AI Data Center Grid Interconnection

The headline finding of this cycle's compile: transmission and interconnection — not generation — is the near-term binding constraint on the AI data center buildout. FERC-filing-level analysis of US interconnection queues finds regional capacity gaps of 50-150+ GW by 2030-2035, with connection timelines now extending 5-10+ years in the most congested regions. This is not a "not enough power plants exist" problem — it is a "power likely exists but can't be delivered to where it's needed, fast enough" problem, and it is the same conclusion this KB's sibling datacenters topic reached independently from a different source base, a rare case of two separately-compiled KB topics corroborating the same structural constraint from opposite ends (grid-supply side here; compute-demand side there).

The industry's workaround, now formalized in research for the first time, is "connect-and-manage" interconnection: a gigawatt-scale AI data center (AIDC) connects to the grid without waiting for conventional transmission upgrades, in exchange for accepting real-time curtailment (power-import reductions) during grid-stress periods. Two companion arXiv papers (same lead author, same submission day) formalize the mechanism and show that a coordinated three-layer request/acceptance protocol between the data center and the transmission system operator (TSO) — plus on-site battery storage as a physical buffer — cuts curtailment from 9.1% to 2.8% while preserving 98.1% of frontier-training workload (batch training absorbs most of the curtailment burden; frontier training stays effectively intact). This reframes on-site battery storage's newest AI-adjacent use case away from "smoothing renewables" and toward a narrower, more urgent one: buffering a data center against its own interconnection limits (see Grid Energy Storage).

On the demand side, Bloom Energy's November-2025 industry survey (n=92-152, the fourth wave of a longitudinal series) independently corroborates the supply-side finding: developers expect power delivery 1.5-2 years earlier than utilities believe they can actually deliver it, and that gap has been widening — specifically in Northern Virginia, the Bay Area, and Atlanta — even as US IT load capacity is projected to roughly double from ~80 GW (2025) to ~150 GW (2028). The strategic response documented in the same survey is that behind-the-meter (onsite) generation is becoming a permanent strategy, not a bridge: ~33% of US data centers are expected to run 100% onsite power by 2030 (up 22% in six months), with fuel cells leading current onsite-technology evaluation at 47%, ahead of reciprocating engines (38%) and mobile turbines (33%). Nuclear/SMR barely registers in that same near-term technology mix — it is a distinct, slower-arriving answer (see Data Center On-Site Nuclear & SMR).

Key Claims

• Transmission/interconnection, not generation, is the near-term binding constraint on AI data center growth. FERC-filing-level analysis across North American grid regions. Evidence: strong (Power Grid Infrastructure for AI Data Centers)
• Regional capacity gaps of 50-150+ GW projected by 2030-2035, varying 2-3x by region — site selection materially changes a developer's time-to-power. Evidence: strong (Power Grid Infrastructure)
• Interconnection-queue timelines now extend 5-10+ years in the most congested regions, per direct FERC-filing data (not survey sentiment). Evidence: strong (Power Grid Infrastructure)
• "Connect-and-manage" interconnection formalized: AIDCs connect without prior transmission upgrades in exchange for accepting real-time curtailment; a hierarchical TSO-acceptance/AIDC-request protocol governs the exchange. Evidence: strong (paper), but validated only in simulation — tested on the IEEE 39-bus system with Australian market data, not a live utility interconnection (Grid Integration under Connect-and-Manage)
• Coordinated curtailment protocol cuts curtailment from 9.1% to 2.8% vs. a naive connect-and-manage baseline, while preserving 98.1% of frontier-training workload (batch training absorbs the curtailment burden instead). Evidence: strong, simulation result (Grid Integration)
• On-site battery storage (BESS) buffers checkpoint-constrained AI training continuity against connect-and-manage power-import limits, substantially increasing the workload a data center can credibly commit to day-ahead. Evidence: strong, simulation result on IEEE 39-bus / Australian market data (Battery-Assisted Operation)
• US IT load capacity projected to roughly double, ~80 GW (2025) to ~150 GW (2028) — more than double 2024-vintage forecasts. Evidence: moderate (industry survey, n=92-152, interested-party-adjacent — Bloom Energy sells onsite fuel cells) (Bloom Energy 2026 Report)
• Developer/utility time-to-power expectations diverge by 1.5-2 years, and the gap is widening in Northern Virginia, the Bay Area, and Atlanta specifically. Evidence: moderate (survey sentiment, not realized deployment) (Bloom Energy 2026 Report)
• Onsite/behind-the-meter generation is becoming a permanent strategy, not a bridge: ~33% of US data centers expected to run 100% onsite by 2030 (up 22% in 6 months), 44% by 2035; fuel cells lead current onsite-technology evaluation (47%), ahead of reciprocating engines (38%) and mobile turbines (33%). Evidence: moderate (survey; Bloom Energy is a fuel-cell vendor, so the fuel-cell-leads finding should be read with that conflict of interest in mind) (Bloom Energy 2026 Report)
• Gigawatt-scale campuses are becoming normal — share of new campuses expected to exceed 1 GW rises from ~1-in-5 (2030) to ~1-in-3 (2035); ERCOT raised its 2030 data-center-growth estimate from 29 GW to 77 GW between 2024 and 2025. Evidence: moderate (survey + grid-operator revision) (Bloom Energy 2026 Report)

