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

Xiaogui Zheng; Yize Chen

arxiv: 2604.05376 · v1 · submitted 2026-04-07 · 📡 eess.SY · cs.SY

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

Yize Chen , Xiaogui Zheng This is my paper

Pith reviewed 2026-05-10 19:42 UTC · model grok-4.3

classification 📡 eess.SY cs.SY

keywords AI data centersload flexibilitygrid interconnectioncapacity expansionpower systems planningtemporal shiftingspatial shiftingdemand response

0 comments

The pith

AI data centers that defer or spatially shift their loads can cut grid investment and operating costs by 3 to 21 percent, though benefits are uneven and do not always grow with more flexibility.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper builds a grid capacity expansion model that lets AI data centers move their electricity use forward or backward in time and across different locations instead of treating them as fixed demands. It tests how these options affect the need for new power plants, daily operating expenses, and transmission congestion under varying grid conditions and data center placements. The results show that flexibility helps lower total costs in many cases but that simply allowing longer deferral windows or wider shifting ranges does not guarantee steadily larger savings. This matters because rapid AI growth is creating interconnection delays that traditional rigid-load planning cannot resolve without expensive overbuilding.

Core claim

Incorporating temporal deferral and spatial shifting of AI computational loads into a quantitative capacity expansion framework reveals that flexibility reduces additional generation capacity and operational costs by 3-21 percent depending on data center location, flexibility range, and existing grid load, yet the reductions are not monotonic with increasing flexibility and longer deferral times exhibit diminishing returns for relieving electricity dispatch pressure.

What carries the argument

Grid capacity expansion model that explicitly represents AI data center temporal deferral and spatial shifting to quantify effects on generation needs, costs, and congestion.

If this is right

Grid planners can sometimes meet AI load growth with less new generation capacity by allowing data centers to adjust timing and location of consumption.
Operational costs fall when AI loads are permitted to defer or shift, with the largest reductions occurring under certain location and load conditions.
Network congestion relief depends on where data centers are sited relative to transmission constraints.
Extending deferral windows beyond moderate lengths yields progressively smaller additional relief on electricity dispatch.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Utilities might redesign interconnection queues to reward or require data-center flexibility commitments in exchange for faster approvals.
The same modeling approach could be applied to other large, schedulable loads such as hydrogen electrolysis or cryptocurrency mining.
Locating new data centers in regions where the model predicts high flexibility value could become part of regional energy planning.
Real operating data from data centers could be used to refine the flexibility parameters and test whether the modeled cost ranges hold.

Load-bearing premise

The chosen ranges for how far and how long AI loads can be deferred or shifted accurately reflect what real data centers can do without harming their own operations.

What would settle it

Field measurements of actual generation additions and cost changes after an AI data center begins deferring or spatially shifting loads on a real grid that matches the modeled conditions.

read the original abstract

The integration of AI data centers into power grid represents one of the most emerging and complex challenges for the energy systems. As computational demand scales at an unprecedented rate, the traditional grid planning study's paradigm of treating data centers as rigid, inflexible loads is becoming economically, mathematically and operationally untenable. This work tries to understand and address the large load interconnection bottleneck by modeling and evaluating AI load flexibility. By examining data center's temporal and spatial shifting capabilities within a grid capacity expansion framework, we build a quantitative grid planning model, and evaluate their impacts on additional generation, operational costs, and network congestion. Numerical study reveals interesting observations, as AI data center flexibility are not felt consistently, and increasing flexibility does not necessarily translate to less generation capacity required. Depending on data center's locations, flexibility range, and grid load conditions, flexible AI load can help reduce grid investment and operational costs by 3-21%. Our work also indicate that longer deferral time of AI compute has diminishing returns for offloading grid electricity dispatch pressure.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper finds that AI data center flexibility cuts grid costs by 3-21% in some scenarios but does not reliably lower generation capacity needs and shows diminishing returns on longer deferrals.

read the letter

The main takeaway is that embedding temporal and spatial shifting for AI loads into a capacity expansion model produces cost savings that vary by location and grid conditions, yet the flexibility does not translate into consistent reductions in required generation capacity and longer deferral windows add little extra relief after a point. This challenges the default view that more flexibility always eases interconnection pressure for hyperscale loads. The work is a straightforward extension of existing capacity-expansion tools to this specific load type, and the numerical observations on non-monotonic effects and location dependence are the clearest contribution. It is useful to see quantified ranges rather than qualitative claims about flexibility helping the grid. The authors also correctly flag that deferral benefits plateau, which aligns with practical limits on compute scheduling. The soft spots sit mainly in the input assumptions. The chosen flexibility fractions, maximum deferral horizons, and spatial relocation rules need explicit justification against real data-center job schedulers and service-level agreements; if the model permits more shifting than operators can actually deliver without penalties, both the 3-21% savings band and the inconsistency result become sensitive to parameterization rather than robust outcomes. The abstract gives no equations, data sources, or validation steps, so the full paper must include sensitivity checks and comparison to measured grid data to make the numbers credible. Without those, the central claims rest on untested ranges. This is for power-systems modelers and utility planners who already work with flexible-load extensions and want a concrete AI-data-center case. Readers outside that group will find the setup familiar but the application timely. The paper deserves peer review because the topic matters for near-term interconnection decisions and the modeling approach is a reasonable incremental step, even though revisions will be needed on parameter grounding and robustness checks.

