, sites had deployed a disturbance-counting scheme: typically, three voltage disturbances within one minute triggers a transfer to backup, held there until manually returned.[4] That is an islanding policy — already installed at gigawatt scale, and badly specified. It separates on the grid's schedule, runs on plant legally constrained in annual hours, and reconnects by hand. Deliberate behind-the-meter generation is the same event with the sign flipped, and the discipline for doing it on purpose is the DER interconnection standard, which treats islanding, abnormal-condition response, and power quality as interconnection requirements rather than site preferences.[6]
The permit ceiling nobody prices in
The most common behind-the-meter plan is also the most common mistake: run the existing standby generators. Federal air rules do not allow it. An engine classified as emergency may operate a maximum of 100 hours per calendar year for maintenance and readiness testing, at most 50 of those hours in non-emergency situations, and those hours "cannot be used for peak shaving or non-emergency demand response, or to generate income for a facility to supply power to an electric grid." Operate outside the limits and the engine is no longer an emergency engine — it must meet every requirement applying to non-emergency engines.[5]
The standby fleet is a bridge measured in hours. Standby systems are classified by how long they carry load and how fast they pick it up,[7] which is a different design question from annual duty cycle, fuel logistics, forced-outage rate, and overhaul intervals. Behind-the-meter prime power is a generation project that happens to sit on your parcel, and it moves reliability risk onto the operator — half of surveyed operators (50%) reported an impactful or serious outage in the prior three years even with the utility carrying supply.[8]
The standby fleet is a bridge measured in hours, not a generation project.
1,500 MW
Data-center load that islanded itself in one 82-second event
Figure 1 · Original analysis
The worked calculation: how much delay must it avoid?
Behind-the-meter generation does not create value by being cheap power. It rarely is. It creates value by moving revenue earlier in time, so the settling question is how many months of delay it must avoid to pay for itself.
Per megawatt of IT load, the breakeven delay D* in months is:
D* = 12 × ( Cbtm + ΔE × 8,760 × LF × T ) ÷ V
Cbtm is the incremental capital cost of the on-site plant per MW, ΔE the energy-cost premium of self-generation over delivered grid power in $/MWh, LF the annual load factor, T the years the premium is paid before grid service arrives, and V the gross contribution earned per MW-year of running compute.
Stated assumptions — placeholders, not quotes or PODOS figures
- C(btm) = $1.2 M per MW incremental for generation, switchgear, and fuel infrastructure.
- ΔE = $25/MWh premium for self-generated energy over delivered grid power.
- LF = 0.85 annual load factor on the IT load.
- T = 5 years of premium paid before grid service would have arrived anyway.
- V = gross contribution per MW-year of operating compute — the input that varies most, and the one you must supply.
At V = $1.0 M per MW-year the energy premium totals $25 × 8,760 × 0.85 × 5 = $930,750 per MW, so the numerator is $1,200,000 + $930,750 = $2,130,750 and D* = 12 × $2,130,750 ÷ $1,000,000 ≈ 25.6 months.
37.7
51.1
73.5
18.9
25.6
36.7
9.4
12.8
18.4
Table 2 · Breakeven delay avoided, D* in months
Reading the result: the spread is the point
Across plausible inputs the breakeven sits between roughly nine and seventy-three months, so the answer is rarely decided by the cost of generation. It is decided by how much compute margin an idle megawatt is losing — one input that swings the result fourfold. Two operators can face the same site, fuel price, and queue position and be correctly opposed.
| V per MW-year | ΔE = $10/MWh | ΔE = $25/MWh | ΔE = $50/MWh |
|---|---|---|---|
| $0.5 M | 37.7 | 51.1 | 73.5 |
| $1.0 M | 18.9 | 25.6 | 36.7 |
| $2.0 M | 9.4 | 12.8 | 18.4 |
Practice
What this means for operators
Price the delay
Price the delay, not the electricity. If the breakeven exceeds the delay you can realistically avoid, the fuel-price debate is irrelevant.
Read the air permit first
Audit the air permit before the single-line diagram. A plan resting on emergency-classified engines is already outside federal limits.
Make islanding a deliverable
Specify islanding as an interconnection design: protection coordination, resynchronization, and reconnection ramp rate are the deliverables.
Curtailability is commercial
Treat verifiable curtailability as a commercial asset — it is what separates cheap non-firm service from expensive firm service.
Honest limits
What this does not prove
The evidence above is strong on regulation and queue statistics, and weaker everywhere the decision actually gets made.
- The breakeven is a structure, not a result. Every input above is a placeholder — no market quote, no vendor price, no PODOS figure — and the output moves almost eightfold across the grid.
- The FERC order covers one region and is not final. It sets a paper hearing rather than final rates, and declines to resolve jurisdictional questions about retail load served through co-location.
- Queue medians do not predict a single project. They describe thousands of heterogeneous generation projects, not one load's service request — which is not even the same process.
- One incident is not a fleet characterization. NERC analysed a single event; it shows the disturbance-counting behaviour exists at scale, not how widespread that configuration is.
- Federal air limits are a floor. State and local permitting, nonattainment status, and nuisance rules can rule out on-site combustion entirely.
- The framework assumes the compute is worth running now. If capacity is built ahead of demand, the value of arriving early collapses, and the case with it.
In the product
Where the PODOS approach sits in this
Gate BTM-06 punishes plant sized for a campus that never materializes, and the breakeven rewards whatever arrives soonest at the smallest committed increment. That is the design premise of the PODOS Pod, which is designed as a standardized 1 MW building block and designed for 128 GPUs; PODOS targets a 90-day window from order to commissioning for a standard unit. A 1 MW increment is a smaller bet against an uncertain interconnection date than a campus-scale commitment, and it lets generation be sized to load that already exists.
The gates map onto the rest of the cluster: BTM-03 and BTM-05 are power architecture questions, BTM-04 depends on monitoring and controls, and V rests on high-density GPU infrastructure. For site prerequisites, use the readiness checklist and the deployment model; for the schedule comparison behind the calculation, see modular vs traditional AI data centers. Terms are defined in the AI infrastructure glossary, and you can size an increment in the configurator.
Sources
- [1] Queued Up: 2026 Edition — Characteristics of Power Plants Seeking Transmission Interconnection (data through end of 2025) — Lawrence Berkeley National Laboratory, Jun 2026
- [2] FACT SHEET | FERC Directs Nation's Largest Grid Operator to Create New Rules (Docket Nos. EL25-49-000, et al.) — Federal Energy Regulatory Commission, 18 Dec 2025
- [3] Order No. 2023 — Improvements to Generator Interconnection Procedures and Agreements — Federal Energy Regulatory Commission, 2023
- [4] Incident Review — Considering Simultaneous Voltage-Sensitive Load Reductions — North American Electric Reliability Corporation, accessed 2026-08-31
- [5] 40 CFR 63.6640 — operation requirements for stationary RICE (NESHAP subpart ZZZZ) — eCFR / U.S. EPA, accessed 2026-08-31
- [6] IEEE Std 1547-2018 — Interconnection and Interoperability of Distributed Energy Resources with Associated Electric Power Systems Interfaces — IEEE Standards Association, 2018
- [7] NFPA 110, Standard for Emergency and Standby Power Systems — National Fire Protection Association, current ed.
- [8] Annual outage analysis 2026 (reporting Global Data Center Survey 2025 results) — Uptime Institute, 2026
Size the increment against your interconnection date
Bring the constraint the utility named, the delay you are actually trying to avoid, and the margin an idle megawatt is losing. Engineering will tell you which gate you fail first.