Flexible Load in ERCOT:
What Texas Built, What It Paid,
and What It's Now Rewriting

On July 22, 2026, ERCOT served 91,133.73 megawatts in the hour ending 6 p.m. It was the highest electricity demand in the grid's history, roughly 5,600 MW above the previous record set on August 10, 2023, and no Energy Emergency Alert was issued.

That fact has been used all summer as evidence that the Texas grid has been fixed. The evidence for that reading is thinner than it looks, and the reason is worth sitting with before going any further.

ERCOT forecast a higher peak than the one that arrived. Its own July 2026 projection was 92,783.59 MW, and actual demand came in about 2% below it. The grid operator's Monthly Outlook for Resource Adequacy for July 2026 put the probability of committed capacity falling below the 2,500 MW emergency threshold at 0.00% for every hour from 1 a.m. through 8 p.m. The record was not a surprise the system survived. It was a Wednesday that landed slightly under forecast.

There's a second problem. The number itself has been reported at least three ways. A widely circulated figure of 91,308 MW originates with a data analytics vendor and matches nothing ERCOT measures on any basis, including ERCOT's own 15-minute interval maximum of 91,262.71 MW for the same day. ERCOT's yearly records page carried 91,089 MW as late as August 18, 2026, a superseded preliminary value. ERCOT states plainly that it uses "the integrated system load for the full hour" and "does not count instantaneous loads for demand records."

None of this makes the record uninteresting. It makes the growth rate the interesting number. ERCOT's July 2025 peak was 81,733.65 MW. July 2026 was 91,133.73 MW. That is a 12% increase in a single year on the same measurement basis, which is the kind of load growth a grid built for slower change has to answer for.

And Texas is answering for it right now, in three separate proceedings that will each change what a flexible load is worth. That is what this document is actually about.

ERCOT is the only electricity market that has run a complete cycle on interruptible large load. It built a registration pathway. It priced participation. It paid operators money that shows up in audited financial statements. And having done all that, it is now rewriting the transmission charge, the interconnection process, and the co-location rules, because the load arriving turned out to be far larger than the terms were written for.

The repricing is not a footnote to the success story. It is the most useful part of it, and it is the part that transfers. Any market about to absorb tens of gigawatts of computational load will face the same sequence: prove the resource works, pay for it, then discover the payment terms don't scale. Texas is simply further along.

Winter Storm Uri, with the numbers kept apart

Every account of Texas grid reform starts in February 2021, and most of them conflate three different measurements.

The federal after-action report, produced jointly by FERC, NERC and the regional entities in November 2021, keeps them separate. ERCOT shed 20,000 MW of firm load "at the worst point of the Event." That is a peak simultaneous figure, corroborated by ERCOT's own February 24, 2021 presentation, which records a maximum load shed request of 20,000 MW sustained across 70.5 hours. Separately, ERCOT averaged 34,000 MW of generation unavailable from 7 a.m. on February 15 to 1 p.m. on February 17, which the report frames as "equivalent to nearly half of its all-time winter peak electric load of 69,871 MW."

That 34,000 MW is a two-day average, not a peak. The report contains no ERCOT peak-unavailable-capacity figure at all, and any sentence presenting 34,000 MW as a maximum is unsupported by the source it usually cites.

The most repeated detail is also the most misunderstood. ERCOT came "four minutes and thirty-seven seconds" from collapse, the story goes, as though a countdown were running. It was the opposite: 4:37 was the margin remaining. Frequency fell below 59.4 Hz at around 1:51 a.m. on February 15. Generator under-frequency relays would have tripped if it stayed there for nine minutes. It stayed for roughly four minutes and twenty-three seconds before rolling blackouts brought it back. The 4:37 was what was left. ERCOT President and CEO Bill Magness presented the figure to the Board of Directors on February 24, 2021, and the federal report corroborates the mechanism without the number: operators "had only nine minutes to prevent approximately 17,000 MW of generating units from tripping due to underfrequency relays."

The human cost is documented by the Texas Department of State Health Services at 246 deaths, counting deaths occurring between February 11 and June 4, 2021 across 77 counties, classified as 148 direct, 92 indirect and 6 possible, with hypothermia the leading cause at 158.

The economic cost is not documented so much as estimated, across two orders of magnitude. The only figure produced by a government body is the Federal Reserve Bank of Dallas calculation of $4.3 billion in value of lost load over 70.5 hours. Insured losses have been put at $10–20 billion and total direct and indirect losses at $80–130 billion, both by private forecasters. The widely quoted "$300 billion" is the Texas Section of the American Society of Civil Engineers saying costs "could reach" that figure, in a report released on the storm's first anniversary. These measure different things and none of them is the cost.

The correction that matters most about Uri

The version of this story told most often holds that ERCOT had no meaningful registered demand response in 2021, and built it afterwards. ERCOT's own event timeline says otherwise, with timestamps:

• 12:15 a.m., February 15: Emergency Operations Level 1, reserves below 2,300 MW
• 1:07 a.m.: Emergency Operations Level 2, reserves below 1,750 MW, “Load Resources Deployed”
• 1:23 a.m. to 1:33 a.m.: EEA3 entered, first 1,000 MW of firm load shed ordered

Demand response was dispatched roughly sixteen to twenty-six minutes before the first firm load shed order. Research by the University of Texas at Austin Energy Institute, authored by Carey King, Joshua Rhodes, Jay Zarnikau and colleagues, puts the delivered quantities at roughly 1,100 MW from Emergency Response Service on the morning of February 15, and maximum Load Resource reductions of over 1,400 MW on February 15, 16 and 17.

