For the state utility commission staffer or consumer advocate who must judge a pending large-load tariff, and for the developer’s finance lead who must price against one. It is a data center tariff benchmark: the central terms of 27 US large-load tariffs, orders and filings, so you can see where a proposal sits.
Bottom line
If the tariff in front of you asks a data center to pay for 80 percent of its contracted capacity for 12 to 15 years, it sits at the middle of the 27 large-load tariffs in this benchmark. The median minimum take is 80 percent in every cut of the data: all 27, the 15 adopted and the 12 pending or non-binding. The median term is 12 years among the 15 adopted tariffs and 15 years across all 27. [High: read from filings, orders and tariff sheets] A minimum take outside 75 to 90 percent is unusual here: 14 of the 15 adopted tariffs and 25 of 27 overall sit inside it. A proposal outside that band should come with evidence from the utility, not a convention. [Juncture judgment]
I tested whether market structure, flexibility options, capacity markets or the year of decision explain the terms. In the full samples, none cleared the bar I set in advance. That is a limit of the rows, not proof that nothing matters. The data describe what regulators wrote. They do not show whether any tariff worked.
What the 27 tariffs show
I collected 49 large-load tariffs, orders, rules and filings. Two Pennsylvania rows rest on one commission model tariff, so they are now one row, which leaves 48. I kept the 27 where I could quote both a numeric minimum take and a numeric minimum term from a source I opened. They cover 19 states and 17 independent families. A family is a group of sister utilities that tend to copy each other, such as American Electric Power (AEP) or Xcel Energy. The other 21 rows are out: 18 gave no number, usually because the terms sit in confidential special contracts or in each customer’s service agreement, and three failed on a second read of the document. Montana-Dakota’s Rate 45 uses an 85 percent expected load factor as an eligibility test, not a billing floor. Entergy Mississippi’s ALGS-10 is a general schedule for customers of 5,000 to 35,000 kW, not a large-load tariff. New Jersey’s Chapter 32 directs the state board to require guarantees but sets no tariff yet. [High]
15 of the 27 are adopted: a commission order, an effective tariff or an approved special contract. 12 are pending or non-binding: filings and proposals that no commission has finally decided, plus Pennsylvania’s model tariff, which the commission issued as guidelines. I report both groups because a reviewer deciding a final tariff should compare it with adopted terms first.
Minimum take. This is the share of contracted capacity the customer pays for whether or not it uses it. Across all 27 rows the median is 80 percent, the mean is 81.9 and the range is 70 to 100. Among the 15 adopted rows the median is 80, the mean is 81.3 and the range is 70 to 90. The single 100 is Texas, which is a proposal (see below).
| Minimum take (%) | All 27 tariffs (count) | Adopted 15 (count) | Pending or non-binding 12 (count) |
|---|---|---|---|
| 70 | 1 | 1 | 0 |
| 75 | 3 | 0 | 3 |
| 80 | 14 | 10 | 4 |
| 85 | 5 | 2 | 3 |
| 90 | 3 | 2 | 1 |
| 100 | 1 | 0 | 1 |
| Total | 27 | 15 | 12 |
Source: Juncture dataset of 48 collected US large-load tariffs, orders and filings (27 verified rows), tariffs_v3.csv and results_v3.txt, built from the commission, utility and legislative documents linked in this post. Accessed and re-run October 8, 2026.
Term. Across all 27 rows the median minimum term is 15 years (mean 14.1, range 5 to 20). Among the 15 adopted rows it is 12 years (mean 13.7, range 8 to 20). The pending rows carry the longer terms: 9 of the 12 ask for 15 or 20 years.
Term is coded as the tariff writes it, and tariffs write it differently. 9 state a period after the load ramp (AEP Ohio: the ramp plus 8 years), 6 include the ramp inside the term, and two give a size-scaled or option-based range, where I coded the lowest figure (Portland General Electric: 10 years at 20 MW, rising to 30 years at 220 MW; Public Service Company of Oklahoma: 10 to 15 years by option). If I add each ramp-additive tariff’s maximum ramp, the median becomes 15 years across all 27 and 17 among the adopted. Read the term figures as a band: 25 of the 27 rows run 10 to 20 years, with a middle near 12 to 15.
| Minimum term (years) | All 27 tariffs (count) | Adopted 15 (count) | Pending or non-binding 12 (count) |
|---|---|---|---|
| 5 | 1 | 0 | 1 |
| 8 | 1 | 1 | 0 |
| 10 | 4 | 2 | 2 |
| 12 | 5 | 5 | 0 |
| 14 | 2 | 2 | 0 |
| 15 | 8 | 2 | 6 |
| 19 | 1 | 1 | 0 |
| 20 | 5 | 2 | 3 |
| Total | 27 | 15 | 12 |
Source: Juncture dataset of 48 collected US large-load tariffs, orders and filings (27 verified rows), tariffs_v3.csv and results_v3.txt, built from the commission, utility and legislative documents linked in this post. Accessed and re-run October 8, 2026. Term is the number in the tariff as written; see the note on ramp periods.
