Grid Access, Large-Load Service & Speed-to-Power

In the chip-bound era, the gating question was how many accelerators you could buy. In the power-bound era it is how many megawatts you can energize, and when — and the answer is governed almost entirely by a bureaucratic process most engineers never learn until it has already cost them eighteen months. The interconnection queue is the scarcest asset in the project. A live cluster needs steel, power, and cooling; steel takes 12–18 months and is rarely the long pole. The grid connection—host-utility service request, applicable load studies and agreements, network-upgrade scope, and the transformer that energizes the substation—can take 3 to 7+ years, and in the densest hubs it has stalled past a decade. The mechanics of that process are the subject here: how it works, where it breaks, how the 2024–2026 reforms reshaped it on four continents, and the fork — firm-but-slow vs flexible-but-fast — that every speed-to-power strategy turns on.

The forks here carry direct costs. Site into the wrong regime and your cost-allocation rule flips from "the network socializes upgrades" to "you pay 100% of everything you touch." Insist on firm power and you take the full queue wait; accept curtailable load and you energize years earlier but inherit a kill-switch and a merchant-tail risk. Order the building before the transformer and the slab sits dark for two years past mechanical completion. The grid-impact process is owned here; the macro load-growth story that created this crisis lives in Chapter 16.1, and turning grid flexibility back into revenue lives in Chapter 15.8.

Do not start from one universal interconnection funnel. A data-center project first identifies its actual path: large-load retail or transmission service under the host utility, ISO/RTO and state tariff; a co-located or behind-the-meter arrangement; and, only when on-site generation exports or otherwise interconnects as a generating facility, the applicable generator-interconnection process. Each path has different deposits, studies, cost allocation, milestones, withdrawal rights and contract endpoints.

Large-load service path. Use the host utility or ISO/RTO tariff and state-jurisdiction process: application, load-flow/short-circuit/dynamic studies as required, facilities and network-upgrade scope, then the actual retail-service, transmission-service or large-load agreement. Co-located / behind-the-meter path. Add the applicable withdrawal, standby, netting, curtailment and operating rules. Generation/export path. If the campus generation exports or otherwise qualifies as a generating facility, run the separate generator-interconnection procedure and agreement. Record the governing tariff/order, study endpoint and contract name instead of substituting LGIA terminology for load service.

The binding upgrade scope and cost emerge from the studies required by the governing load-service tariff or utility process. Where that process studies interacting requests, another project's withdrawal can trigger re-study; where it does not, generator-queue mechanics must not be imported. Record the actual study dependency, cost-allocation rule, milestone, and contract endpoint for the site.

FERC Order 2023 reforms generating-facility interconnection. Its cluster studies, readiness requirements, withdrawal penalties and transmission-provider deadlines do not create a national large-load queue or a default data-center study process. They matter directly only to a generating facility covered by the tariff; they are supply-side context for a load project.

Large-load screening is instead being revised through utility tariffs, state rules and market-specific proceedings. Some utilities and RTOs use readiness deposits, site control, duplicate-request screening or flexible-service products, but the authority and contract differ by jurisdiction. Use ERCOT's published large-load population only for ERCOT, and use LBNL's generation/storage queue only for generation/storage. The United States has no comparable national large-load queue dataset; do not transfer generator attrition rates to data-center load requests.

Read the table as a siting screen. The regime is a decision variable you select, and the choice propagates into your cost of upgrades, your firmness, and your schedule. ERCOT remains the structural-advantage market: a single energy-only grid, an interconnection process that historically let large loads energize fast, and an explicit "energize-then-curtail" pathway under SB6 that converts the queue wait into an operating constraint. The cost is the mandate — loads interconnecting after 31 December 2025 must accept curtailment during firm load shed and demonstrate ride-through, with a regulator-imposed kill switch. PJM is the opposite lesson: deep, congested, and slow, with application-to-commercial-operation stretching past eight years and a generation shortfall that has made every new large load politically fraught. The December 2025 FERC PJM order — four transmission-service options and a behind-the-meter cost-causation reform — is the regime trying to make co-location workable rather than a speed promise. SPP's HILLGA/CHILLS/PALS stack is the quiet outlier: explicit non-firm and price-responsive large-load products that reward flexibility. The vertically-integrated utilities are the wild card — one counterparty means it can move fast or capture you entirely, and the line-extension policy often puts 100% of the distribution upgrade on your bill.