Against a 20,000 MW shed, that is small. But it was there, it went first, and it is documented. The honest lesson from Uri is not that Texas had no demand response. It is that Texas had roughly 2,500 MW of it against a shortfall eight times larger, and spent the next five years finding out how much more the market would actually build.

The answer, as Part one now shows, is less than the rhetoric suggests.

Load Resources, and the distinction almost everyone misses

ERCOT's Nodal Protocols define a Load Resource as "a Load capable of providing Ancillary Service to the ERCOT System and/or energy in the form of Demand response and registered with ERCOT as a Load Resource."

Within that category sits a narrower one. A Controllable Load Resource is "a Load Resource capable of controllably reducing or increasing consumption under Dispatch control by ERCOT." Note the bidirectionality. A CLR is not simply a load that can switch off. It follows a continuous ERCOT dispatch set point and provides proportional primary frequency response, in both directions.

Everything else is a Non-Controllable Load Resource, which trips a block of load either on a manual instruction or autonomously through an under-frequency relay. The NCLR is a switch. The CLR is a dial.

That distinction determines what a load can sell:

The scale nobody reports

Here is where the public narrative and ERCOT's own numbers part company. In April 2026 ERCOT's Board of Directors received a strategic discussion paper on resource adequacy and the role of demand response. Table 1 of that paper reports qualified capacity by ancillary service product for 2025:

Controllable Load Resources are roughly 3% of qualified Responsive Reserve capacity, provided by six resources. The overwhelming majority of load participating in ERCOT's ancillary service markets is non-controllable. It is blocky, relay-tripped, and closer to a large industrial interruptible contract than to a dispatchable asset.

Demand response in aggregate

Across all programs, ERCOT's April 2026 board materials put demand response at 4.8% of the grid's 2024 peak demand, roughly 4,099 MW. That is a survey-derived enrollment figure spanning ERCOT-administered and retail programs, not firm dispatchable capacity, and it describes reporting year 2024. The maximum recorded peak reduction during summer 2024 was 4,949 MW.

Registered Load Resources have been reported at "600+, approximately 7,000 MW" in September 2021 and "575+, approximately 8,300 MW" in April 2023. Utility-run programs added roughly 290 MW for summer 2025. ERCOT's aggregated distributed energy resource pilot had 107.7 MW qualified as of February 2026.

ERCOT procures five reserve products. Their differences are entirely about how fast and for how long.

ECRS launched in June 2023 and was the first new ERCOT ancillary service in roughly two decades.

Now the error. Emergency Response Service is not an ancillary service, and describing it as one is the single most common technical mistake in writing about ERCOT demand response.

ERS is authorised under 16 Texas Administrative Code § 25.507 as "a special emergency response service to be deployed by ERCOT to help prevent or alleviate an actual or anticipated Energy Emergency Alert event." It is procured by request for proposals under four fixed contract terms a year, not co-optimised in the day-ahead or real-time market. Its spending is capped at $75 million per twelve months, plus up to $25 million for renewals. It has been deployed eleven times since its inception in 2007.

ERS-10 and ERS-30 are response times, ten minutes and thirty minutes. They are not megawatt quantities, though they are frequently reported as if they were. Cleared ERS capability for the December 2025 to March 2026 term ranged from 2,323 to 2,885 MW across eight time periods, at clearing prices between $0.10 and $26.87 per MW against an $80 per MW-hour cap.

And this is where the crypto-mining share actually belongs: cryptocurrency mining loads made up 64% of ERS resources procured for that December 2025 to March 2026 term. That figure is real. It describes ERS, not Controllable Load Resources, and the two get conflated constantly.

The new product that excludes load entirely

In June 2026 the ERCOT Board adopted NOGRR283 and NPRR1309, creating Dispatchable Reliability Reserve Service. The Public Utility Commission of Texas approved both at its July 9, 2026 open meeting. The Legislature had directed ERCOT to create the product through House Bill 1500 in 2023, with an original statutory deadline of December 1, 2024 that was missed.

DRRS competitively procures reserves for the hours when net demand runs above forecast, which in practice means evenings when solar output falls and demand does not. That is precisely the problem flexible load is supposed to be good at.

ERCOT's own workshop materials are unambiguous about who may provide it: "DRRS can be provided by eligible Off-Line Generation Resources and On-Line Generation Resources." Energy storage may be added later through a separate revision request. Load Resources are not eligible. And the service is not live: ERCOT's June 2026 update to its Board places implementation in 2028.

So the newest reliability product in the market, created by legislative direction to solve the evening net-load problem, is one that flexible load is shut out of. That is a real limit, and any honest account of the opportunity has to carry it.

What real-time co-optimization changed

On December 5, 2025 ERCOT launched Real-Time Co-optimization plus Batteries, the largest change to its market design in roughly two decades. Energy and ancillary services are now cleared together every five minutes, with Ancillary Service Demand Curves setting both the quantity procured and the clearing price.

Under the new design, ERCOT states that “participation of CLRs in DAM will be similar to Generation Resources.” That is the clearest statement available of where controllable load sits in ERCOT’s own conception: not as a special category of demand, but as a resource that clears alongside generators.

For load resources specifically, four things changed. All ancillary service dispatch became resource-specific, replacing group assignments, which ERCOT describes as the single largest operational change for participants. Qualified scheduling entities can now update ancillary service offers for an operating hour within that hour. Proxy offers are generated where a resource shows online status with headroom but no submitted offer. And a new self-provision mechanism exists for under-frequency-relay load, validated through real-time telemetry.