Collateral and flexibility. Collateral is required in 26 of the 27 rows, and Portland General Electric has no entry. Every row but one records an exit-fee rule; El Paso Electric’s tariff states none and defers to each service agreement. Because nearly every row has both, neither can be tested against the size of the take. 21 rows offer an interruptible, curtailable or capacity-reduction option, 3 do not (AEP Ohio, Portland General Electric, Florida Power & Light) and 3 are not recorded.
Examples a reviewer can check against the source.
- AEP Ohio’s Schedule DCT, filed in compliance with the commission’s July 9, 2025 order (this copy is hosted by the Ohio Capital Journal), sets a bracket formula and says “the minimum demand will not exceed 85% of the total contract capacity.” The term is “the Load Ramp Period plus 8 years,” which is 12 years for a four-year ramp. [High]
- Kentucky Power’s order approved its proposal as modified, for customers adding 150 MW or more. The proposal sets a minimum billing demand of at least 90 percent of the customer’s highest established demand, a 20-year term and a one-time payment equal to five years of minimum billing on permanent closure. [High]
- Florida Power & Light’s large-load rates carry a 70 percent take-or-pay demand charge and a minimum term of no less than 20 years from the in-service date, including the ramp, per the commission staff recommendation and the company overview. The commission unanimously approved the settlement containing them on November 20, 2025, as Order PSC-2026-0117-FOF-EI recites. It is the lowest take in the sample, paired with a 20-year term, the top of the range. [High]
- Dominion Virginia’s order sets 85 percent for transmission and distribution but 60 percent for generation, and a 14-year term. [High]
- Michigan’s Consumers Energy terms, per the commission’s release, include an 80 percent minimum, a 15-year term, and default collateral equal to half of the exit fee. [High]
- DTE Electric’s special-contract order sets the minimum billing demand at 80 percent of contract capacity for about 19 years. It is a contract approval for one customer, not a general tariff. [High]
Not one number. “Minimum take” is not defined the same way everywhere. Dominion splits it by service type. AEP Ohio’s 85 percent is a cap on a bracket formula. Kentucky Power’s 90 percent applies to the customer’s highest established demand, not to contract capacity. Texas’s proposed floor is the greater of contracted peak, highest non-coincident peak and 12-coincident-peak demand. PPL Electric Utilities in Pennsylvania pays 80 percent of the ramp-schedule load for the first five years and 50 percent for the next five, under Rate Schedule LP-6. In 17 of the 27 rows the floor is a plain share of contract capacity. The clustering at 80 and 85 partly reflects drafting convention. [High for each tariff; Juncture judgment on the inference]
Tests I committed to before running them
I wrote six hypotheses and the decision rule into a pre-registration file before any test. A result counted as supported only with p below 0.05 and a bootstrap 95 percent interval that excluded zero in the predicted direction. A dated amendment on October 1, 2026 corrected two counts and the coding of the Texas row. A second, on October 8, 2026, followed my re-reading of the filings: it removed three rows, merged two, moved one row’s regional market from MISO to SPP, recoded two terms to the figure the tariff states, and corrected the script’s exit-fee test, its handling of missing collateral entries, the direction of its one-sided H6 p-value and its family list. I made those changes after seeing the first results, which were all null, and they are listed in the dataset’s change log. The hypotheses and decision rules did not change. None of the six cleared the bar in either sample.
| Test | All 27: estimate | All 27: p | Adopted 15: estimate | Adopted 15: p |
|---|---|---|---|---|
| H1 restructured states demand more (median take, %) | 80 vs 80 (n = 7, 20) | 0.161 | 80 vs 80 (n = 4, 11) | 0.531 |
| H2 longer term, lower take (Spearman rho) | 0.036 | 0.857 | -0.265 | 0.340 |
| H3 flexibility, shorter term (median years) | 14 vs 10 (n = 21, 3) | 0.478 | 12 vs 10 (n = 11, 3) | 0.424 |
| H3 flexibility, lower take (median %) | 80 vs 85 | 0.673 | 80 vs 85 | 0.585 |
| H4 collateral and exit fee vs take | Collateral: no variation. Exit fee: 1 of 27 lacks one | 1.000 | No variation to test | n/a |
| H5 lower take with a capacity market (median %) | 80 vs 80 (n = 14, 13) | 0.479 | 80 vs 80 (n = 11, 4) | 0.734 |
| H6 later decision, higher take (Spearman rho) | 0.171 | 0.197 | 0.083 | 0.384 |
Source: Juncture analysis, results_v3.txt, script analyze_v3.py run October 8, 2026 (seed 20261001, 20,000 bootstrap resamples). One-sided p for H1, H5 and H6; two-sided for H2 and H3. In the 15 adopted rows every row records both collateral and an exit fee, so H4 cannot be tested there. In all 27, one row (El Paso Electric) records no exit-fee rule, and Fisher exact on exit fee against take is p = 1.000.