Outside the US the process differs in kind, not just degree — and the binding constraint is frequently a hard moratorium rather than a queue you can pay to jump. The decision is no longer "firm vs flexible" but "is new large-load connection legally available at all, and on what conditions."

EU / UK. Great Britain's connections queue, managed by the National Energy System Operator (NESO, the former National Grid ESO) under Ofgem, ballooned to hundreds of GW and prompted a 2025 reform that culls stalled "zombie" projects and re-orders the queue by readiness — the same first-ready logic as Order 2023, arrived at independently. Ireland is the cautionary extreme: EirGrid and the CRU imposed a de-facto moratorium on new Dublin-region data-center grid connections in 2021 as the sector approached ~20% of national electricity demand. In December 2025 the CRU lifted the blanket ban — but replaced it with arguably the world's strictest conditional-access regime: a new connecting data center must bring on-site generation or storage capable of meeting its full demand, must be able to export power back to the grid when the system needs it, and must procure renewable generation covering at least 80% of its annual demand. Ireland did not re-open the door so much as redefine "interconnection" to mean "bring your own firm power and grid services." The Nordics (Norway, Sweden, Finland) remain the relief valve — firm hydro, cold-climate free cooling, and comparatively available transmission — though Norwegian and Swedish grid operators have begun prioritizing connections and questioning data-center load. The Middle East (Saudi HUMAIN, UAE) and India offer state-coordinated, fast-tracked access where sovereign-AI ambition aligns connection policy with the project. APAC is moratorium-shaped: Japan faces transmission constraints around Tokyo/Osaka; Singapore ran a hard moratorium from 2019, and its successor regime — the December 2025 DC-CFA2 call allocating at least 200 MW — rations capacity by efficiency and sustainability bid rather than first-come; Malaysia (Johor) absorbed the spillover and is now itself power- and water-constrained.

A 2026-specific risk sits on top of the regime map: the rules may change mid-queue. In October 2025 the Secretary of Energy invoked Section 403 of the DOE Organization Act — a rarely-used authority to direct FERC — ordering the Commission to open a rulemaking to accelerate the interconnection of large loads (generally those above 20 MW) to the interstate transmission system. FERC created docket RM26-4, took comments through late 2025, and on 18 June 2026 — rather than issue a nationwide rule — opened Section 206 show-cause proceedings against all six RTOs/ISOs, directing each to justify or reform its large-load interconnection, cost-allocation, and transmission-service tariff, with the RM26-4 docket left open for possible later action. The clock then slipped: all six RTOs won ~90-day abeyances in mid-August 2026, pushing tariff filings toward late 2026. PJM's Board did not wait — on 2026-07-27 it directed a Large Load Registry (sites ≥50 MW, location and supply status tracked) and an Interim Resource Adequacy Service under which new large loads that do not bring their own generation by 2027-06-01, and have not otherwise secured supply, are curtailed before Pre-Emergency Load Management — a regional first-to-curtail stack, Board-directed but not yet a FERC-accepted tariff. And Texas showed the rules can freeze mid-queue as well as change: the governor's 2026-08-03 directive paused ERCOT's Batch Zero large-load process outright pending a project-by-project audit (~250–300 projects, several months). The contemplated reforms include standardized large-load study procedures, a target study timeline for flexible loads, and joint co-located load-plus-generation filings.

The collision is jurisdictional. Large-load interconnection has traditionally been state and utility territory; the DOE/FERC move asserts federal interest in standardizing it — and several states have pushed back, protective of their authority over retail load, siting, and cost allocation. For a developer the practical consequence is regulatory tail risk: an interconnection agreement negotiated under today's state or RTO rules could be reshaped by a federal rule landing mid-construction, changing your cost-allocation, your flexibility obligations, or your study timeline. The mitigation is to build optionality into terms (flexibility you can offer rather than have imposed), to track RM26-4 as a live schedule item, and to avoid betting the project on a cost-allocation rule that a 2026 final rule might overturn. → cross-reference the macro driver in Chapter 16.1.