Worth noting what the launch announcement did not say. ERCOT's December 5, 2025 release projects "wholesale market savings exceeding one billion dollars annually" and does not mention load resources or demand response at all. The market redesign that most improves the position of flexible load was not sold on that basis.

One piece of Lancium history that is better than the myth

A claim repeated widely holds that no electricity consumer qualified as a Controllable Load Resource until 2020, when Lancium qualified a Bitcoin mining load, and that crypto miners were the only load type ever to hold the designation.

None of that is supported. Lancium's own release claims it qualified "the first datacenter CLR in ERCOT's history," which is a narrower statement, and ERCOT's records show Load Resources deploying as early as April 2006 with roughly 2,400 MW registered by 2011. ERCOT publishes no breakdown of Controllable Load Resources by industry at all.

The real Lancium story is more useful. In Market Notice M-A041125-01, dated April 11, 2025, ERCOT announced it had licensed "all of Lancium's U.S. patents… that are potentially applicable to Load Resources' registration as Controllable Load Resources," on terms that are royalty-free, perpetual and irrevocable within the ERCOT region, and sublicensed automatically upon registration. ERCOT's language: "No further action beyond registration with ERCOT as a CLR is required in order for any entity to avail itself of the license."

A patent barrier to CLR registration existed, and the grid operator bought it out on behalf of the entire market. That is a more interesting fact than the exclusivity legend it replaces, and it points the other way.

What registration actually takes

For anyone weighing the path, the requirements are specific. A qualified scheduling entity at level 3 or 4, with level 4 required for ancillary service market participation. A designated Decision Making Entity for CLRs seeking security-constrained economic dispatch qualification. ICCP telemetry established by the QSE under ERCOT's communications handbook. Registration through RIOO at $500, with a planned load date at least 45 days out. Qualification testing under ERCOT's Controllable Load Qualification Test Procedure for Ancillary Services. Mandatory primary frequency response for any resource qualified for Regulation or Responsive Reserve. And registration data including maximum deployment time, maximum weekly energy, and normal and emergency ramp rate curves across ten segments.

None of that is exotic. All of it is work, and the fact that six operators have completed it says something about the gap between the concept and the execution.

The only clean public record

Riot Platforms is, as far as public filings show, the only ERCOT-exposed miner that discloses curtailment credits as a discrete, audited line item in its financial statements. Searches across Cipher Mining, CleanSpark, MARA, Core Scientific, Bitdeer, Hut 8 and Galaxy found peers addressing curtailment qualitatively in risk factors and earnings commentary without quantifying it in a comparable recurring disclosure.

That matters more than it sounds. It means every market-wide estimate of monetised flexible load in ERCOT is extrapolated from a single company. This document extrapolates from it too, and says so.

The series, which is not a growth story

The 2024 figure fell 53% against 2023 on a mild Texas summer. 2025 rose 68% against 2024. The series is volatile and non-monotonic, because it tracks weather and scarcity pricing rather than a steadily growing program. Any presentation of it as a trend line is misleading. For scale: Riot's full-year 2025 power and direct costs were $339.011 million before credits and $282.282 million after.

Two accounting traps that appear in almost every write-up

Power curtailment credits are not revenue. In Riot's Condensed Consolidated Statements of Operations the line appears with negative values, and in the company's results releases it is rendered as "Less: power curtailment credits," subtracted from total power cost. They reduce cost of revenue. CNBC, the Houston Chronicle, and a 2022 Senate investigation letter have all described these credits as money "made" or "earned." The filings are authoritative and the press is wrong.

The headline cents-per-kilowatt-hour figure already includes the credits. Riot's own footnote defines its all-in power cost as "Inclusive of all transmission and distribution charges, fees, adders, and taxes. Net of Total Power Credits." So quoting both a 3.0¢/kWh power cost and a $21 million credit total as separate benefits counts the same money twice. Riot's stated all-in power cost was 3.4¢/kWh for 2024 and 3.0¢/kWh for the first quarter of 2026.

The anatomy of the most famous number in the industry

In September 2023 Riot reported "a new monthly record for Power and Demand Response Credits, totaling $31.7 million in August." The coverage that followed framed it as Texas paying a Bitcoin miner $31.7 million to shut down. Two days later, on September 8, 2023, Riot published a memorandum responding directly to that coverage. Its own breakdown:

Riot adds that the $7 million was less than one percent of an ERCOT program that "administered nearly $1 billion during this time period." So roughly 22% of the number came from an ERCOT program. Roughly 76% came from Riot monetising energy it had already bought forward under a private retail contract, in exchange for credits against future energy bills.

That is the most important sentence in this document. The money was in the power contract, not the grid program. And it points directly at where the value in flexible load actually sits, which is a conclusion the independent research reaches by a completely different route.

Riot's quarterly production updates split the total into two components the financial statements combine. In Q1 2026 that was $13.5 million in "Power Credits" and $7.5 million in "Demand Response Credits." The first is energy resale through the provider. The second is being paid to be dispatchable. It is the cleanest available separation of the two mechanisms, and it appears in no press coverage we could find.

Two scope cautions

Riot's 2024 and 2025 annual reports describe curtailment across both ERCOT and MISO. Consolidated credit totals from 2024 onward are therefore not purely ERCOT figures, and any per-megawatt ERCOT ratio derived from them is overstated by an unquantified amount.

