This case exists to complicate the cluster's own emerging narrative before it hardens into 'AI data centers strain the grid,' full stop. Meta and Williams Companies are building a dedicated gas-fired power plant — two 200 MW units, 400 MW total — in New Albany, Ohio, approved by the Ohio Power Siting Board on June 9, 2025, and targeted online in late 2026.[1] It sits inside PJM's own territory, in AEP Ohio service area, the same grid documented elsewhere in this cluster missing its own capacity-reliability target twice in a row. But the plant runs behind the meter — generating power directly for Meta's adjacent data-center campus, not feeding or drawing from PJM's shared system.[1] Separately, Ohio's Public Utilities Commission approved a data-center-specific tariff in July 2025 requiring large data-center customers (over 25 MW) to commit to paying for at least 85% of their subscribed capacity for up to 12 years, with steep exit fees — a regulatory fix that doesn't reduce physical grid demand but insulates residential ratepayers from the stranded-cost risk of a data center overbuilding and then leaving.[2] Neither example eliminates AI's grid impact broadly. Both show that the strain this cluster documents isn't an inevitable feature of AI buildout — it's a choice some operators and regulators are actively opting out of.
Meta and Williams Companies' New Albany, Ohio project — approved by the Ohio Power Siting Board on June 9, 2025 — pairs two 200 MW natural-gas generating units, 400 MW total, directly with a Meta data-center campus.[1] The critical design choice is that it runs behind the meter: the plant's output goes to Meta's facility directly, not onto PJM's shared transmission system. Targeted to come online in late 2026, it sits inside the exact grid — PJM, AEP Ohio territory — documented elsewhere in this cluster as missing its own capacity-reliability target in two consecutive auctions.
This matters precisely because it's the same region under strain, not a different one. A 400 MW load with its own dedicated generation doesn't draw down PJM's shared capacity margin the way a grid-connected data center of the same size would — it's effectively invisible to the shortfall this cluster's at-risk case measures. One hyperscaler, one specific project, choosing to bring its own power rather than compete for a share of an already-short capacity market.
A second, complementary example addresses the cost side rather than the physical-demand side. Ohio's Public Utilities Commission approved an AEP Ohio tariff in July 2025 requiring data centers over 25 MW to commit to paying for at least 85% of their subscribed capacity for up to 12 years, with substantial exit fees for early termination.[2] This doesn't reduce how much power a data center draws from the shared grid — it protects residential and other ratepayers from the stranded-infrastructure cost risk if a data center operator overcommits and later leaves, a distinct but related failure mode this cluster's other cases don't address.
The honest limit of this case: one 400 MW behind-the-meter plant and one state's ratepayer-protection tariff don't offset the multi-gigawatt capacity shortfall documented elsewhere in this cluster — they're a genuine complication, not a solution at scale. Most AI data-center growth in PJM's territory is still grid-connected, still drawing on the same strained capacity market. What this case supports precisely is narrower and still important: the industry has real, working examples of AI buildout that doesn't add to shared-grid strain, proving the strain is a design choice as much as an inevitability.
A real, approved, dated project — inside the exact grid under strain — built specifically to not draw on it.[1]
How one hyperscaler and one state regulator both opted out of the grid-strain dynamic this cluster documents.
The Ohio Power Siting Board approves Meta and Williams' 400 MW dedicated generation project for the New Albany data-center campus.[1]
The ApprovalPUCO approves AEP Ohio's data-center tariff, requiring large customers to commit to minimum capacity payments for up to 12 years — a separate, cost-side fix.[2]
The TariffPJM's capacity market misses its reliability target twice in the region where Meta's project sits — the counterexample and the shortfall coexist in the same grid.
The ContrastMeta's dedicated generation is expected to begin serving its data-center campus directly, drawing nothing from PJM's shared system.[1]
TargetedAs of this writing, no data yet shows whether other hyperscalers in PJM territory are adopting comparable behind-the-meter arrangements at scale.
