⚡ electrical engineering

To Defer or To Shift? The Role of AI Data Center Flexibility on Grid Interconnection

This paper proposes a quantitative grid planning model to evaluate AI data center flexibility, revealing that while spatial and temporal load shifting can reduce grid investment and operational costs by 3–21%, the benefits are location-dependent and exhibit diminishing returns with longer deferral times.

Original authors: Yize Chen, Xiaogui Zheng

Published 2026-04-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yize Chen, Xiaogui Zheng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the power grid as a massive, complex highway system, and AI data centers as a sudden, explosive influx of super-heavy trucks (like 747s filled with gold) that need to get to their destinations immediately.

Right now, the grid is built for regular cars and occasional delivery vans. Suddenly, we have thousands of these massive trucks demanding to enter the highway at the exact same time. The result? Traffic jams, road collapses, and a massive bill to build new lanes and bridges just to keep up.

This paper asks a simple but revolutionary question: What if these trucks didn't have to arrive right now or stay in one specific lane?

Here is the breakdown of the research using everyday analogies:

1. The Problem: The "Rigid" Truck

Currently, we treat AI data centers like rigid, unmovable trucks. If a company wants to train a new AI model, the grid must promise, "We will give you 100% of the power you need, 24/7, no matter what."

The Consequence: To guarantee this, the utility company has to build massive new power plants and upgrade transmission lines just for these peak moments. It's like building a 10-lane highway just because a parade might happen once a year. It's incredibly expensive and slow.

2. The Solution: The "Flexible" Truck

The authors propose treating AI data centers as flexible, smart trucks that can do two things:

Time Shifting (Deferral): "Hey, I don't need to deliver this package right this second. I can wait 2 hours until traffic dies down." (This is delaying AI tasks).
Space Shifting (Geographic Shifting): "This highway is jammed. I'll take a detour to a different city where the roads are empty, deliver the package there, and then send it back." (This is moving AI workloads to different data centers in different locations).

3. The Big Surprise: Flexibility Doesn't Always Mean "Less"

You might think, "If the trucks are flexible, we need fewer roads and fewer power plants."
The paper says: Not necessarily.

Here is the twist:

The "Unlock" Effect: Sometimes, making the trucks flexible actually makes the existing highway system work so efficiently that the grid operator realizes, "Hey, if we just add one more power plant here, we can unlock huge savings elsewhere."
The Analogy: Imagine a crowded elevator. If everyone is rigid, you need a bigger elevator. But if people are flexible (some wait for the next one, some take the stairs), the current elevator works better. However, because the elevator is now working so well, you might decide to add a new floor to the building because it's finally profitable to do so.
The Result: The total cost of the system (building + running) goes down significantly (by 3% to 21%), even if the amount of new power plants built doesn't always go down. The flexibility makes the whole system smarter and cheaper to run.

4. The "Sweet Spot" (Diminishing Returns)

The researchers found a limit to how much flexibility helps:

Time: If you delay an AI task by 1 or 2 hours, it's a huge help. But if you say, "Wait 10 hours," the benefit stops growing. It's like waiting for a bus; waiting 10 minutes saves you time, but waiting 2 hours just makes you late without saving much more.
Space: Moving data around helps most when the roads (power lines) are already jammed. If the roads are empty, moving the truck doesn't save much money.

5. Why This Matters

This research is a roadmap for the future. Instead of just screaming, "Build more power plants!" or "Stop building AI!", it offers a middle ground:

For Grid Operators: You don't always need to build a new power plant immediately. You can ask AI companies to be flexible first.
For AI Companies: You can get your power connection faster if you agree to be flexible (wait a bit or move your work).
For Everyone: It saves money and keeps the lights on without needing to pave over the entire countryside with new power lines.

The Bottom Line

The paper argues that flexibility is the new currency. By treating AI data centers not as rigid demands but as flexible partners that can shift their time and location, we can solve the energy crisis without breaking the bank. It's about making the grid "dance" with the AI, rather than trying to force the AI to march in lockstep with a rigid grid.

1. Problem Statement

The rapid proliferation of Generative AI (GenAI) and Large Language Models (LLMs) has created an unprecedented surge in electricity demand from AI Data Centers (AIDCs). Projections suggest US AI power demand could grow 30-fold by 2035, reaching 123 GW. This growth has overwhelmed traditional grid interconnection pipelines, causing delays of over five years and necessitating massive, costly infrastructure upgrades.

Current grid planning paradigms treat data centers as rigid, firm loads, requiring the grid to supply power at any moment to meet Service Level Objectives (SLOs). This approach is becoming economically and operationally untenable. The paper addresses the critical question: Given large load interconnection targets, what type and magnitude of data center flexibility are required to mitigate grid bottlenecks, reduce costs, and accelerate interconnection?

2. Methodology

The authors propose a quantitative two-stage capacity expansion model that integrates AI load flexibility directly into grid planning.