And Riot's capacity is routinely inflated. The company's own term is developed capacity: 700 MW at Rockdale, 400 MW at Corsicana with approximately 1 GW expected on completion, and 137 MW in Kentucky. That is roughly 1,237 MW developed. The "1.7 GW" figure in circulation adds Rockdale's actual capacity to Corsicana's planned capacity.

Where the mining company went next

In January 2026 Riot announced a ten-year data centre lease with AMD for an initial 25 MW of critical IT load at Rockdale, with expansion options up to a further 75 MW and a right of first refusal on 100 MW more. By the second quarter of 2026 the initial 25 MW was commissioned and AMD had exercised another 25 MW, bringing contracted capacity to 50 MW, with further phases scheduled for November 2026 and May 2027.

The larger one gets less attention. Riot's Q2 2026 release describes a lease with a frontier AI lab for 191 MW of critical IT capacity at Rockdale, a twenty-year term running through June 2048, and a total contract value of $9.1 billion. The counterparty is not named. Initial delivery is December 2027.

A company that built the clearest public record of monetised flexible load in ERCOT is now signing twenty-year contracts to host compute that will not be flexible in the same way. That is not a contradiction. It is what happens when the value of a well-sited, well-interconnected megawatt rises faster than the value of curtailing it.

The best independent estimate, and its ceiling

The most credible independent assessment of demand-side flexibility available is not about Bitcoin at all. O'Shaughnessy, Shah, Parra and Ardani, "The demand-side resource opportunity for deep grid decarbonization," published in Joule on May 18, 2022, was authored in part by the National Renewable Energy Laboratory under contract to the US Department of Energy.

Its conclusion is careful and it does not flatter anyone:

Ten to thirty percent, and secondary. That is the honest ceiling, from a national lab with no stake in the outcome, and it should be the boundary on any claim made about flexible load anywhere. Three further findings from the same paper matter for how the opportunity is understood.

The resource is barely used. "Only about 25% of demand-side resource capacity is enrolled in demand-side service programs in the US and even those services are rarely used." The authors put roughly 200 GW of flexible load in the United States today capable of providing grid services. The gap between what exists and what participates is the actual opportunity, and it is a gap of execution rather than technology.

The value is not where most people look for it. Citing work by Dyson and colleagues, the paper reports that about 69% of the grid value of load flexibility is attributable to capacity services, 23% to energy services, and 8% to ancillary services. Eight percent. The ancillary service market, which absorbs nearly all the attention in writing about flexible load, is the smallest of the three value streams by a wide margin.

And siting is a form of flexibility. The paper makes an argument that describes a business model more precisely than most business models describe themselves: "Although historic industrial siting was driven by proximity to raw materials and demand centers, future siting patterns could be driven by access to low-cost, renewable electricity. More flexible industrial siting could help balance variable renewable energy, such as by reducing curtailment in resource-rich regions." Putting a flexible industrial load where the stranded power is, is itself a grid service. NREL said so in 2022.

The paper is equally clear about the costs. Demand-side resources "cannot always provide the same quality of grid services as the same capacity of supply-side resources," and "many thousands of flexible loads must be aggregated to provide the same service as a single centralized generator." Supply-side solutions are typically more cost effective because of economies of scale. Whenever demand-side substitutes for supply-side, the authors write, "critical analysis is required."

The study that shows flexibility is a choice

The most useful empirical work on cryptocurrency data centres specifically is Paez, Mohammadi and Taylor, "Aligning load flexibility with emissions reduction: empirical insights from a multi-site study of cryptocurrency data centers" (Georgia Institute of Technology, School of Civil and Environmental Engineering; preprint arXiv:2509.04380). It draws on hourly energy data from 21 North American cryptocurrency data centres matched against locational marginal emissions over roughly three months in 2023.

Its headline result is a negative one. Uptime alone explains only about 19% of the variance in avoided emissions (adjusted R² = 0.19), and its coefficient is negative. Add two engineered metrics, the facility's maximum energy and the variability of marginal emissions on its grid, and explanatory power rises to adjusted R² = 0.59, which the authors describe as a 210% improvement. In plain terms: whether a facility curtails tells you almost nothing. When, how deeply, and where it curtails tells you most of it.

The paper then sorts the 21 facilities into performance quadrants, and one of them is empty:

The authors note it directly: "Theoretically, a fourth category, High Uptime & High Avoided Emissions, was possible, but none of the facilities aligned to this group." No facility achieved both. There is a real trade-off between running constantly and delivering grid or emissions value, and it shows up in the data.

Two facilities, same size, ninety-three times the difference

The paper’s case studies contain the single clearest illustration of the point. Two facilities in the study have essentially identical maximum energy draw.

Same size box. Roughly ninety-three times the avoided emissions.

Facility 10 recorded a single curtailment event across the entire study period, lasting 19 hours, which the authors suspect may not have been a response to grid conditions at all. Their read: "a steady load profile suggests that this facility is likely on a fixed price contract, which suggests that it is not likely to be curtailing in response to price signals," reflecting "an operational strategy that maximizes uptime over emissions reduction or grid balancing." Facility 11, in Texas, curtails to near zero when it curtails, at 98.57% magnitude. The authors conclude it is "likely price responsive but the price signals are not well-aligned with the power generator's carbon intensity."

The lesson is not that Bitcoin mining is flexible. It is that a Bitcoin mine is flexible if it is contracted and operated to be, and identical hardware under a fixed-price contract behaves like any other data centre. Flexibility is a commercial arrangement that happens to be enforced by physics, rather than a property of the machines.

