pith. sign in

arxiv: 2605.14109 · v1 · pith:PAWGWSC2new · submitted 2026-05-13 · 📡 eess.SY · cs.SY

Grid Integration of Gigawatt-Scale AI Data Centers under Connect-and-Manage

Xin Lu , Qianwen Xu This is my paper

Pith reviewed 2026-05-15 05:01 UTC · model grok-4.3

classification 📡 eess.SY cs.SY
keywords AI data centersgrid integrationconnect-and-managecurtailment reductionhierarchical architecturebattery energy storageworkload flexibilitypower systems
0
0 comments X

The pith

Gigawatt-scale AI data centers can connect to transmission grids without upgrades using a hierarchical coordination protocol that slashes curtailment while maintaining training workloads.

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

This paper develops a coordination framework for AI data centers to integrate into power systems under connect-and-manage policies, which permit immediate connection but impose real-time power curtailment during grid congestion. The approach models distinct workload types—frontier training, batch training, and inference—with shared battery storage to exploit different flexibilities. A three-layer system uses learning for planning requests, robust evaluation by the operator, and optimization for execution. Readers should care because it demonstrates how massive computing loads can become grid assets rather than liabilities by shifting flexible tasks and storing energy on-site.

Core claim

The paper claims that the opaque TSO acceptance can be handled by a hierarchical architecture consisting of a learning-based planning layer that generates power requests, a robust acceptance mechanism at the TSO, and a single-step execution optimizer that ensures feasibility under the allocated budget. Case studies on the IEEE 39-bus system with Australian data show curtailment dropping from 9.1% to 2.8%, 98.1% frontier training preserved, batch training providing the largest flexibility swing, and batteries buffering via discharge and deferral.

What carries the argument

The three-layer hierarchical architecture for sequential request-acceptance protocol with curtailment variable and information boundary between AIDC and TSO.

If this is right

  • Batch training acts as the primary grid-elastic resource with the largest throughput swing during peak demand.
  • The on-site battery provides curtailment buffering through active discharge and charge deferral.
  • The framework reduces curtailment from 9.1% to 2.8% while preserving 98.1% frontier training workload in case studies.
  • This is achieved on the IEEE 39-bus system using Australian market data.

Where Pith is reading between the lines

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

  • The model could be extended to account for multiple AIDCs interacting with the same TSO.
  • Data center operators might benefit from investing in more battery capacity to further reduce curtailment impacts.
  • Similar protocols could be applied to other large-scale flexible loads in the grid such as cryptocurrency mining operations.

Load-bearing premise

The TSO's acceptance mapping is completely opaque to the AIDC and can be treated as a robust black-box mechanism whose worst-case behavior is known in advance.

What would settle it

A real-world implementation on a gigawatt-scale AIDC where the actual curtailment exceeds 2.8% or the preserved frontier workload falls below 98.1% under the proposed framework would falsify the performance claims.

Figures

Figures reproduced from arXiv: 2605.14109 by Qianwen Xu, Xin Lu.

Figure 3
Figure 3. Figure 3: shows the weekly operation of the SAC policy on the evaluation period. The upper panel displays the power request and accepted exchange at the PCC. Throughout the week, 𝑃𝑟𝑒𝑞 fluctuates between 600 and 1,100 MW, reaching approximately 1,200 MW during off-peak periods on February 23, and dropping as low as 700 MW and 600 MW during the most severe grid stress events on February 23 and February 29. The accepte… view at source ↗
read the original abstract