Referee Report

2 major / 2 minor

Summary. The paper develops a grid capacity expansion model incorporating temporal deferral and spatial shifting of AI data center loads. Numerical experiments show that these flexibilities do not reduce required generation capacity monotonically and can lower combined investment and operational costs by 3-21% depending on data-center location, flexibility range, and background grid conditions; longer deferral horizons exhibit diminishing returns on dispatch relief.

Significance. If the modeling assumptions prove representative, the work supplies quantitative evidence that AI-load flexibility can ease interconnection bottlenecks and reduce system costs, an issue of growing importance for power-system planning. The non-monotonic and location-dependent results usefully temper expectations that “more flexibility is always better.” The forward modeling approach from stated flexibility bounds to cost outcomes avoids circularity.

major comments (2)

[Numerical study] Numerical study section: the reported 3-21% cost reductions and the observation that “increasing flexibility does not necessarily translate to less generation capacity” rest on specific ranges for deferrable load fraction and maximum deferral horizon. No justification, literature reference, or sensitivity sweep against measured AI workload deferrability (e.g., inference-job SLAs, training checkpointing constraints) is supplied; if the chosen bounds exceed realistic values, both the headline savings interval and the non-monotonic claim become parameterization artifacts rather than robust grid outcomes.
[Model formulation] Model formulation section: the capacity-expansion optimization embeds temporal and spatial shifting variables, yet the manuscript provides neither the explicit mathematical formulation of the shifting constraints nor the data sources used to calibrate grid load profiles and network parameters. Without these, independent verification of the 3-21% figures or the diminishing-returns result on deferral time is impossible.

minor comments (2)

[Abstract] Abstract: the quantitative claims (3-21% savings, non-monotonic effects) are presented without any accompanying model equation, data source, or error bar; adding one or two key equations or a one-sentence description of the optimization framework would improve readability.
[Numerical results] Figure captions and tables: several numerical results are shown without accompanying units, baseline definitions, or confidence intervals; clarifying these would aid interpretation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments and positive evaluation of the paper's significance. We address each of the major comments below and outline the revisions we plan to make.

read point-by-point responses

Referee: [Numerical study] Numerical study section: the reported 3-21% cost reductions and the observation that “increasing flexibility does not necessarily translate to less generation capacity” rest on specific ranges for deferrable load fraction and maximum deferral horizon. No justification, literature reference, or sensitivity sweep against measured AI workload deferrability (e.g., inference-job SLAs, training checkpointing constraints) is supplied; if the chosen bounds exceed realistic values, both the headline savings interval and the non-monotonic claim become parameterization artifacts rather than robust grid outcomes.

Authors: We acknowledge the validity of this concern. The manuscript does not currently include justification or references for the chosen parameter ranges. These ranges were intended to capture a spectrum of possible AI load flexibilities to illustrate the potential impacts. In the revised manuscript, we will add a dedicated subsection on parameter selection, including references to studies on AI workload deferrability (such as those discussing SLA requirements for inference jobs and constraints in training processes), and perform additional sensitivity analyses to show how the results vary with different bounds. This will strengthen the claim that the observed cost savings and non-monotonic effects are robust. revision: yes
Referee: [Model formulation] Model formulation section: the capacity-expansion optimization embeds temporal and spatial shifting variables, yet the manuscript provides neither the explicit mathematical formulation of the shifting constraints nor the data sources used to calibrate grid load profiles and network parameters. Without these, independent verification of the 3-21% figures or the diminishing-returns result on deferral time is impossible.

Authors: We agree that providing the explicit mathematical formulation and data sources is necessary for full transparency and verifiability. The model formulation section in the current manuscript describes the overall framework but omits the detailed equations for the shifting variables. We will revise the manuscript to include the complete mathematical model, specifying the constraints for temporal deferral and spatial shifting. Furthermore, we will detail the sources of the grid load profiles and network parameters used in the case studies. These changes will enable independent replication and verification of the numerical results. revision: yes

Circularity Check

0 steps flagged

No circularity: cost and capacity outcomes derive from forward optimization under explicit flexibility bounds.

