arxiv: 2605.03090 · v1 · submitted 2026-05-04 · 💻 cs.DC · cs.AI· cs.SY· eess.SY

From Barrier to Bridge: The Case for AI Data Center/Power Grid Co-Design

Noman Bashir , Rob Sherwood , Le Xie , Minlan Yu This is my paper

Pith reviewed 2026-05-08 17:19 UTC · model grok-4.3

classification 💻 cs.DC cs.AIcs.SYeess.SY

keywords AI data centerspower gridload diversityco-designgrid stabilityhyperscale computingcompute-power integration

0 comments p. Extension

The pith

AI training data centers break the grid's load diversity assumption, requiring explicit co-design with power systems.

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

AI training data centers create power demands that swing by hundreds of megawatts in seconds from a single synchronized job, unlike the smooth aggregates from diverse small users that grids have relied on for over a century. The paper argues this breaks the statistical foundation of grid stability and requires the data center and power industries to shift from separate operations to joint development of planning, controls, and markets. Readers should care because continued separation risks unreliable electricity supply or limits on AI scaling. The work details differing design principles and incentives that make coordination hard and outlines research needs in capacity planning and protocols.

Core claim

AI training data centers break the load diversity assumption because a single hyperscale training campus draws power comparable to a mid-sized city with rapid swings driven by one tightly synchronized job. This entanglement requires a shift from implicit coexistence to explicit co-development between the data center and electric power industries, with new approaches to capacity planning, control, protocols, and markets to support sustainable and reliable AI.

What carries the argument

The load diversity assumption in electric grids, which produces predictable aggregate demand from uncorrelated small loads, now violated by synchronized large AI workloads and requiring co-design.

If this is right

Joint capacity planning becomes necessary to handle AI load peaks without excessive overbuilding.
Multi-timescale control systems must be developed to manage rapid power swings from compute jobs.
A compute-power protocol stack is needed for coordination across different operational timescales.
Market innovations are required to align the economic incentives of the two sectors.

Where Pith is reading between the lines

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

Successful co-design could allow faster scaling of AI training without running into grid bottlenecks.
Existing demand-response programs may prove too slow or limited for the speed and magnitude of AI-induced swings.
This shift might influence how future large computing facilities are sited, permitted, and operated.

Load-bearing premise

Power swings from individual hyperscale AI campuses are large and rapid enough to threaten grid stability under existing practices, and misalignments between industries cannot be overcome with current coordination tools.

What would settle it

Operational data from a large AI training facility demonstrating that its power demand fluctuations do not exceed the grid's ability to maintain stability using standard reserves and controls.

read the original abstract

For over a century, the electric grid has relied on a single statistical assumption: \emph{load diversity}, the principle that the uncorrelated demands of millions of small consumers produce a smooth, predictable aggregate. AI training data centers break that assumption. A single hyperscale training campus can draw power comparable to a mid-sized city, driven by one tightly synchronized job whose demand swings by hundreds of megawatts in seconds. This paper argues that the resulting entanglement of compute and power infrastructure requires a shift from implicit coexistence to explicit co-development between the historically decoupled data center and electric power industries. We introduce the distinct design principles, operational philosophies, and economic incentives of each sector, and show why their cultural and technical misalignment makes coordination difficult. We identify key research directions, from joint capacity planning, multi-timescale control, a compute--power protocol stack, to market innovation, that must be pursued to power the future of AI sustainably and reliably.

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 flags how synchronized AI training loads break the grid's load diversity assumption and pushes for explicit co-design, but it stays qualitative with no numbers or simulations to show the scale of the risk.

read the letter

The core point is that a single hyperscale training campus can swing hundreds of megawatts in seconds because the workload is tightly synchronized, unlike the uncorrelated small loads the grid has always relied on. The authors lay out why this creates real entanglement between compute and power systems that current separate planning cannot handle well. That framing is the main new angle here, even if calls for better coordination have appeared before in grid and data-center work.

Referee Report

1 major / 1 minor

Summary. The paper claims that AI training data centers break the electric grid's longstanding load diversity assumption, as a single hyperscale training campus can draw power comparable to a mid-sized city with synchronized demand swings of hundreds of megawatts on seconds-scale timescales. This creates an entanglement between compute and power infrastructure that necessitates a shift from implicit coexistence to explicit co-development between the historically decoupled data center and electric power industries. The manuscript introduces the distinct design principles, operational philosophies, and economic incentives of each sector, explains the resulting cultural and technical misalignment, and identifies key research directions including joint capacity planning, multi-timescale control, a compute-power protocol stack, and market innovation.