Reading the table. Every null is “not supported at this n,” not “refuted.” Take clusters at 80 and 85, so the tests have little variation to work with. Four points sharpen it:
- H1: coding Michigan as vertically integrated moves the restructured median to 85, with p = 0.073 (all 27 rows). Still not supported.
- H2: on the 10 independent families among the adopted rows the correlation is negative, rho = -0.547, with an interval from -0.92 to 0.19 and p = 0.102. It points the predicted way and does not clear the bar.
- H6: every row is dated 2025 or 2026, so the window is too short to detect a trend. Among adopted rows the slope is 1.6 points per year, p = 0.590.
- H6 in one exploratory cut: on the 17 rows whose floor is a plain share of contract capacity, rho = 0.466 (one-sided p = 0.030, interval 0.07 to 0.75), which meets the rule. I chose that cut after seeing the main results, it is one of several, and the dates mix decision dates with filing dates for pending rows. I do not count it as support.
I repeated the tests on one row per family (17 families for all rows, 10 for the adopted rows) and without the DTE special contract. No conclusion changed. [High: computed by script, output in results_v3.txt]
12 of the 27 are pending or non-binding
12 rows are filings, proposals or guidelines, not final decisions. The table lists them.
| Utility | State | Minimum take (%) | Term (years) | Instrument |
|---|---|---|---|---|
| Duke Energy Carolinas and Progress | NC | 75 | 15 | Technical-conference filing |
| ERCOT large loads (Texas) | TX | 100 | 20 | PUCT proposal for publication |
| OG&E | OK | 85 | 15 | Filed application |
| Public Service Company of Oklahoma | OK | 80 | 10 | Filed stipulation |
| Xcel Energy (Northern States Power Wisconsin) | WI | 75 | 15 | Filed tariff |
| NorthWestern Energy | MT | 75 | 15 | Filed application |
| Pennsylvania PUC model tariff | PA | 80 | 5 | PUC model tariff, guidelines |
| Potomac Edison | MD | 85 | 10 | Filed tariff |
| El Paso Electric | TX | 85 | 20 | Filed tariff, suspended |
| Xcel Energy Colorado | CO | 80 | 15 | Filed application, suspended |
| Xcel Energy Southwestern Public Service | TX | 80 | 15 | Filed tariff |
| Duke Energy Florida | FL | 90 | 20 | Filed petition |
Source: Juncture dataset, tariffs_v3.csv, with the Texas row checked against the PUCT Project 58000 docket on October 8, 2026. Other rows are as filed; docket outcomes after each filing were not re-checked. The PSO term is the lowest of its three options (10 to 15 years).
Three examples:
- OG&E’s filing in Oklahoma, filed June 17, 2026 as testimony with the tariff, sets the billing demand floor at no less than 85 percent. Its testimony puts the minimum term at 15 years, which follows a load ramp of up to five years. [High: as filed]
- PSO’s joint stipulation exhibit says billing demand “shall be not less than 80% of Contract Demand.” The term is not less than 10 years under Option 1 and 15 years under Option 3, which may include a ramp of up to five years. [High: as filed]
- Pennsylvania’s final order of April 30, 2026 issues a model tariff as guidelines (“should”), not a tariff any utility must file. It sets a minimum demand charge of 80 percent of contracted demand and an initial term of not less than five years after any three-to-five-year load ramp. [High]
The adopted-only columns of the tables do not depend on these rows, and the test results are null with them in or out.
Texas. The Texas row is coded as a 100 percent minimum for 20 years, from the Public Utility Commission of Texas (PUCT) rule text in Project 58000. It is a proposal, not an adopted rule. The document I opened is headed “Project No. 58000 Proposal for Publication” for the July 9, 2026 open meeting. It would require a large load customer’s minimum billing demand to be the greater of its contracted peak demand, its highest non-coincident peak demand in the past year, or its 12 coincident peak demand, billed “for a period of no less than 20 years.” Comments were due August 11, 2026, and the notice itself asks whether the floor should be “reduced in phases over a period of 20 years.” The Project 58000 docket, checked on October 8, 2026, lists 70 filings, the latest dated October 1, 2026, and no order adopting the rule. A separate rulemaking, Project 58481, produced an order adopting new 16 TAC section 25.194 on large load interconnection standards, filed September 18, 2026. The minimum billing demand is in the proposed amendments to section 25.193, not in that order. [High on the text and the docket; the final rule may differ]
The 100 is the one extreme value in the all-row sample. Because it is a proposal, it is excluded from the adopted-only range, which tops out at 90.