full rationale

The paper formulates a grid capacity expansion model that treats AI load deferral and spatial shifting as decision variables or constraints with chosen ranges. Numerical experiments then solve for generation investment, operational costs, and congestion under those ranges, producing the reported 3-21% savings and non-monotonic observations. No equation or result reduces by construction to a fitted parameter renamed as prediction, no self-citation supplies a load-bearing uniqueness theorem, and no ansatz is smuggled in; the chain remains self-contained from stated modeling assumptions to simulation outputs.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

Only abstract available; ledger is therefore incomplete. Model relies on standard optimization assumptions for capacity expansion and on unstated representations of data-center flexibility ranges and grid load conditions.

free parameters (2)

flexibility range and deferral time limits
Chosen values that determine how much load can be shifted; directly affect the reported 3-21% cost savings.
data center location and grid load condition parameters
Location-specific inputs that modulate whether flexibility reduces or does not reduce required generation capacity.

axioms (1)

domain assumption Grid capacity expansion can be represented as a mathematical optimization problem with temporal and spatial shifting variables for AI loads.
Invoked to build the quantitative planning model described in the abstract.

pith-pipeline@v0.9.0 · 5479 in / 1333 out tokens · 49587 ms · 2026-05-10T19:42:51.465717+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

6 extracted references · 6 canonical work pages · 1 internal anchor

[1]

The role of flexible connection in accelerating load interconnection in distribution networks,

Google. A new milestone for smart, afford- able electricity growth. URL https://blog. google/innovation-and-ai/infrastructure-and-cloud/ global-network/demand-response-data-center-milestone/. Blog post on data center demand response and grid flexibility. N. Gu, G. Chen, and J. Qin. The role of flexible connection in accelerating load interconnection in di...

work page arXiv
[2]

X. He, Z. Fang, J. Lian, D. H. Tsang, B. Zhang, and Y. Chen. Freesh: Fair, resource-and energy-efficient schedul- ing for llm serving on heterogeneous gpus.arXiv preprint arXiv:2511.00807,

work page arXiv
[3]

Y. Li, M. Mughees, Y. Chen, and Y. R. Li. The unseen ai disruptions for power grids: Llm-induced transients.arXiv preprint arXiv:2409.11416,

work page arXiv
[4]

Accessed: 2026-04-06

URL https://www.wecc.org/ 7 program-areas/reliability-planning-performance-analysis/ reliability-modeling/loads-resources. Accessed: 2026-04-06. P. Xiong and C. Singh. Optimal planning of storage in power systems integrated with wind power generation.IEEE Trans- actions on Sustainable Energy, 7(1):232–240,

work page 2026
[5]

M. Yan, T. Joswig-Jones, B. Zhang, Y. Chen, and W. Cu. Switching-reference voltage control for distribution sys- tems with ai-training data centers.arXiv preprint arXiv:2603.15588,

work page arXiv
[6]

T. V. Zuluaga, S. Pang, and J.-P. Watson. Nodal capacity expansionplanningwithflexiblelarge-scaleloadsiting.arXiv preprint arXiv:2510.19781,

work page internal anchor Pith review Pith/arXiv arXiv

[1] [1]

The role of flexible connection in accelerating load interconnection in distribution networks,

Google. A new milestone for smart, afford- able electricity growth. URL https://blog. google/innovation-and-ai/infrastructure-and-cloud/ global-network/demand-response-data-center-milestone/. Blog post on data center demand response and grid flexibility. N. Gu, G. Chen, and J. Qin. The role of flexible connection in accelerating load interconnection in di...

work page arXiv

[2] [2]

X. He, Z. Fang, J. Lian, D. H. Tsang, B. Zhang, and Y. Chen. Freesh: Fair, resource-and energy-efficient schedul- ing for llm serving on heterogeneous gpus.arXiv preprint arXiv:2511.00807,

work page arXiv

[3] [3]

Y. Li, M. Mughees, Y. Chen, and Y. R. Li. The unseen ai disruptions for power grids: Llm-induced transients.arXiv preprint arXiv:2409.11416,

work page arXiv

[4] [4]

Accessed: 2026-04-06

URL https://www.wecc.org/ 7 program-areas/reliability-planning-performance-analysis/ reliability-modeling/loads-resources. Accessed: 2026-04-06. P. Xiong and C. Singh. Optimal planning of storage in power systems integrated with wind power generation.IEEE Trans- actions on Sustainable Energy, 7(1):232–240,

work page 2026

[5] [5]

M. Yan, T. Joswig-Jones, B. Zhang, Y. Chen, and W. Cu. Switching-reference voltage control for distribution sys- tems with ai-training data centers.arXiv preprint arXiv:2603.15588,

work page arXiv

[6] [6]

T. V. Zuluaga, S. Pang, and J.-P. Watson. Nodal capacity expansionplanningwithflexiblelarge-scaleloadsiting.arXiv preprint arXiv:2510.19781,

work page internal anchor Pith review Pith/arXiv arXiv