Significance. If the premise holds that AI training power characteristics exceed the response capabilities of existing grid mechanisms, the paper could catalyze valuable interdisciplinary work on sustainable AI scaling by clearly framing the problem and outlining a research roadmap. It earns credit for articulating sector differences and proposing concrete directions without relying on invented entities or circular derivations, though its impact is tempered by the lack of quantitative grounding.

major comments (1)

[Abstract] Abstract: The assertion that AI training campuses produce power swings large and fast enough to 'materially threaten grid stability under current operating practices' (and that standard mechanisms such as spinning reserves, demand response, or existing curtailment contracts are insufficient) is presented without any supporting quantitative modeling, power-flow simulations, frequency-response analysis, or comparisons to NERC/ISO reserve margins, inertia constants, or AGC bandwidth. This quantification is load-bearing for the central claim that explicit co-design is required.

minor comments (1)

The discussion of current industry practices would be strengthened by citing specific prior studies on data-center demand response programs and grid integration efforts.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We appreciate the referee's detailed review and the recognition of the paper's contribution in framing the co-design challenge between AI data centers and the power grid. We address the major comment below and outline our planned revisions.

read point-by-point responses

Referee: The assertion that AI training campuses produce power swings large and fast enough to 'materially threaten grid stability under current operating practices' (and that standard mechanisms such as spinning reserves, demand response, or existing curtailment contracts are insufficient) is presented without any supporting quantitative modeling, power-flow simulations, frequency-response analysis, or comparisons to NERC/ISO reserve margins, inertia constants, or AGC bandwidth. This quantification is load-bearing for the central claim that explicit co-design is required.

Authors: We agree with the referee that the central claim would benefit from more explicit quantitative support. The manuscript is a position paper intended to articulate the problem and outline a research roadmap rather than to conduct power systems engineering analysis. The described power swings are based on documented characteristics of hyperscale AI training workloads, where synchronized GPU clusters can exhibit rapid power ramps on the order of hundreds of MW within seconds, as reported in industry disclosures and technical literature on large language model training. Standard grid mechanisms such as automatic generation control (AGC) operate on timescales of seconds to minutes but are calibrated for smaller, diverse load fluctuations, while spinning reserves address contingency events rather than routine operational variability. We acknowledge the absence of original simulations in the current version. In revision, we will update the abstract to include specific scale references and add citations to studies on data center power profiles and grid reserve requirements, thereby providing a clearer foundation for the need for co-design while noting that detailed modeling and simulations represent important future research directions identified in the paper. revision: partial

Circularity Check

0 steps flagged

No circularity: position paper relies on domain facts without self-referential derivation

full rationale

The paper is an argumentative position piece that asserts AI training workloads break the long-established load diversity principle of the electric grid and therefore require explicit co-design. It describes industry differences, operational philosophies, and lists research directions. No equations, fitted parameters, derivations, or mathematical reductions appear. The central claim is an empirical characterization presented as input rather than a result derived from prior steps within the paper. No self-citations, ansatzes, or renamings of known results are load-bearing in a way that creates circularity. This is the expected non-finding for a non-mathematical advocacy paper.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the established engineering premise that uncorrelated consumer loads produce aggregate predictability; no free parameters, new entities, or ad-hoc axioms are introduced.

axioms (1)

domain assumption Load diversity from uncorrelated small consumers produces a smooth, predictable aggregate demand that the grid can reliably serve.
Invoked in the opening sentence as the century-old statistical foundation of grid operation.

pith-pipeline@v0.9.0 · 5474 in / 1271 out tokens · 31042 ms · 2026-05-08T17:19:16.264520+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

38 extracted references · 13 canonical work pages

[1]

Amazon.com, Inc. 2025. Amazon.com Announces First Quarter Results. Press Release. https://ir.aboutamazon.com/ news-release/news-release-details/2025/Amazon-com-Announces-First-Quarter-Results/default.aspx

2025
[2]