You can measure this from the outside

The same author’s doctoral dissertation, defended at Georgia Tech on April 20, 2026, extends the finding in a direction that matters commercially. It introduces what it calls a data-light behavioral signature, a threshold derived from a facility’s load curve, which “separates both HPC and CDC facilities into statistically distinct operational types and provides a practical proxy for benchmarking flexibility and mitigation performance using only time-series energy data.”

No proprietary hardware specifications. No workload logs. No cooperation from the operator. A facility’s willingness and ability to flex is legible from its power profile alone, and the dissertation suggests the measure could be “integrated into real-time grid operator dashboards or regulatory monitoring frameworks as a simple performance indicator.”

That cuts both ways, which is why it belongs here. It means a counterparty can verify a flexibility claim rather than accept it. It also means a regulator can.

The honesty this section requires

Every facility in the Paez study induced far more emissions than it avoided. Facility 11's ratio of avoided to induced emissions was 0.20. Facility 10's was 0.0024. None of these operations is net-negative, and the paper does not claim otherwise. It is a study of relative effectiveness at emissions alignment among facilities that all consume substantial power. The study window was also confined to summer months, which the dissertation names as a limit on year-round generalisability.

And the dissertation is blunt about the conditional nature of the whole thesis:

That is the argument of this entire document stated by someone with no commercial interest in it. Flexibility is a capability. Whether it produces anything worth having depends on what the market pays it to do, which is exactly what Texas is in the middle of redrafting.

Its summary puts the point more broadly: the energy and climate impacts of large-scale compute “are not fixed properties of demand, but emergent outcomes of facility behavior, grid conditions, and system design.”

A disclosure that belongs with the citation. The dissertation’s acknowledgments record support from the Bitcoin Policy Institute, a Graduate Bitcoin and Energy Fellowship supported by Peter McCormack, and financial support from the Human Rights Foundation. That funding does not invalidate the work, which is committee-reviewed doctoral research with published methods and reported statistics. It does mean the research should not be described as disinterested, and readers weighing it should know where it came from. The same standard applies to every industry-funded source cited anywhere in this document.

How 4CP works, and why it has mattered so much

ERCOT's transmission costs are recovered through a mechanism called Four Coincident Peak. Under 16 Texas Administrative Code § 25.192(b), wholesale transmission charges are updated annually "based on the prior year's average of the four-coincident-peak (4CP) demand that is coincident with the ERCOT 4CP." The four peaks are the highest fifteen-minute system demand intervals in June, July, August and September. A customer's average load across those four intervals sets its share of the following year's transmission costs.

Two structural points get lost. First, ERCOT files 4CP demand for each distribution service provider, and the transmission charge matrix is used by transmission service providers to bill transmission customers. A large load's 4CP exposure reaches it as a pass-through under its utility's retail tariff, not as a direct bill from the grid operator. Second, there is a one-year lag. Curtailing across the 2026 peaks affects 2027 charges.

The 2025 rate, set in PUCT Docket No. 57491 and approved by Commission order on June 5, 2025, is exact. From the Commission Staff's Final Transmission Charge Matrix of March 20, 2025:

• Total ERCOT postage stamp rate: $68.5473 per kW-year, or $68,547.30 per MW-year at full exposure.
• Total transmission cost of service: $5,446,864,794.70.
• Total average 4CP load: 81,042,656.556 kW.

Individual utility access fees vary widely beneath that average, from roughly $1.25 per kW at Austin Energy to $9.08 at AEP Texas.

The four intervals that made the point for us

The same filing lists the actual 2024 coincident peak intervals:

Look at the first two. The June and July coincident peaks fell on consecutive calendar days. An operator curtailing on instinct through a hot mid-July afternoon missed the July interval entirely, because it had already happened on the first day of the month.

This is not a hypothetical failure mode. It is the finding of the only peer-reviewed study of 4CP response. Zarnikau and Thal, "The response of large industrial energy consumers to four coincident peak (4CP) transmission charges in the Texas (ERCOT) market," Utilities Policy, Vol. 26, pp. 1–6 (2013), found that industrial customers served at transmission voltage reduce consumption by roughly 4% in response to 4CP charges, that the response persists two to three hours because consumers cannot identify which interval will set the peak, and that some coincident peaks show minimal response at all. Customers served at primary voltage showed negligible responsiveness.

Four percent, smeared across a two-to-three-hour window, from the largest industrial consumers in Texas. That is the measured performance of conventional heavy industry guessing at a peak. It is also the clearest possible statement of what a load that can reach near-zero within a dispatch interval is actually worth, and why the gap between the two is the entire business.

And now the mechanism is going away

On August 1, 2025 the PUCT opened Project No. 58484, Evaluation of Transmission Cost Recovery. A draft report issued March 16, 2026 proposed moving from 4CP "to a methodology utilizing a greater number of coincident peaks" and "lengthening the interval over which each coincident peak is measured."

The rulemaking that followed is Project No. 58000. Commission Staff filed a proposed rule on June 11, 2026. The Commission approved it for publication on July 9, 2026. It published in the Texas Register on July 24, 2026, and comments closed on August 11, 2026. Senate Bill 6 requires the resulting rule changes by December 31, 2026.

The proposed rule defines the replacement precisely: "ERCOT 12CP intervals — The set of 12 30-minute intervals composed of the ERCOT system monthly peak demand interval for each of the 12 calendar months from October 1 of one year through September 30 of the subsequent year, inclusive." Twelve monthly peaks instead of four summer ones, measured over thirty minutes instead of fifteen. That alone triples the number of intervals a load must correctly anticipate.

