What Is Demand Response Actually Worth?
One gigawatt of data center demand response — agreeing to curtail load during grid stress — is worth $122M per year in avoided PJM capacity payments. At 30% curtailment across 10 GW of DC load, demand response alone extends the grid tipping point by approximately five years. Zero capital cost. Immediate impact. The least-discussed option on the table.
The Value of Flexible Load
Investigation 3 showed interruptibility is at least 5× cheaper than new gas in every weather year tested. This investigation puts a dollar figure on that finding using PJM's actual capacity auction prices. The sensitivity analysis identified demand response as a “High” impact lever (109 reliability-impact hours). Here we quantify exactly what that means in capital terms.
Avoided Cost by Curtailment Level
PJM's 2027/28 Base Residual Auction cleared at $333/MW-day. Each megawatt of demand response enrolled avoids that capacity obligation. The math is straightforward: the question is how much load data centers are willing to curtail, and for how long.
Avoided cost = curtailed GW × 1,000 MW/GW × $333/MW-day × 365 days. Tipping point extension estimated from Investigation 1 sensitivity: each GW reduction in peak DC load extends the reliability cliff by ~1.5–2 years (tipping point range 14–24 GW; 1 GW ≈ 1.7 year shift at median).
Demand Response vs. Self-Generation
Investigation 4 established the self-generation economics. This table places demand response in direct comparison. The capital cost difference is the key finding.
| Strategy | Capital Cost | Annual Value | Timeline |
|---|---|---|---|
| 10% curtailment (1 GW DR) | $0 | $122M avoided | Immediate |
| 20% curtailment (2 GW DR) | $0 | $244M avoided | Immediate |
| 30% curtailment (3 GW DR) | $0 | $366M avoided | Immediate |
| On-site Gas CCGT (Investigation 4) | $2–4B | $37/MWh savings | 2–4 year build |
| Nuclear PPA / Clean Self-Gen (Inv. 4) | $4–8B | $7/MWh savings | 3–5 year build |
The self-gen vs. demand response comparison. Investigation 4 showed on-site generation requires $2–8B in capital to provide reliability independence. Demand response provides comparable reliability value at near-zero capital cost — the tradeoff is operational flexibility vs. capital certainty. For data centers that can tolerate brief load reductions (for example, deferring non-time-critical ML training jobs during off-peak hours), demand response dominates economically.
When Does Curtailment Actually Happen?
The key misconception about demand response is that it requires constant interruption. In practice, PJM emergency signals are rare and brief.
Historical Stress Hours
PJM issues emergency demand response signals when real-time LMP exceeds ~$200/MWh or reserve margins fall below operating minimums. Historically, this occurs 50–150 hours per year — roughly 0.6–1.7% of total operating hours. In most years, it is zero.
The sensitivity analysis identified demand response as a “High” impact lever at 109 hours — that is, demand response affects reliability outcomes during approximately 109 hours of the simulation year. Those are the hours that matter for capacity planning. Curtailing during those hours eliminates the grid stress without affecting the other 8,651 hours of data center operation.
What Hyperscale Operators Can Actually Do
ML training workloads are not real-time. A 10% reduction in DC load during a 4-hour evening peak event means delaying a subset of batch jobs by 4–8 hours. For inference workloads, geographic load shifting (routing traffic to DCs outside the stressed zone) achieves the same result without any compute delay. The operational constraint is real but manageable — especially with advance notice signals, which PJM provides 1–4 hours ahead.
18% curtailment for 1.2% of hours was the specific finding from Investigation 3. That is the operational ask: curtail load during ~105 hours per year. For a 1 GW data center, that means shedding 180 MW for roughly 100 hours — or about 18,000 MWh per year out of 8.76 million total MWh consumed. The curtailment share of total energy is 0.2%.
Why Isn’t This Already Standard Practice?
The math favors DR by a wide margin. The operational ask is 0.2% of compute hours. Yet most hyperscale data center agreements with PJM are firm load — not interruptible. Three barriers explain the gap.
Regulatory Structures Are Lagging
PJM's demand response programs were designed for industrial and commercial customers (HVAC, refrigeration, manufacturing). The market structures for data center interruptibility — with automated load shedding, workload migration protocols, and advance-notice guarantees — do not yet exist at scale. FERC Order 2222 opened the door for distributed energy resource aggregation, but implementation is still underway.
Hyperscale SLAs Create Perceived Risk
Cloud providers sell 99.99% uptime SLAs to customers. Engineering and legal teams perceive demand response participation as SLA risk, even when the actual exposure is <0.2% of compute hours. The contractual and operational frameworks to de-risk DR participation have not been standardized.
Capital Projects Are Easier to Justify
A $2B gas plant is a discrete capital project with an identifiable owner, depreciation schedule, and balance sheet treatment. Demand response contracts are operational commitments that cut across infrastructure, operations, and customer commitments teams. The organizational complexity of DR exceeds the engineering complexity of self-generation — which is why the more expensive option often wins.
The most valuable intervention is blocked by organizational friction, not economics or physics. The capacity market price signal ($333/MW-day) already prices DR at $122M/GW/year. The missing piece is the operational and contractual framework to act on it.
How This Was Calculated
Capacity value: Based on PJM 2027/28 Base Residual Auction clearing price ($333/MW-day, the most recent completed auction). Annual value per GW = 1,000 MW × $333/MW-day × 365 days = $121.5M, rounded to $122M. This is the avoided capacity payment — what a data center would pay PJM if it does not provide interruptibility, versus what it pays if it enrolls as a demand response resource.
Tipping point extension: Derived from Investigation 1 sensitivity analysis. The reliability tipping point spans 14–24 GW of DC load across 200 stochastic draws (median ~19 GW). Each 1 GW reduction in effective DC peak load shifts the tipping point distribution by approximately 1.5–2 years at median, based on the load growth trajectory used in the MC simulation (~1.5–2 GW/year DC growth rate). Curtailment modeled as load reduction during hours when PJM LMP > $200/MWh or emergency signals are issued (historically 50–150 hours/year in PJM; 109 hours in the simulation sensitivity analysis).
What is not included: Demand response contract costs (interruptibility premiums, control systems, automated load shedding infrastructure) are not included in the avoided cost figure. These vary by operator and contract structure but are typically <10% of the avoided capacity value at current auction prices. Geographic load shifting costs (network bandwidth, cross-region replication) are also not included.