2019.The Datacenter As a Computer: Designing Warehouse-scale Machines

Luiz André Barroso, Urs Hölzle, and Parthasarathy Ranganathan. 2019.The Datacenter As a Computer: Designing Warehouse-scale Machines. Springer Nature

2019
[3]

Noman Bashir, Tian Guo, Mohammad Hajiesmaili, David Irwin, Prashant Shenoy, Ramesh Sitaraman, Abel Souza, and Adam Wierman. 2021. Enabling Sustainable Clouds: The Case for Virtualizing the Energy System. InProc. of the ACM Symposium on Cloud Computing(Seattle, WA, USA)(SoCC ’21). Association for Computing Machinery, New York, NY, USA, 350–358. https://doi...

work page doi:10.1145/3472883.3487009 2021
[4]

2025.Today’s Outlook

California Independent System Operator. 2025.Today’s Outlook. https://www.caiso.com/todays-outlook Accessed: 2025-09-05

2025
[5]

Diane Cardwell. 2025. Data center activity ‘exploded’ in Texas, spiking reliability risks: monitor. Utility Dive, https: //www.utilitydive.com/news/data-center-activity-has-exploded-in-ercot-spiking-grid-reliability-risk/752780/. Quotes David Penny, Director of Reliability Services for Texas RE

2025
[6]

Esha Choukse, Brijesh Warrier, Scot Heath, Luz Belmont, April Zhao, Hassan Ali Khan, Brian Harry, Matthew Kappel, Russell J. Hewett, Kushal Datta, Yu Pei, Caroline Lichtenberger, John Siegler, David Lukofsky, Zaid Kahn, Gurpreet Sahota, Andy Sullivan, Charles Frederick, Hien Thai, Rebecca Naughton, Daniel Jurnove, Justin Harp, Reid Carper, Nithish Mahalin...

work page arXiv 2025
[7]

Electric Power Research Institute. 2025. Flex MOSAIC: A Flexibility Classification Framework for Large Electric Loads. DCFlex Initiative. https://www.tdworld.com/distributed-energy-resources/demand-side-management/news/ 55366867/epri-launches-framework-to-reduce-time-to-power-for-data-centers. Framework developed with more than 65 utilities, system operat...

2025
[8]

Energy Systems Integration Group. 2025. Engaging with Large Loads. ESIG briefing paper. https://www.esig.energy/ engaging-with-large-loads/. Documents voltage-sensitive behavior of large electronic loads in ERCOT and the resulting reliability concerns. Accessed: 2026-04-19

2025
[9]

Melissa Farney. 2025. The Power Play: How Data Centers and Utilities Are Reinventing Energy Strate- gies. https://www.datacenterfrontier.com/energy/article/55277929/the-power-play-how-data-centers-and-utilities- are-reinventing-energy-strategies

work page arXiv 2025
[10]

2025.Major energy rate case decisions in the US January-June 2025: Quarterly update on decided rate cases

Lisa Fontanella. 2025.Major energy rate case decisions in the US January-June 2025: Quarterly update on decided rate cases. Technical Report. S&P Global Market Intelligence, Regulatory Research Associates. https://psc.ky.gov/pscecf/2025- 00113/kyle.j.smith124.civ%40army.mil/09232025010656/MPG_Copyright_Protected_WP_27.pdf Contributors: Jim Davis, Heike Do...

work page arXiv 2025
[11]

Grid Status. 2025. How Large Data Center Loads Are Surfacing New Issues. https://blog.gridstatus.io/byte-blackouts- large-data-center-loads-new-issues-pjm/. Analysis of the July 2024 PJM / Dominion data center load-loss event. , Vol. 1, No. 1, Article . Publication date: May 2018. 12 Bashir, Sherwood, Xie, and Yu Accessed: 2026-04-18

2025
[12]

https://www.datacenters.com/. [n. d.]. Microsoft’s $80B Capex Signals Continued Cloud Expansion and Strategic AI Infrastructure Bet. https://www.datacenters.com/news/microsoft-s-80b-capex-signals-continued-cloud-expansion- and-strategic-ai-infrastructure-bet
[13]

2022.Global EV Outlook 2022

International Energy Agency. 2022.Global EV Outlook 2022. Technical Report. IEA, Paris. https://www.iea.org/ reports/global-ev-outlook-2022 Licence: CC BY 4.0