Benchmarks & Data

• Regional transmission/interconnection capacity gap: 50-150+ GW by 2030-2035 (Power Grid Infrastructure)
• Interconnection-queue timeline in congested regions: 5-10+ years (Power Grid Infrastructure)
• Curtailment under coordinated connect-and-manage protocol: 9.1% → 2.8% (simulation, IEEE 39-bus) (Grid Integration)
• Frontier-training workload preserved under curtailment: 98.1% (simulation) (Grid Integration)
• US IT load capacity: ~80 GW (2025) → ~150 GW (2028 projected) (Bloom Energy 2026 Report)
• Developer vs. utility time-to-power expectation gap: 1.5-2 years (Bloom Energy 2026 Report)
• Onsite/100%-behind-the-meter expectation by 2030: ~33% of US data centers (up 22% in 6 months); 44% by 2035 (Bloom Energy 2026 Report)
• Onsite-technology evaluation share: fuel cells 47%, reciprocating engines 38%, mobile turbines 33% (Bloom Energy 2026 Report)
• ERCOT 2030 data-center-growth estimate revision: 29 GW → 77 GW (2024 to 2025 vintage) (Bloom Energy 2026 Report)
• Survey sample: n=92-152 core respondents, November 2025 fielding, 84% US-based (Bloom Energy 2026 Report)

Open Questions

• Does connect-and-manage curtailment tolerance become standard PJM/ERCOT/MISO practice beyond the IEEE-39-bus-scale research validation seen so far? No source yet documents a live RTO implementation.
• Will the 50-150+ GW regional capacity-gap projection hold as permitting/interconnection-process reform (not modeled in the FERC-filing analysis) takes effect?
• How much of the projected 80→150 GW US IT-load growth actually materializes given that time-to-power is a demand-side survey expectation, not a confirmed capacity buildout?
• Does Bloom Energy's fuel-cell-led onsite-technology mix (47%) hold as gas-turbine lead times — a KB knowledge gap, see frontier.md — compress relative to fuel cells?
• What is the specific onsite gas-turbine cost-per-MW and lead-time data (xAI Colossus, Homer City, GE Vernova) that would let this concept quantify the dominant current onsite technology, not just the survey-ranked evaluation share?

Related Concepts

• Grid Energy Storage — On-site BESS curtailment-buffering for AI data centers is a new, narrower demand driver distinct from the renewables-integration/grid-firming storage economics this KB has tracked to date.
• Data Center On-Site Nuclear & SMR — The generation-side, multi-year-lagged answer to the same power constraint: nuclear/SMR deals cluster 2030-2035, arriving well after the near-term gas/fuel-cell behind-the-meter response this concept documents.

Backlinks

Pages that reference this concept:

• Data Center On-Site Nuclear & SMR — nuclear as the slower, generation-side counterpart to the interconnection/curtailment mechanisms described here

Changelog

• 2026-07-23 — New concept, first KB coverage of AI-datacenter-specific grid interconnection mechanics. Compiled from 4 sources (2 FERC/simulation-grade papers, 1 companion arXiv pair on connect-and-manage + BESS buffering, 1 industry survey). Headline finding: transmission/interconnection — not generation — is the near-term binding constraint, corroborated independently by this KB's sibling datacenters-topic compile. Cross-linked to Grid Energy Storage (BESS buffering use case) and the new Data Center On-Site Nuclear & SMR concept.