An analyst’s comparison with take-or-pay in power purchase agreements
This section is my own view. It is not a finding from the dataset and it makes no claim about any specific project. [Juncture judgment]
A take-or-pay contract makes the buyer pay for contracted capacity or energy whether or not it takes delivery, so that a lender can finance the plant against that revenue. Independent power producers have long used this structure in power purchase agreements. The data-center tariffs borrow its logic: a long term, a payment floor and credit support behind it.
Two differences matter for a reviewer. First, the lender’s role. In a project-financed power purchase agreement the lender prices the buyer’s credit and tests the terms before committing. In a tariff the party exposed if the customer fails is the utility’s other customers, and the commission stands in for their consent. Second, price discovery. A lender-tested negotiation tends to produce reasoned terms. A tariff minimum is set in a proceeding where, the number can be a negotiating position rather than a measured loss. Pennsylvania’s commission was explicit that it “tentatively adopted approximately 80% of contracted demand as an opening position,” in a final model-tariff order dated April 30, 2026 that notes there was no consensus on the number. [High on the quote]
What to watch
- Final adoption of the Texas rule. Decision frame: if the final rule keeps a floor at 100 percent of contracted peak demand for 20 years, treat Texas as the upper outlier in your benchmark and ask whether your docket should copy it. If the commission phases the floor down, as its own notice asks about, move Texas into the 80 to 90 percent cluster and recompute. Do not cite Texas as adopted until the final text is in hand.
- Whether the other 11 pending rows are adopted as filed. Decision frame: if most are approved within 75 to 90 percent, the benchmark hardens and a proposal at the median needs little extra scrutiny. If commissions cut them materially, treat 80 percent as an opening position, as Pennsylvania called it, and ask for the utility’s loss evidence before accepting it.
- Outcome reporting. No dataset I know of records how many contracted customers reached full load, how many paid exit fees, or how much collateral was drawn. Decision frame: if your docket can require annual reporting of these items, add it now. If it cannot, record in your order that the terms rest on convention and that no outcome test exists.
- Term length in new filings. Decision frame: the adopted median is 12 years and the all-row median is 15, and the gap is the pending rows. If they are adopted as filed, the adopted median moves toward 15. If new filings cluster at 15 years or longer, ask each utility how the asset life of the plant built for the customer compares with the term. If they cluster at 10 to 12, ask the same question the other way.
For the wider cost-allocation fight, see FERC Large-Load Tariff Reform: Who Pays for the Grid? and the AI Infrastructure and Sovereignty stream.
What this is not
This is not a finding that any tariff is too weak or too strong, a forecast of how any pending case will be decided, or a legal opinion on any commission’s authority. It is a benchmark of written terms.
Limits
- Small n. 27 rows, 7 restructured, 3 with no flexible option, 14 with a capacity market; 15 adopted and 10 independent adopted families. Every null is “not supported at this n.”
- Dates. All 27 rows are dated 2025 or 2026. For pending rows the date is the filing date, not a decision date. The sample cannot speak to longer-term change.
- Proposals. 12 rows are proposed tariffs, proposed rules or guidelines, and the Texas row could change on adoption. Status for each pending row is as of its filing; only Texas was re-checked against its docket on October 8, 2026.
- Publication selection. Tariffs with clean numeric terms are the ones regulators publish. Negotiated special contracts (Entergy Louisiana, Entergy Arkansas, Rocky Mountain Power in Utah) are confidential and failed my verification. The sample over-represents tariffed deals.
- Families. AEP supplies six rows (including Public Service Company of Oklahoma), Xcel four, and Evergy and Duke two each.
- Coding. Minimum take and term are coded values and the definitions differ, as the “Not one number” paragraph shows. The coding rules are fixed in the pre-registration file and its amendments.
- Terms, not outcomes. Nothing here measures whether any tariff covered a cost or protected a ratepayer.
- Verification. On October 8, 2026 a script matched the recorded minimum-take and term quote for each row to the text of the source document, and I opened the documents behind every row I changed or excluded. Documents the script could not fetch were opened separately. Only the Texas dockets were re-checked for status. An independent reviewer re-checked the dataset coding, the script output and the tables on October 8, 2026. The post has had no external peer review.
Drafted with AI assistance. Quotes were matched against the linked sources by script and by reading on October 8, 2026.