But the change that actually matters is elsewhere in the same rule. Proposed § 25.193(d)(2) sets billing demand for a large load customer at the greater of:

• (A) the customer’s contracted peak demand;
• (B) the highest non-coincident peak demand placed on the system in the past year; or
• (C) the customer’s 12CP demand.

And § 25.252(e)(2)(B): "For the first 20 years after commencement of billing, the large load customer must complete 240 consecutive payments." Billing begins at the level of available service "irrespective of whether the large load customer has energized or is fully using available capacity."

Read limbs (A) and (B) carefully. Neither can be lowered by curtailing. Contracted peak demand is a contract term. Non-coincident peak is, by construction, indifferent to whether a peak happened to coincide with the system’s. Curtailing during the twelve coincident intervals reduces only limb (C), and cannot bring the bill below the higher of the other two.

Twelve peaks instead of four makes avoidance harder. The minimum billing demand removes avoidance as a strategy, for twenty years.

The proposed rule is not adopted as of this writing, and the specific figures in it may change. But the direction is set by statute with a December 31, 2026 deadline, and the Commission's own chairman has stated the reasoning plainly. Interviewed by Dr. Joshua Rhodes of the Webber Energy Group at the University of Texas at Austin, PUC Chairman Thomas Gleeson said:

Anyone underwriting a Texas flexible load site on the basis of 4CP avoidance is underwriting a mechanism with a scheduled end date.

What the statute actually says

Senate Bill 6, 89th Legislature, Regular Session, authored by Senator Phil King, passed the Senate 31–0 on March 19, 2025 and was signed and took effect on June 20, 2025. The provision most often cited is new Utilities Code § 39.170(a). Here is what it actually requires:

• The duty runs to the utility, not the customer. Each electric cooperative, transmission and distribution utility and municipally owned utility serving a transmission-voltage customer must develop a protocol, including installing necessary equipment before interconnection, to allow the load to be curtailed during firm load shed.
• It applies to transmission-voltage customers interconnected after December 31, 2025.
• It contains no megawatt threshold. The 75 MW figure that circulates belongs to § 37.0561(c) and to § 39.170(b), which are different provisions.
• It exempts critical load industrial customers and designated critical natural gas facilities.
• It is a firm-load-shed mechanism. The load is curtailed when the grid is already shedding firm load. It is not a general dispatch obligation.

The word "controllable" does not appear anywhere in Senate Bill 6. The bill adds §§ 35.004(c-1), 35.004(c-2), 37.0561, 39.169 and 39.170, and none requires registration as a Controllable Load Resource or a Provisional Controllable Load Resource. No law firm analysis of the bill mentions such a mandate, because there isn't one.

That should settle a claim now widespread in industry commentary, including in an earlier version of this document: that every AI data centre seeking ERCOT interconnection must demonstrate controllable load capability. It has no statutory basis.

The demand response service that doesn't exist yet

Section 39.170(b) directs ERCOT to "develop a reliability service to competitively procure demand reductions from large load customers with a demand of at least 75 megawatts to be deployed in the event of an anticipated emergency condition." Rules must specify usable periods, at least 24 hours notice, and one more thing worth reading twice: exclusion of customers that curtail in response to wholesale price, or that participate in other reliability services.

ERCOT's own legislative status tracker, dated July 2026, gives the state of play:

This is a direction, not a product. Two of the four tasks ERCOT lists have not begun. It has no rules, no price, and no participants. And when it does arrive, it will exclude loads that already curtail on price, which is to say it will exclude the operating model this entire document describes.

The specific dollar figures are proposed, not law

The statute requires a flat study fee of at least $100,000 for initial transmission screening studies, with unused portions credited toward interconnection obligations. It leaves financial commitment requirements to the Commission. The widely quoted figures come from proposed 16 TAC § 25.194 in PUCT Project 58481. That rule was approved for publication on March 12, 2026, comments closed April 17, 2026, and it has not been adopted. Treat the numbers as proposals.

What the first real case decided

On July 24, 2026 the PUCT issued its order in Docket No. 59220, the first net metering case under Senate Bill 6. Goodnight Wind, a 265.5 MW facility in Armstrong County, paired with Crusoe Load Two, a 260 MW AI data centre whose interest was later transferred to Ensign Infrastructure. The Commission approved the arrangement subject to conditions:

• Full load curtailment within 30 minutes during grid emergencies, without compensation
• Sixty minutes advance notice “when practicable”
• The load is barred from participating in ERCOT demand response and ancillary service programs
• Mandatory regulatory review in 36 to 60 months
• A cumulative 525.5 MW curtailment obligation across the two co-located data centres at the site

And the Commission rejected proportional mitigation. Emergency curtailment imposed on a co-located data centre is not capped by the capacity of the behind-the-meter generator it sits next to.

Sit with that third condition. A co-located large load in Texas now accepts an uncompensated curtailment obligation and is prohibited from earning demand response or ancillary service revenue. For that structure, curtailability has become a cost of interconnection rather than a source of income. If the business case for a site depends on stacking grid payments on top of a cheap co-located power price, Docket 59220 is the document to read before signing anything.

The machinery being built underneath

ERCOT is building the control-room system to execute all this. Its Large Load Curtailment Manager was described in a May 24, 2026 presentation to the Board as a tool giving operators "advance warning that system conditions may require Large Load curtailment to avoid EEA as allowed under SB6," monitoring projected supply margins over a forward horizon to provide "actionable lead time to initiate curtailment before a shortage materializes." A first phase focused on net metering arrangements was targeted for June 2026.

