How many hours of demand response can a Texas data center actually offer tenants?

AI tenants doing diligence on a Texas data center site increasingly ask one specific question: how many hours of demand response can you actually offer us? The answer is not one number. It is three constraints stacked on top of each other, each with a different operational meaning and a different ceiling. Most operators do not separate them in the tenant pitch, which makes the “flexible” claim either vague or unprovable when a procurement team runs the math.

The combined number is what underwrites the LOI. Here are the three.

Constraint 1: PCLR binding curtailment

ERCOT’s Provisional Controllable Load Resource is a binding dispatch obligation a load accepts in exchange for early energization through Batch Zero. When ERCOT calls a PCLR event, the load actually curtails. No genset compensation, no override. ERCOT dispatches PCLRs through SCED based on local congestion at the load’s node. The frequency depends almost entirely on shift factor at the site against historical shadow-price profiles on the binding constraint.

Expected curtailment varies wildly by location. A site in the Panhandle near the wind-export bottleneck plausibly sees hundreds of hours/year. A site in the Coastal Bend or East Texas may see fewer than 50. The geographic dispersion is large enough that “PCLR enrollment” means very different things at different parcels. The technical breakdown lives in the companion post on ERCOT PCLR and Batch Zero.

Constraint 2: ERS and voluntary CLR

Emergency Response Service (ERS) is ERCOT’s voluntary demand-response program. Loads bid availability per four-month Standard Contract Term and are paid for being on call. ERS deployment is rare in practice: an annual 15 to 30-minute test event is guaranteed, and real deployments historically average closer to zero than to ten per year. Total annual deployment hours for an enrolled site are typically well under twenty.

Voluntary (non-provisional) Controllable Load Resources participate in ERCOT’s ancillary services markets — Responsive Reserve Service, ECRS, Non-Spin. Calls are more frequent than ERS but still infrequent operationally; deployment events for properly-bid CLRs typically cluster in the tens, not hundreds, per year. For a 60 MW Texas data center, voluntary CLR plus ERS revenue can run well into seven figures annually with operational disruption measured in tens of hours, not hundreds.

ERS and voluntary CLR are stackable with PCLR in principle, though the operational envelope and event-overlap accounting matter. They are not a substitute for PCLR — they trade revenue for optionality, where PCLR trades a binding obligation for energization speed.

Constraint 3: Title V air permit headroom

This is the constraint most often forgotten. Backup gensets are subject to Title V of the Clean Air Act when potential to emit exceeds 100 tons/year of any criteria pollutant in an attainment area (lower in non-attainment). Williamson County is in the Austin-Round Rock MSA and currently attainment/unclassifiable for the 2015 ozone NAAQS, so the federal 100 ton/year threshold applies to NOx.

For a typical AI-HPC data center with 100 MW of Tier 4 final diesel gensets — the current EPA non-road emission standard at 0.40 g/bhp-hr NOx — the runtime budget at full load before NOx alone triggers Title V is:

100 MW × 1.34 hp/kW × 1000 = 134,000 hp
134,000 hp × 0.40 g/bhp-hr = 53,600 g/hr NOx
53,600 g/hr ÷ 907,185 g/ton = 0.059 tons/hr
100 tons/year ÷ 0.059 tons/hr ≈ 1,690 hours/year

Roughly 1,500 to 1,700 hours/year of full-load runtime headroom on NOx alone, before Title V applicability is triggered. PM2.5, CO, and HAP thresholds bind at different ratios; in practice NOx is usually the first constraint for diesel-genset data centers, and dispersion modeling can fail before annual mass totals do for sites near sensitive receptors. The number is also a function of the genset emission tier — Tier 2 gensets emit roughly 16x more NOx per bhp-hr than Tier 4, and the runtime budget shrinks proportionally.

What the tenant pitch can credibly say

Stack the three constraints and the operator-side numbers fall out in plain ranges:

The credible offer to an AI tenant in Texas is roughly 100 to 300 hours/year of curtailable load via PCLR plus voluntary CLR, with several hundred hours of remaining genset runtime headroomunder the Title V cap for non-curtailment use cases (training-pause windows backed by generation, planned maintenance windows, scheduled inference ramps). That is a real flexibility budget. It is not “unlimited shed for unlimited duration,” which is what the marketing word “flexible” sometimes implies.

For a tenant procurement team, the two numbers that matter are: (a) committed curtailable hours/year through ERCOT programs, and (b) remaining genset runtime headroom under the air permit. Both are knowable. Most operators do not compute or expose them, which is why “flexible AI/HPC” in a marketing one-liner often does not survive procurement review.

What Cliff is building

Cliff computes these numbers per site. Given a parcel, a generation stack, and a target enrollment posture, we return expected PCLR curtailment frequency (shift factor against historical shadow-price profile), ERS and voluntary CLR revenue and event modeling, and Title V runtime headroom by pollutant under the local attainment-status profile. Every input is cited to a primary source. If you have a Texas project and tenants asking how many hours of flex you can actually offer, send us the location.

Primary sources

• ERCOT PGRR145 (Batch Zero / PCLR mechanism)
• PUC Project 58481 (16 TAC §25.194 large-load interconnection standards)
• ERCOT Emergency Response Service program
• TCEQ Austin-Round Rock current attainment status
• EPA Title V operating permits program
• EPA Tier 4 nonroad diesel engine emission standards