A. Modeling AI Load Flexibility

The paper translates AI SLOs (e.g., latency thresholds for inference, throughput for training) into two specific flexibility mechanisms:

Temporally Deferrable Load: Workloads that can be delayed within a specific time window ( $W_d$ ). This is modeled using backlog dynamics ( $b_t$ ) where unserved work accumulates and must be cleared within the delay window.
Geographically Shiftable Load: Workloads that can be moved between data center locations (buses) to alleviate local congestion. This is modeled as a shifting vector ( $\delta_t$ ) subject to conservation constraints ( $\sum \delta_t = 0$ ) and bus-specific shifting limits.

B. The Optimization Framework

The problem is formulated as a Two-Stage Stochastic-like Deterministic Expansion Model:

First Stage (Investment): Decides on the capacity addition ( $x_g$ ) for candidate generators to minimize total investment costs ( $C_{inv}$ ).
Second Stage (Operation): Optimizes hourly dispatch ( $p_t$ ), voltage angles ( $\theta_t$ ), and flexible load decisions ( $s_t, \delta_t$ ) to minimize operational costs (generation cost, delay penalties, and shifting costs).
Constraints: The model incorporates DC power flow approximations, transmission line limits, nodal power balance, and the specific temporal/spatial constraints of the AI workloads.

C. Simulation Setup

The model was tested on two systems:

IEEE 14-bus system: Used as a transparent benchmark to visualize geographical shifting behaviors and validate the model.
WECC 240-bus system: Used to evaluate planning implications in a realistic, large-scale transmission network with heterogeneous load and congestion patterns.

Data: Hourly load profiles based on WECC data, with AI loads superimposed on specific buses.
Scenarios: Experiments varied the delay window size, the proportion of shiftable load, and the level of AI load penetration (up to 1.5x nominal load).

3. Key Contributions

Novel Integration: First work to explicitly couple AI compute flexibility (deferral and shifting) with grid capacity expansion planning, moving beyond simple demand response models.
Counter-Intuitive Insight: The study challenges the assumption that flexibility always reduces the need for new generation capacity. It demonstrates that flexibility can sometimes increase the optimal capacity investment if it unlocks a more efficient, lower-cost operating regime that was previously inaccessible due to congestion.
Diminishing Returns Analysis: Quantifies the "sweet spot" for flexibility, showing that deferral benefits plateau after a specific time threshold.
Decision Support Tool: Provides a quantitative framework for utilities and policymakers to determine the specific flexibility levels needed to avoid costly upgrades or accelerate interconnection.

4. Key Results

Cost Reduction: Both temporal and geographical flexibility can reduce total grid operational costs by 3% to 21%, depending on grid conditions and flexibility levels.
Capacity Investment Paradox:
- In some scenarios, introducing flexibility leads to an increase in required generation capacity. This occurs because flexibility allows the system to utilize cheaper generation sources more effectively, making the expansion of capacity economically viable where it previously was not.
- However, this increase in capacity is offset by significant reductions in operational costs, leading to a net reduction in total system cost.
Geographical Shifting:
- Highly effective in congested networks. When lines are binding (1.0x capacity), shifting load helps redistribute demand to regions with headroom.
- In uncongested networks (1.2x capacity), increasing flexibility monotonically reduces capacity investment by over 21%.
Temporal Deferral (Delay Window):
- Diminishing Returns: The benefits of deferring load diminish significantly once the delay window exceeds 3 hours. Short-term flexibility (1–3 hours) provides the majority of the grid relief.
- Extending delays beyond 3 hours offers negligible additional value for grid dispatch pressure.
Congestion Mitigation: Geographical shifting effectively mitigates line congestion by moving load away from "hot" buses (e.g., Bus 9 in the 14-bus case) to non-congested regions during peak hours.

5. Significance and Implications

Policy & Planning: The paper argues that treating AI loads as flexible resources is essential for modern grid planning. Relying on "firm load" models leads to over-investment in infrastructure and prolonged interconnection queues.
Strategic Trade-offs: Utilities and data centers must balance the cost of implementing flexibility (e.g., storage, software throttling, multi-site coordination) against the savings in grid upgrades. The model suggests that short-term flexibility (3 hours) is the most cost-effective lever for grid operators.
Future Grid Design: The findings suggest that future grid expansion decisions should not be made in isolation from compute flexibility. A coordinated approach where data centers actively participate in demand response can unlock cheaper generation assets and reduce the need for expensive transmission upgrades.
Limitations & Future Work: The current study uses a deterministic framework. Future work needs to address the stochastic nature of renewable generation and AI load uncertainty, potentially using two-stage stochastic programming. Additionally, the distinct hardware profiles of training vs. inference workloads require further granular modeling.

In summary, this paper provides a rigorous mathematical foundation for understanding how AI data centers can transition from being a grid bottleneck to a flexible asset, offering a roadmap for optimizing the co-evolution of the power grid and the AI industry.