2022
[14]

2023.Electricity Grids and Secure Energy Transitions

International Energy Agency. 2023.Electricity Grids and Secure Energy Transitions. Technical Report. IEA, Paris. https://www.iea.org/reports/electricity-grids-and-secure-energy-transitions Licence: CC BY 4.0

2023
[15]

2025.Energy and AI

International Energy Agency. 2025.Energy and AI. Technical Report. IEA, Paris. https://www.iea.org/reports/energy- and-ai Projects global data center electricity consumption to more than double from 415 TWh in 2024 to approximately 945 TWh by 2030, with AI-focused data centers driving most of the growth. Licence: CC BY 4.0

2025
[17]

Srinivasan Keshav and Catherine Rosenberg. 2010. How internet concepts and technologies can help green and smarten the electrical grid. InProceedings of the First ACM SIGCOMM Workshop on Green Networking(New Delhi, India)(Green Networking ’10). Association for Computing Machinery, New York, NY, USA, 35–40. https://doi.org/10. 1145/1851290.1851298

work page arXiv 2010
[18]

Xuhui Lin, Yue Wang, Xin Chen, Nan Li, Enrique Mallada, and Mung Chiang. 2025. AI data centres as grid-interactive assets.Nature Energy(2025). https://doi.org/10.1038/s41560-025-01927-1 Demonstrates that workload orchestration alone can enable data centers to act as grid-flexible assets without hardware modifications or energy storage

work page doi:10.1038/s41560-025-01927-1 2025
[19]

Alexandra Sasha Luccioni, Yacine Jernite, and Emma Strubell. 2024. Power Hungry Processing: Watts Driving the Cost of AI Deployment?. InProceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT). 85–99. https://doi.org/10.1145/3630106.3658542 Empirical measurements of inference energy across task types and model sizes. ht...

work page doi:10.1145/3630106.3658542 2024
[20]

Cy McGeady. 2024. Powering the Commanding Heights: The Strategic Context of Emergent U.S. Electricity Demand Growth. https://www.csis.org/analysis/powering-commanding-heights-strategic-context-emergent-us-electricity- demand-growth. Accessed: November 20, 2025

2024
[21]

2024.Meta’s Richland Parish Data Center

Meta Platforms, Inc. 2024.Meta’s Richland Parish Data Center. Technical Report. Meta. https://datacenters.atmeta. com/wp-content/uploads/2024/12/Metas-Richland-Parish-Data-Center.pdf Accessed: 2025-09-05

2024
[22]

Microsoft. 2025. Microsoft Earnings Release FY25 Q3. Press Release. https://www.microsoft.com/en-us/investor/ earnings/fy-2025-q3/press-release-webcast

2025
[23]

National Renewable Energy Laboratory. 2022. ResStock: A Highly Granular Analysis of the U.S. Residential Building Stock. https://www.nrel.gov/buildings/resstock.html. Hourly residential load simulations used to estimate coincidence and diversity factors across U.S. feeders. Accessed: 2026-04-19

2022
[24]

Norris, Tim Profeta, Dalia Patino-Echeverri, and Adam Cowie-Haskell

Tyler H. Norris, Tim Profeta, Dalia Patino-Echeverri, and Adam Cowie-Haskell. 2025.Rethinking Load Growth: Assessing the Potential for Integration of Large Flexible Loads in US Power Systems. Technical Report. Nicholas Institute for Energy, Environment & Sustainability, Duke University. Estimates that existing U.S. power systems could accommodate roughly ...

2025
[26]

Effectiveness and Safety of Lattice Radiotherapy in Treating Large Volume Tumors: A Systematic Review and Meta -Analysis,

North American Electric Reliability Corporation. 2025.2025 State of Reliability Report. Technical Report. NERC. Documents the July 2024 Northern Virginia event in which approximately 1.5 GW of voltage-sensitive data center load tripped offline within seconds. https://www.nerc.com/pa/RAPA/PA/Performance%20Analysis%20DL/NERC_SOR_ 2025.pdf

work page arXiv 2025
[27]

North American Electric Reliability Corporation. 2025. Industry Recommendation on Large Load Interconnections. NERC Alert / Industry Recommendation. First formal NERC recommendation specifically addressing data center and computational-load interconnection risk. Discussed at https://www.jdsupra.com/legalnews/nerc-continues-focus-on- large-load-8238000/