ERCOT stated at the same meeting that it was "nearing filing" of a revision request to establish the rule framework, including registration and performance requirements. No such filing has been located, and delivery of the first phase is unconfirmed. It is an operations tool under development, not an approved market mechanism, and it should be described that way.

In August, the Governor stopped it

On August 3, 2026, Governor Abbott directed the PUCT and ERCOT to conduct a comprehensive verification and audit of every data centre project in the interconnection queue, covering public financial support, power arrangements, on-site generation, water use and sourcing, cooling technology, community impact mitigation, and ownership and control. Projects that fail to comply "must be denied" grid connection.

ERCOT’s own market notices record what followed. On the day of the directive, ERCOT announced it would miss the August 7 deadline for issuing interconnection classifications and would seek a good-cause exception. The PUCT granted one on August 20, moving the date to August 31. On August 31 ERCOT missed that deadline too, citing the need for further "data validation and due diligence." No classification results have been published.

Two audit workstreams now run in parallel: a Batch Zero Audit verifying eligibility attestations for large loads at or above 75 MW, and a Community Impact Audit covering computational loads at or above 25 MW. ERCOT expects verification to take several months. The study deadline has moved from late 2026 to April 9, 2027, with a final eligibility verification report scheduled for December 10, 2026.

The interconnection pathway that Texas spent three years building is, as of this writing, halted and under state audit. That is the third of the three rewritings, and it is the one nobody saw coming.

What was paused

Until mid-2026 ERCOT studied large load interconnection requests one at a time, under a process established by Planning Guide Revision Request 115. That process ran "through end of day on July 10, 2026."

Its replacement is Batch Zero, created by PGRR145 and NPRR1325, adopted by the ERCOT Board on June 2, 2026, approved by the PUCT on June 18, 2026, and effective July 11, 2026. ERCOT describes it as the most-commented policy change in the history of its stakeholder process, and as the first batch interconnection process used by any grid operator for large loads.

Batch Zero groups projects of 75 MW or larger into a single coordinated study and sorts each request into one of three classifications. Base Load covers projects that completed prior studies or are already in approved planning studies; they are modelled as background load and receive their allocation without restudy. Studied Load enters without a predetermined allocation, with capacity determined through a system-wide reliability assessment; it may receive full capacity, partial capacity, or deferral. Excluded Load qualifies as neither and must wait for a later batch.

A related mechanism arrived with it. The Provisional Controllable Load Resource is an election, made through a notarised declaration on Planning Guide Section 23 Form W, by which a project facing transmission constraints may energise earlier in exchange for accepting dispatch above its firm consumption limit. A load that elects it does end up registering as a CLR, with a qualified scheduling entity, telemetry and qualification testing. But that is a voluntary trade for earlier energisation, not a requirement.

Two corrections are needed here, because both errors are widespread. PCLR was not created by PGRR134; that revision request was tabled on December 4, 2025 and withdrawn on May 27, 2026, and never took effect. And PCLR is not yet operational: the Board motion adopting PGRR145 defers its implementation to "upon system implementation," and ERCOT’s August 2026 stakeholder materials still list it among planned future enhancements.

What the queue is, and what it isn’t

Three numbers describe ERCOT’s large load pipeline, and they measure completely different things.

Requests, approvals, actual consumption. A ratio close to 100 : 2 : 1.

ERCOT’s own materials note that the vast majority of tracked requests lack submitted studies, and the preliminary volumes circulating before the pause, roughly 150 GW screened as eligible Base Load, come from an eligibility screen rather than a completed study. ERCOT received about 290 dynamic model submissions by the July 10 deadline and found problems with a significant number of them.

The forecast has the same character. In April 2026 ERCOT published a protocol-required projection of 367,790 MW of peak demand by 2032, roughly four times the current record. Within weeks it filed at the PUCT to re-base the number, describing it in Board materials as a figure "that few stakeholders considered credible for market planning purposes." The PUCT agreed on June 18, 2026 that the forecast should be rebuilt on Batch Zero classifications instead. Those classifications are the ones now stuck behind the audit, so no revised figure exists.

None of this means the load isn’t coming. ERCOT’s June 2026 system peak of 82,772 MW ran 5,376 MW above the June 2025 peak, and July 2026 came in 12% above July 2025. The growth is real and it is fast. What the queue number cannot tell you is how much of the remainder arrives, or when, and every planning decision downstream of it inherits that uncertainty.

A document that only presents the case for flexible load is not worth citing. Here are the strongest arguments against, with their sources.

Large computational loads may be a reliability liability, not an asset

In June 2026 the ERCOT Board adopted NOGRR282 and NPRR1308, establishing frequency and voltage ride-through requirements for Large Computational Loads, defined as one or more facilities at a single site with aggregate peak demand at or above 75 MW where at least half of site demand is power-electronic-based computational load. The PUCT approved both on July 9, 2026, effective August 1, 2026.

The requirements are specific. Continuous operation between 58.8 and 61.2 Hz. At least 299 seconds in the bands either side. Voltage ride-through continuous between 0.90 and 1.10 per unit, 2.0 seconds at 0.80–0.90, 0.5 seconds at 0.50–0.80, and down to 0.15 seconds below 0.20 per unit. Through all of it, the load must continue consuming current and hold active power within 10% of its pre-disturbance level.