UnresolvedApproved for construction of a 400 MW behind-the-meter generation facility. — Ohio Power Siting Board, Case No. 24-1449-EL-BGN, June 9, 2025
| Dimension | Evidence |
|---|---|
| Operational (D6) Origin · 82 | The lever is a real, physical infrastructure choice — a 400 MW plant built specifically to serve one data center directly, avoiding the shared grid entirely.[1] D6 is the origin because this entire case is about a concrete operational design decision, not a policy abstraction.Dedicated Generation, By Design |
| Revenue (D2) L1 · 72 | AEP Ohio's tariff is a direct financial mechanism protecting ratepayers from a specific cost-risk this cluster's other cases don't address.[2] D2 amplifies from D6 as the complementary cost-side response to the same underlying pressure.The Ratepayer Fix |
| Regulatory (D4) L1 · 70 | Both examples exist because state regulators — the Ohio Power Siting Board and PUCO — approved them, real institutional decisions enabling the counterexample.[1][2] D4 amplifies alongside D2 as the regulatory enabler. |
| Quality (D5) L2 · 56 | The honest boundary this case insists on — real but partial complications, not a solution at scale — keeps the counterexample from overclaiming a broader trend the evidence doesn't yet support. D5 sits here as that discipline. |
| Customer (D1) L2 · 48 | Residential ratepayers protected by the AEP Ohio tariff are a direct beneficiary population, distinct from the broader exposed population in this cluster's other cases.[2] D1 sits here as that specific, protected group. |
| Employee (D3) 28 | Deliberately the thinnest dimension. This is an infrastructure and regulatory-design cascade; no comparable workforce-level finding exists in the research. |
The cascade originates in D6 — Operational — because the lever is a real, physical infrastructure choice: dedicated generation built specifically to avoid drawing on the shared grid.[1] From D6 it moves to D2 (the ratepayer-protection mechanism addressing the cost side of the same underlying risk) and D4 (the state regulatory approvals — Ohio Power Siting Board and PUCO — that made both examples possible).[2] It then reaches D5 (the honest boundary — these are real but partial complications, not a solution at scale) and D1 (residential ratepayers protected by the tariff, and the broader population this cluster's other cases show exposed). D3 is deliberately thin — an infrastructure and regulatory-design cascade, not a workforce one. Cross-references: [UC-285] documents the emergency-generation mechanism this project's design specifically avoids needing; [UC-286] is the capacity shortfall this load doesn't contribute to; [UC-288] must weigh this counterexample honestly rather than assume all AI buildout strains the grid uniformly.
-- UC-287: The Data Center That Left the Grid: 6D Amplifying Cascade (Counterexample)
-- Meta's 400MW behind-the-meter plant + Ohio's ratepayer-protection tariff complicate the AI-buildout-strains-the-grid narrative (cluster: UC-285/286/288)
FORAGE data_center_left_the_grid
WHERE behind_meter_project_confirmed = true
AND located_inside_strained_grid = true
AND ratepayer_protection_tariff_confirmed = true
ACROSS D6, D2, D4, D5, D1, D3
DEPTH 3
SURFACE data_center_left_the_grid
DIVE INTO strain_as_choice
WHEN dedicated_generation_avoids_shared_grid_draw = true
AND same_region_as_documented_shortfall = true
TRACE opt_out_cascade
EMIT behind_meter_signal
DRIFT data_center_left_the_grid
METHODOLOGY 78
PERFORMANCE 34
FETCH data_center_left_the_grid
THRESHOLD 1000
ON WATCH CHIRP medium 'Meta/Williams New Albany Ohio project: Ohio Power Siting Board approved Jun 9 2025 (Case 24-1449-EL-BGN), 2x 200MW gas units (400MW total), behind-the-meter, targeted online late 2026, inside PJM/AEP Ohio territory - same grid with the documented capacity shortfall. Separately, PUCO approved AEP Ohio data-center tariff Jul 2025: >25MW customers pay for >=85pct subscribed capacity up to 12yrs + exit fees, protecting ratepayers from stranded-cost risk. Neither offsets the multi-GW shortfall at scale, but both show grid strain is a design choice, not an inevitability'
SURFACE analysis AS jsonRuntime: @stratiqx/cal-runtime · Spec: cal.semanticintent.dev · DOI: 10.5281/zenodo.18905193
This isn't an example from a different, less-strained grid — it's inside PJM, in the same AEP Ohio territory feeding the capacity shortfall this cluster documents. The complication is direct, not a comparison across regions.[1]
The behind-the-meter plant addresses physical grid demand. The ratepayer tariff addresses stranded-cost risk. Neither solves the other's problem, and both are real, working regulatory responses.[1][2]
Against a shortfall measured in thousands of megawatts, one project's 400 MW is a genuine but partial complication — this case doesn't claim it resolves the structural gap, only that the gap isn't inevitable.
A hyperscaler chose dedicated generation. A state regulator chose a ratepayer-protection tariff. Both decisions show the strain-vs-no-strain outcome depends on choices being made now, not an unavoidable property of AI infrastructure.
Two sources, held two-sided by design: the Ohio Power Siting Board's own approval order for Meta's behind-the-meter plant, and the Public Utilities Commission of Ohio's own approval of AEP Ohio's data-center ratepayer-protection tariff.
A real, dated, approved project. Not a solution at scale — but proof the strain is a choice, not a law of physics.