Emerging connect-and-manage interconnection practices allow gigawatt-scale artificial intelligence data centers (AIDCs) to connect to the transmission network without prior network upgrades, at the cost of real-time curtailment during grid stress. This paper formalizes the resulting AIDC-transmission system operator (TSO) coordination as a sequential request-acceptance protocol with an explicit curtailment variable and a strict information boundary between the two parties. Physical models are developed on both sides of the point of common coupling: the AIDC is decomposed into frontier training, batch training, and inference serving subclasses sharing on-site battery energy storage, capturing differentiated temporal flexibility; the transmission network is modeled via DC power flow with generator constraints and budget-constrained demand uncertainty. Because the TSO's acceptance mapping is opaque to the AIDC, a three-layer hierarchical architecture is formulated in which a learning-based planning layer generates power requests, the TSO evaluates each request through a robust acceptance mechanism, and a single-step execution optimizer enforces internal feasibility under the realized power budget. Case studies with a gigawatt-scale AIDC on the IEEE 39-bus system with Australian market data show that the framework reduces curtailment from 9.1% to 2.8% while preserving 98.1% frontier training workload, that batch training acts as the primary grid-elastic resource with the largest throughput swing during peak demand, and that the on-site battery provides curtailment buffering through active discharge and charge deferral.

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.

Referee Report

2 major / 2 minor

Summary. The paper formalizes AIDC-TSO coordination under connect-and-manage as a sequential request-acceptance protocol with an explicit curtailment variable and information boundary. It decomposes the AIDC into frontier training, batch training, and inference workloads sharing on-site BESS, models the transmission network via DC power flow with generator constraints and budget-constrained demand uncertainty, and proposes a three-layer hierarchy: a learning-based planning layer that generates requests, a robust TSO acceptance mechanism, and a single-step execution optimizer that enforces feasibility under the realized budget. IEEE 39-bus case studies with Australian market data report curtailment reduction from 9.1% to 2.8% while preserving 98.1% frontier training workload, with batch training identified as the primary grid-elastic resource and the battery providing buffering via discharge and charge deferral.

Significance. If the central claims hold, the work supplies a concrete, hierarchical protocol for integrating gigawatt-scale AI loads into transmission networks without prior upgrades, with quantified trade-offs between curtailment, workload preservation, and resource flexibility. The use of an external test system and market data, together with differentiated workload modeling, strengthens the practical relevance for power-system operators facing rapid AI-driven demand growth.

major comments (2)
  1. [three-layer architecture and robust acceptance mechanism] The performance guarantees (curtailment drop from 9.1% to 2.8% and 98.1% frontier-workload preservation) rest on the assumption that the TSO acceptance mapping is a known robust black-box whose worst-case behavior can be encoded in advance. If actual TSO decisions incorporate private generator status, forecast errors, or non-robust criteria outside this model, the single-step execution optimizer cannot guarantee internal feasibility, undermining the reported IEEE 39-bus gains.
  2. [case studies and demand uncertainty formulation] The budget-constrained demand uncertainty model and its propagation through the planning layer are not shown to be tight; it is unclear whether the reported curtailment reductions remain stable when the uncertainty set is enlarged or when the learning-based planner is retrained on different Australian market traces.
minor comments (2)
  1. [introduction and model sections] Notation for the curtailment variable and the information boundary between AIDC and TSO should be introduced earlier and used consistently across the three layers.
  2. [results] The abstract states concrete numerical outcomes; the main text should include a table or figure that directly compares the baseline (no coordination) against the three-layer framework for all key metrics.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the practical relevance of the hierarchical AIDC-TSO coordination protocol. We address each major comment below and outline targeted revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [three-layer architecture and robust acceptance mechanism] The performance guarantees (curtailment drop from 9.1% to 2.8% and 98.1% frontier-workload preservation) rest on the assumption that the TSO acceptance mapping is a known robust black-box whose worst-case behavior can be encoded in advance. If actual TSO decisions incorporate private generator status, forecast errors, or non-robust criteria outside this model, the single-step execution optimizer cannot guarantee internal feasibility, undermining the reported IEEE 39-bus gains.