Why the rule exists is the uncomfortable part. At the Large Load Working Group on June 19, 2026, ERCOT presented an interim voltage ride-through assessment with a finding it stated in one line:

Four groups of large loads that could drop off the system simultaneously during a fault. ERCOT notes none currently operates above that threshold, and that it is working with transmission providers on mitigation. But the exposure is real, quantified by the grid operator, and it points the opposite way from the flexible-load thesis.

It also produces the sharpest distinction in this whole subject. ERCOT now wants large loads to be curtailable on instruction and simultaneously incapable of falling over on their own. Those are different engineering problems. A load that can be told to stop is an asset. A load that trips is a contingency. The industry has spent years talking about the first and comparatively little about the second, and the regulation has now caught up.

The pivot to AI may remove the flexibility

The Paez dissertation closes on a warning that reads directly onto the Riot leases described earlier in this document:

Set that against the record. The operator with the clearest public history of monetised curtailment in ERCOT has committed 191 MW at Rockdale to a twenty-year AI lease running to 2048, plus 50 MW contracted to AMD. Those megawatts are not going to behave like the load that earned $71.2 million in curtailment credits in 2023.

The flexible-load thesis and the AI buildout are usually presented as the same story. On this evidence they may be in tension, with the more valuable contract winning. That is not a reason to dismiss flexible load. It is a reason to be precise about which megawatts are actually flexible, and to notice that the answer changes when a better-paying tenant appears.

The value may be smaller than the enthusiasm

O'Shaughnessy and colleagues put demand-side resources at 10–30% of what deep decarbonization requires, and secondary to supply-side solutions. Within that, only about 8% of the grid value of load flexibility is attributable to ancillary services. Supply-side solutions are generally more cost effective because of economies of scale, and demand-side resources cannot always deliver the same quality of service.

The programs are narrowing, not widening

Within ERCOT specifically, over the past twelve months: DRRS was created and excludes load. Section 39.170(b) remains undeveloped and will exclude price-responsive loads when it arrives. Docket 59220 barred a co-located load from demand response and ancillary programs outright. And the transmission charge that made curtailment valuable is scheduled for replacement by December 31, 2026 with a mechanism that includes a twenty-year minimum billing demand.

Every one of those is a narrowing. An honest reading is that the specific revenue stack available to an ERCOT flexible load in 2023 is smaller in 2026 and will be smaller again in 2027.

The counterweight literature

The critical academic work deserves engagement rather than dismissal. Menati, Lee and Xie at Texas A&M modelled cryptocurrency mining as demand flexibility on a synthetic Texas grid and found benefits, but within the scope of a simulation, and the paper should not be cited for claims about reserve requirements or system-wide ancillary costs that it does not make. Work from NBER and MIT's Center for Energy and Environmental Policy Research examines the cost and emissions consequences of flexible data centres and reaches less favourable conclusions under some assumptions. Any operator making an investment decision should read the sceptics.

The temptation with a document like this is to end by claiming the Texas model generalises. It doesn't, and the reasons are specific. Registration categories, telemetry standards, accreditation methods, settlement rules and retail contract structures differ market by market. ERCOT is an energy-only market with no capacity construct, which changes the economics of being available. It is electrically islanded, which changes what balancing is worth. Its retail market is competitive in most of the territory, which is what made Riot's TXU arrangement possible in the first place. Change any of those and the arithmetic changes with them.

What does transfer is the sequence, and Texas is simply ahead on it:

• Prove the resource works. ERCOT did this over roughly fifteen years, from Load Resources deploying in 2006 through six Controllable Load Resources qualified in 2025.
• Pay for it, and discover what you're paying for. Riot's filings show the honest answer: most of the money came through the power contract, not the grid program.
• Watch the load arrive faster than the rules anticipated. From roughly 8.9 GW approved to energize against 474 GW requested, in a queue that has outgrown serial study.
• Rewrite the terms. The transmission charge, the interconnection process, the co-location rules, and the ride-through standard, all inside twelve months.

Any market about to absorb large computational load will run this sequence. The useful question for an operator or a developer is not whether flexible load works. It demonstrably does, at a scale the data supports. The question is which terms will still be in place when a site built today reaches its third year, and that question has to be answered market by market.

Drawn from the evidence above, not from a product pitch.

Texas did something no other market has finished doing. It built a way for large electrical loads to be dispatched like generators, it found operators willing to be dispatched, and it paid them enough that the money is visible in audited accounts. That happened, and it works, and the evidence for it is public.

It is also smaller than the story suggests. Six controllable load resources. Roughly 3% of qualified responsive reserve. Ten to thirty percent of what deep decarbonization needs, and secondary to supply. One company with a clean public record of what it earned, most of it from a power contract rather than a grid program.

And the terms are being rewritten as this is published. The transmission charge that made curtailment pay is scheduled to be replaced within months. The newest reserve product excludes load. The first co-location case barred a data centre from the demand response market as a condition of connecting. The interconnection process is paused under a state audit with its deadline pushed to April 2027.

That combination is not a failure of the model. It is what success looks like when it arrives faster than the rules were built for, and it is the most useful thing about the Texas experience for anyone watching from another market. The question was never whether interruptible load is worth something. ERCOT settled that. The question is what it will be worth under the terms being written right now, and in Texas those terms are still in the drafting.

Type 3 holds no proprietary data on the subjects covered here. Every figure is drawn from public sources and can be independently checked. Where sources disagree, the disagreement is stated rather than resolved silently.