Quantifying AI data center flexibility as a resource adequacy asset

Technical Summary: Quantifying AI Data Center Flexibility as a Resource Adequacy Asset

Problem Statement
Current resource adequacy planning in U. S. electricity grids treats AI data center loads as "firm," meaning every megawatt of nameplate capacity requires a corresponding megawatt of procured generation and reserves to guarantee uninterrupted service. This approach assumes local power supply must be consistent at every site, driven by service-level agreements (SLAs) guaranteeing uninterrupted compute. As data center demand is projected to grow by 33.7 GW in PJM alone by 2030, this practice drives capacity procurement requirements that exceed the physical limits of the interconnection queue. While policy research suggests load flexibility could integrate projected growth, there is a lack of technical frameworks that translate operational flexibility mechanisms into the specific metrics required for capacity accreditation.

Methodology
The paper proposes an eleven-parameter framework to quantify the "commitment depth" of AI inference loads—the megawatt-equivalent grid value per megawatt of nameplate load eligible for capacity accreditation. The framework decomposes flexibility into three operational axes, combining parameters multiplicatively within mechanisms and additively across mechanisms with a sequential ordering to prevent double-counting:

1. Spatial Migration: Moving inference workloads to unstressed grid regions.
2. Temporal Deferral: Delaying non-interactive workloads (NIW) during stress events.
3. Dynamic Voltage-Frequency Scaling (DVFS): Reducing power on residual compute without violating SLAs.

The study utilizes four years (2022–2025) of day-ahead locational marginal pricing (DA LMP) data across five U.S. Regional Transmission Organizations (PJM, ERCOT, CAISO, MISO, NYISO) to analyze stress decorrelation. It also analyzes 49.4 million production inference requests from Azure (DynamoLLM and BurstGPT datasets) to characterize workload drain times.

The analysis employs a conditional Monte Carlo simulation that conditions on observed per-hour grid states. It evaluates two distinct scenarios reflecting the projected 2030 fleet heterogeneity:

• Mixed-use: Hyperscaler colocation facilities where inference is ~43% of IT compute.
• Inference-dominant: Purpose-built AI campuses where inference exceeds 70% of IT compute.

Key Results

• Commitment Depth: At the single-facility level (500 MW), the mean commitment depth is 40.0% for inference-dominant facilities and 24.6% for mixed-use facilities.
• Scale Dependence: As the fleet scales to 10 GW, commitment depth declines only modestly to 38.5% (inference-dominant) and 24.1% (mixed-use). This stability occurs because deferral and DVFS mechanisms contribute additively and do not saturate, even as spatial migration faces increased destination contention.
• Stress Decorrelation: Empirical analysis shows high dynamic availability for cross-region migration. For a source in ComEd (Northern Illinois), unstressed destinations were available in 99.0% of stress hours. Cross-interconnection regions (e.g., ERCOT, CAISO) exhibit stress overlap below 30%, whereas same-interconnection regions show significantly higher overlap (43–66%).
• Workload Drain Times: The 99th-percentile drain time for inference requests ranges from 4.5 to 12.12 seconds, nearly two orders of magnitude faster than the 10-minute minimum dispatch window of PJM. This supports a conservative parameter setting where 95% of active requests can finish within an emergency dispatch window.
• Economic Impact: Using a P5-based accreditation method (accounting for the worst 5% of stress hours), a 10 GW inference-dominant fleet could provide approximately 2,322 MW of accredited capacity. This translates to avoided capacity costs of roughly $638 million per year (based on $180,000/MW-yr). However, the paper notes that the Net Present Value (NPV) of accelerating interconnection (avoiding compute revenue loss) far exceeds these avoided capacity costs.

Significance and Claims
The paper claims to bridge the gap between systems literature on operational flexibility and the capacity-accreditation metrics required by grid planners. Its primary contribution is demonstrating that the constraints on grid value for AI data centers are not physical limitations but institutional challenges.

The analysis identifies three institutional parameters that dominate uncertainty and residual risk:

1. DESTabsorb: The fraction of GPU headroom operators maintain and hardware compatibility.
2. MIGelig: The fraction of workloads eligible for migration, currently determined by contracts.
3. QUEUEok: Post-stress queue absorption, dependent on whether facilities coordinate recovery windows.

The paper concludes that cross-region stress decorrelation is a physical property of the North American grid and that facility-level power management is already established at major hyperscalers. Therefore, realizing the full potential of AI data centers as resource adequacy assets requires policy instruments focused on flexible interconnection for queued facilities and accredited demand response for operating facilities, rather than further infrastructure investment. The framework reallocates reliability risk, suggesting that data center operators can absorb part of the risk through flexibility mechanisms, reducing the need for grid operators to procure capacity at full facility ratings.