    Authors: We appreciate this observation on the scope of our robustness guarantees. The protocol explicitly enforces an information boundary, with the AIDC observing only the accepted power budget rather than private TSO data such as generator status or internal forecasts. The robust acceptance mechanism encodes worst-case behavior strictly within the modeled budget-constrained demand uncertainty set, ensuring that the single-step execution optimizer maintains internal feasibility for any budget realization inside that set. We agree that if the TSO applies non-robust or private criteria outside the modeled uncertainty, the reported performance cannot be guaranteed. In the revision we will add a dedicated discussion subsection clarifying the assumptions, the conditional nature of the guarantees, and possible extensions (e.g., online adaptation) for non-robust TSO behavior. revision: partial

  2. Referee: [case studies and demand uncertainty formulation] The budget-constrained demand uncertainty model and its propagation through the planning layer are not shown to be tight; it is unclear whether the reported curtailment reductions remain stable when the uncertainty set is enlarged or when the learning-based planner is retrained on different Australian market traces.

    Authors: We agree that additional validation of tightness and sensitivity is warranted. The current results use Australian market data to parameterize the budget-constrained uncertainty set and train the learning-based planner. In the revised manuscript we will augment the case studies with (i) enlarged uncertainty sets obtained by scaling the budget parameter and (ii) retraining and evaluation on additional market traces drawn from different periods or regions. These experiments will quantify the stability of the curtailment reduction (2.8 %) and frontier-workload preservation (98.1 %) metrics. revision: yes

Circularity Check

0 steps flagged

No significant circularity; results derived from external IEEE 39-bus simulations and Australian market data

full rationale

The paper's central claims (curtailment reduction from 9.1% to 2.8%, 98.1% frontier workload preservation) are generated via case studies on the IEEE 39-bus system using external Australian market data. The three-layer architecture (learning-based planning, robust TSO acceptance, single-step execution) treats the TSO mapping as an opaque black-box with assumed known worst-case behavior; this is an explicit modeling assumption, not a self-referential definition or fitted input renamed as prediction. No equations reduce outputs to inputs by construction, and no load-bearing steps rely on self-citations whose content is unverified or tautological. The derivation remains self-contained against the external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The framework rests on standard DC power flow assumptions and robust optimization under budget-constrained uncertainty; no new physical constants or entities are postulated beyond the three workload subclasses and the hierarchical layers themselves.

axioms (2)
  • standard math DC power flow approximation holds for the transmission network under the studied operating conditions
    Invoked in the transmission network model section of the abstract
  • domain assumption TSO acceptance decisions can be treated as a robust black-box mapping whose worst-case behavior is known
    Central to the three-layer architecture description

pith-pipeline@v0.9.0 · 5561 in / 1496 out tokens · 39149 ms · 2026-05-15T05:01:18.062233+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages · 2 internal anchors

  1. [1]

    NVIDIA Launches Omniverse DSX Blueprint, Enabling Global AI Infrastructure Ecosystem to Build Gigawatt -Scale AI Factories,

    J. Mills, “NVIDIA Launches Omniverse DSX Blueprint, Enabling Global AI Infrastructure Ecosystem to Build Gigawatt -Scale AI Factories,” NVIDIA Blog, 2025

    work page 2025

  2. [2]

    The cost of compute: A $7 trillion race to scale data centers,

    M. G. Jesse Noffsinger , et. al. Areita Bhan, “The cost of compute: A $7 trillion race to scale data centers,” McKinsey Quarterly, 2025

    work page 2025

  3. [3]

    Energy and AI,

    I. E. Agency, “Energy and AI,” World Energy Outlook Special Report, 2025

    work page 2025

  4. [4]

    Structural alignment for energy –computation co-design,

    X. Lu, J. Qiu, X. Wang, J. Gu, J. Lin, S. An, and J. Zhao, “Structural alignment for energy –computation co-design,” Nature Reviews Electrical Engineering, 2026

    work page 2026

  5. [5]

    PJM, stakeholders begin work on Board's plan to reliably integrate large loads,

    PJM, “PJM, stakeholders begin work on Board's plan to reliably integrate large loads,” PJM Inside Lines, 2026

    work page 2026

  6. [6]