The firm-but-slow / flexible-but-fast fork resolves, in practice, into an energization sequence. Rather than waiting years for firm service to the full campus peak, the operator energizes in tranches and covers the gap with flexibility and bridge generation. The pattern, now standard at the GW-scale frontier:

• Phase 1 — bridge power. While the interconnection and network upgrades grind through study, energize the first 50–100 MW tranche on behind-the-meter generation (gas reciprocating engines or aeroderivative turbines), often standing up the first IT load on a project-specific schedule before firm grid service would arrive. → the bridge-power thesis and generation choices are Chapter 3.5; electrical integration is Chapter 4.8.
• Phase 2 — flexible grid tie. Take a curtailable / non-firm interconnection (ERCOT SB6 energize-then-curtail, SPP CHILLS, a utility interruptible tariff) that connects without waiting for the firm upgrade, accepting curtailment during system stress and covering curtailed hours from the bridge plant.
• Phase 3 — firm grid. When the network upgrade and the long-lead transformer finally arrive, transition to firm service and re-purpose or retire the bridge generation (or keep it as backup / grid-services capacity → Chapter 15.8).

Energized capacity leads commissioned IT load. A campus may have 100 MW energized while only 40 MW of IT is commissioned, because energization is staged ahead of the fit-out. The whole sequence exists to defeat one number — the multi-year queue wait — by converting it from a hard schedule gate into a manageable operating constraint. Offering flexibility proactively is almost always cheaper than having it imposed: the operator who designs for curtailment captures the speed; the one who insists on firm and is later forced to curtail anyway pays twice.

Winning the queue does not energize the building. The host utility's applicable large-load studies identify facilities and network upgrades—new lines, reconductoring, or substation expansions—and those upgrades can themselves be multi-year builds. US high-voltage (345 kV-plus) transmission additions have collapsed from ~1,700 circuit-miles/year in 2010–2014 to ~350 miles/year in 2020–2023 — and just 55 miles in 2023, even as load growth demands the opposite. The structural answer is higher-voltage backbones. ERCOT's 765 kV program is the headline example: the PUCT-approved Permian Basin import paths (three 765 kV paths, ~$13.8B, roughly 800–1,000 route-miles, approved April 2025) sit inside a statewide ~2,500-mile, ~$33B 765 kV expansion, whose proposed 2024-RTP Phases I+II total ~4,481 miles and ~20 substations. A single 765 kV circuit moves roughly the power of three double-circuit 345 kV lines in less right-of-way — the only way to add bulk transfer capacity fast enough. Where new lines are infeasible on the timeline, grid-enhancing technologies (dynamic line rating, advanced reconductoring) unlock 30–50% more capacity on existing corridors at roughly 10% of new-line cost — the speed-to-power tactic when you cannot wait 4–8 years for a new line.

And even with the line built, the project waits on iron. The long-lead transformer is a co-equal critical path with the queue itself: HV power-transformer lead times have stretched to ~128 weeks standard, ~144 weeks for generator step-up units, and up to ~60 months in constrained markets — with incumbents (GE, Siemens, Hitachi/Mitsubishi) quoting multi-year backlogs. Order the building and discover the transformer is three years out, and the slab sits dark long past mechanical completion. Order the gate, not the building, first: place the transformer and switchgear orders against the interconnection in parallel with — or ahead of — civil works, and treat the long-lead register as a board-level schedule artifact. → the full equipment supply chain and lead-time register is Chapter 2.3.

The forks above compound into a single speed-to-power posture, and mismatching them is expensive. Three pairings recur:

• Regime × firmness. ERCOT-flexible is a coherent posture (fast queue + energize-then-curtail + bridge gas); PJM-firm is coherent but slow; PJM-flexible is constrained by the co-location order; an Irish or Singaporean site is firmness-mandated regardless of preference. Picking a regime implies a firmness option set — you do not get to choose freely once the site is fixed.
• Flexibility offered vs imposed. The operator who designs for curtailment and ride-through from day one captures the speed and can monetize the flexibility (Chapter 15.8); the one who insists on firm and is later subjected to SB6-style mandates pays for the firm upgrade and eats the curtailment. Offer it; do not wait to have it taken.
• Gate ordering. The queue, the transmission upgrade, and the long-lead transformer are three serial gates that must be worked in parallel. Sequence them around the slowest — usually the transformer or the network upgrade — not around the building.

The downstream economics of all three — nodal pricing, congestion exposure, network-upgrade cost allocation, and the "who pays" question — are quantified in Chapter 3.3; the supply strategy that monetizes the firm/flexible split (PPAs, behind-the-meter islands, co-location) is Chapter 3.4.