2025
[28]

2024.Stargate Advances with Partnership with Oracle

OpenAI. 2024.Stargate Advances with Partnership with Oracle. https://openai.com/index/stargate-advances-with- partnership-with-oracle/ Accessed: 2025-09-05

2024
[29]

OpenAI. 2025. OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites. https://openai.com/ index/five-new-stargate-sites/

2025
[30]

Dylan Patel, Jeremie Eliahou Ontiveros, and Daniel Nishball. 2024. Datacenter Anatomy Part 1: Electrical Systems. https://newsletter.semianalysis.com/p/datacenter-anatomy-part-1-electrical. , Vol. 1, No. 1, Article . Publication date: May 2018. From Barrier to Bridge: The Case for AI Data Center/Power Grid Co-Design 13

2024
[31]

David Patterson, Joseph Gonzalez, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David So, Maud Texier, and Jeff Dean. 2022. The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink.Computer 55, 7 (2022), 18–28. https://doi.org/10.1109/MC.2022.3148714 Argues that inference dominates long-run ML energy once models are deplo...

work page doi:10.1109/mc.2022.3148714 2022
[32]

Patterson, Garth Gibson, and Randy H

David A. Patterson, Garth Gibson, and Randy H. Katz. 1988. A Case for Redundant Arrays of Inexpensive Disks (RAID). InProceedings of the 1988 ACM SIGMOD International Conference on Management of Data. ACM, New York, NY, USA, 109–116. https://doi.org/10.1145/50202.50214

work page doi:10.1145/50202.50214 1988
[33]

2025.2024 State of the Market Report for the ERCOT Electricity Markets

Potomac Economics. 2025.2024 State of the Market Report for the ERCOT Electricity Markets. Technical Report. Public Utility Commission of Texas. See Section VI.E.1 for details on the $5,000/MWh System-Wide Offer Cap

work page arXiv 2025
[34]

Ana Radovanović, Ross Koningstein, Ian Schneider, Bokan Chen, Alexandre Duarte, Binz Roy, Diyue Xiao, Maya Haridasan, Patrick Hung, Nick Care, Saurav Talukdar, Eric Mullen, Kendal Smith, Mariëlle Cottman, and Walfredo Cirne. 2022. Carbon-Aware Computing for Datacenters.IEEE Transactions on Power Systems38, 2 (2022), 1270–1280. https://doi.org/10.1109/TPWR...

work page doi:10.1109/tpwrs.2022.3173250 2022
[35]

Mahadev Satyanarayanan. 2017. The Emergence of Edge Computing.Computer50, 1 (2017), 30–39. https://doi.org/10. 1109/MC.2017.9

2017
[36]

2024.2024 Work Trend Index Annual Report

Arman Shehabi, Alex Hubbard, Alex Newkirk, Nuoa Lei, Md Abu Bakkar Siddik, Billie Holecek, Jonathan Koomey, Eric Masanet, Dale Sartor, et al. 2024.2024 United States Data Center Energy Usage Report. Technical Report. Lawrence Berkeley National Laboratory. https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data- center-ene...

work page arXiv 2024
[37]

MacKenzie Sigalos. 2025. Amazon opens $11 billion AI data center in rural Indiana as rivals race to break ground. https://www.cnbc.com/2025/10/29/amazon-opens-11-billion-ai-data-center-project-rainier-in-indiana.html

2025
[38]

Marc Spieler. 2025. How AI Factories Can Help Relieve Grid Stress. https://resources.nvidia.com/en-us-energy- utilities/ai-factories-flexible

2025
[39]

Energy Information Administration

U.S. Energy Information Administration. 2025. Hourly Electric Grid Monitor. https://www.eia.gov/electricity/ gridmonitor/dashboard. Accessed: November 10, 2025

2025
[40]

Lee Willis

H. Lee Willis. 2004.Power Distribution Planning Reference Book(2nd ed.). CRC Press, Boca Raton, FL. Defines load diversity and diversity factor; typical residential feeder diversity factors of 1.4–3.0 are discussed in Chapter 3.. , Vol. 1, No. 1, Article . Publication date: May 2018

2004