    AI data centres as grid - interactive assets,

    P. Colangelo, A. K. Coskun, J. Megrue, C. Roberts, S. Sengupta, V . Sivaram, E. Tiao, A. Vieaykar, C. Williams, and D. C. Wilson, “AI data centres as grid - interactive assets,” Nature Energy, vol. 11, no. 2, pp. 254-261, 2026

    work page 2026

  7. [7]

    Blowing off steam: How power -fleoible AI factories can stabilize the global energy grid,

    J. Parker, “Blowing off steam: How power -fleoible AI factories can stabilize the global energy grid,” NVIDIA Blog, 2026

    work page 2026

  8. [8]

    Electricity demand and grid impacts of ai data centers: Challenges and prospects,

    X. Chen, X. Wang, A. Colacelli, M. Lee, and L. Xie, “Electricity demand and grid impacts of AI data centers: Challenges and prospects,” arXiv preprint arXiv:2509.07218, 2025

    work page arXiv 2025

  9. [9]

    Internet data centers participating in electricity network transition considering carbon -oriented demand response,

    T. Wan, Y . Tao, J. Qiu, and S. Lai, “Internet data centers participating in electricity network transition considering carbon -oriented demand response,” Applied Energy, vol. 329, pp. 120305, 2023

    work page 2023

  10. [10]

    Unlocking spatio-temporal fleoibility of data centers in multiple regional peer -to-peer energy transaction markets,

    T. Jin, L. Bai, M. Yan, and X. Chen, “Unlocking spatio-temporal fleoibility of data centers in multiple regional peer -to-peer energy transaction markets,” IEEE Trans. on Power Systems, vol. 40, no. 5, pp. 3914-3927, 2025

    work page 2025

  11. [11]

    Mitigating power grid impact from proactive data center workload shifts: A coordinated scheduling strategy integrating synergistic traffic -data-power networks,

    Y . Zhang, B. Zou, X. Jin, Y . Luo, M. Song, Y . Ye, Q. Hu, Q. Chen, and A. C. Zambroni, “Mitigating power grid impact from proactive data center workload shifts: A coordinated scheduling strategy integrating synergistic traffic -data-power networks,” Applied Energy, vol. 377, pp. 124697, 2025

    work page 2025

  12. [12]

    Fleoible energy storage system and renewable energy planning for sustainable internet data center considering temporal and spatial load regulation,

    T. Wan, J. Qiu, Y . Tao, S. Lai, and R. Mao, “Fleoible energy storage system and renewable energy planning for sustainable internet data center considering temporal and spatial load regulation,” IEEE Trans. on Industry Applications, 2025

    work page 2025

  13. [13]

    Eoploiting internet data centers as energy prosumers in integrated electricity -heat system,

    X. Yin, C. Ye, Y . Ding, and Y . Song, “Eoploiting internet data centers as energy prosumers in integrated electricity -heat system,” IEEE Trans. on Smart Grid, vol. 14, no. 1, pp. 167-182, 2022

    work page 2022

  14. [14]

    Supply restoration of data centers in fleoible distribution networks with spatial -temporal regulation,

    J. Jian, J. Zhao, H. Ji, L. Bai, J. Xu, P. Li, J. Wu, and C. Wang, “Supply restoration of data centers in fleoible distribution networks with spatial -temporal regulation,” IEEE Trans. on Smart Grid, vol. 15, no. 1, pp. 340-354, 2023

    work page 2023

  15. [15]

    Wide-Area Power System Oscillations from Large-Scale AI Workloads

    M.-S. Ko, and H. Zhu, “Wide -area power system oscillations from large - scale AI workloads,” arXiv preprint arXiv:2508.16457, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  16. [16]

    Rethinking load growth: Assessing the potential for integration of large fleoible loads in us power systems,

    T. Norris, T. Profeta, D. Patino -Echeverri, and A. Cowie -Haskell, “Rethinking load growth: Assessing the potential for integration of large fleoible loads in us power systems,” 2025

    work page 2025

  17. [17]

    Grid Operational Benefit Analysis of Data Center Spatial Fleoibility: Congestion Relief, Renewable Energy Curtailment Reduction, and Cost Saving,

    H. Wan, L. Fang, and X. Li, “Grid Operational Benefit Analysis of Data Center Spatial Fleoibility: Congestion Relief, Renewable Energy Curtailment Reduction, and Cost Saving,” arXiv preprint arXiv:2511.08759, 2025

    work page arXiv 2025

  18. [18]

    Proactive demand response for data centers: A win -win solution,

    H. Wang, J. Huang, X. Lin, and H. Mohsenian -Rad, “Proactive demand response for data centers: A win -win solution,” IEEE Trans. on Smart Grid, vol. 7, no. 3, pp. 1584-1596, 2015

    work page 2015

  19. [19]

    Data -driven risk -adeusted robust energy management for microgrids integrating demand response aggregator and renewable energies,

    Z.-P. Yuan, P. Li, Z.-L. Li, and J. Xia, “Data -driven risk -adeusted robust energy management for microgrids integrating demand response aggregator and renewable energies,” IEEE Trans. on Smart Grid, vol. 14, no. 1, pp. 365-377, 2022

    work page 2022

  20. [20]

    HPC data center participation in demand response: An adaptive policy with QoS assurance,

    Y . Zhang, D. C. Wilson, I. C. Paschalidis, and A. K. Coskun, “HPC data center participation in demand response: An adaptive policy with QoS assurance,” IEEE Trans. on Sustainable Computing, vol. 7, no. 1, pp. 157-171, 2021

    work page 2021

  21. [21]

    Characterizing power management opportunities for llms in the cloud

    P. Patel, E. Choukse, C. Zhang, Í. Goiri, B. Warrier, N. Mahalingam, and R. Bianchini, "Characterizing power management opportunities for llms in the cloud." in Proc. ASPLOS, La Jolla, CA, USA, 2024

    work page 2024

  22. [22]

    Empirical measurements of AI training power demand on a GPU-accelerated node,

    I. Latif, A. C. Newkirk, M. R. Carbone, A. Munir, Y . Lin, J. Koomey, X. Y u, and Z. Dong, “Empirical measurements of AI training power demand on a GPU - accelerated node,” arXiv preprint arXiv:2412.08602, 2024

    work page arXiv 2024

  23. [23]

    LLaMA: Open and Efficient Foundation Language Models

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachauo, T. Lacroio, B. Rozière, N. Goyal, E. Hambro, and F. Azhar, “Llama: Open and efficient foundation language models,” arXiv preprint arXiv:2302.13971, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  24. [24]

    Efficient memory management for large language model serving with pagedattention

    W. Kwon, Z. Li, S. Zhuang, Y . Sheng, L. Zheng, C. H. Y u, J. Gonzalez, H. Zhang, and I. Stoica, "Efficient memory management for large language model serving with pagedattention." in Proc. ACM SOSP, Koblenz, Germany, 2023

    work page 2023

  25. [25]

    2024 united states data center energy usage report,

    A. Shehabi, A. Newkirk, S. J. Smith, A. Hubbard, N. Lei, M. A. B. Siddik, B. Holecek, J. Koomey, E. Masanet, and D. Sartor, “2024 united states data center energy usage report,” 2024

    work page 2024

  26. [26]

    The price of robustness,

    D. Bertsimas, and M. Sim, “The price of robustness,” Operations research, vol. 52, no. 1, pp. 35-53, 2004

    work page 2004

  27. [27]

    Soft actor -critic: Off - policy maoimum entropy deep reinforcement learning with a stochastic actor

    T. Haarnoea, A. Zhou, P. Abbeel, and S. Levine, "Soft actor -critic: Off - policy maoimum entropy deep reinforcement learning with a stochastic actor." in Proc. ICML, Stockholm, Sweden, 2018

    work page 2018