Pith. sign in

REVIEW 5 major objections 5 minor 17 references

This paper claims that when grid dispatch is simulated with ten-minute, time-coupled optimal power flow, fast-ramping datacenter loads in a congestible load pocket force slow, expensive generators to become effectively must-run, raising ave

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 01:32 UTC pith:H7K3QSOU

load-bearing objection Useful modeling study, but the headline 8% cost increase isn't uniquely attributable to ramp constraints without a proper control. the 5 major comments →

arxiv 2607.28833 v1 pith:H7K3QSOU submitted 2026-07-30 eess.SY cs.SY

Hidden Economic Consequences of Adapting to Fast Ramping Datacenter Loads

Willa Gutowski , Charles Foltz , Nicholson Koukpaizan , Slaven Peles , Eve Tsybina This is my paper
classification eess.SY cs.SY
keywords datacenter load rampingACOPFtime-coupled optimizationramp constraintsmust-run generatorsload pocketsmarginal cost5000-bus test system
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tests how a 500 MW datacenter that can ramp at up to 5%/min affects dispatch on a congestible 5000-bus system, comparing hourly decoupled simulations with ten-minute time-coupled simulations that enforce ramping limits. The central claim is that the coupled simulations make slow, expensive generators in the West region behave as though they are must-run: they must be kept at higher output in early intervals to be ready for the datacenter's later ramp. As a result, average generator loading in the West rises by 3.3–4.0 percentage points and average marginal cost by 5.6–6.9 $/MWh, a 7–8% increase, concentrated in the peak-demand window. The paper also reports that all 100 South-area placements failed to converge in coupled runs, which it reads as evidence of a locally weak grid, and excludes them. The stakes are whether off-peak prices, not just peak prices, hide the true cost of firm datacenter connections.

Core claim

When nodal, transmission-constrained ACOPF must remain feasible over a sequence of ten-minute intervals, ramping constraints couple dispatch decisions across time. Under that coupling, a slow-ramping generator that is needed for a future datacenter ramp cannot wait until the ramp begins; it must be pre-positioned at higher output, and once ramped it often cannot come down quickly. In the West-area simulations, this 'must run' effect raises average loading from about 79% to about 83% and average marginal cost from about 88–90 $/MWh to about 95 $/MWh, a 7–8% increase, while the North shows no difference. The paper concludes that time-decoupled or relaxed optimization misses these costs and tha

What carries the argument

The organizing object is the time-coupled ACOPF (called TCOPFLOW in the paper) with a ten-minute discretization and generator ramping limits, contrasted with an hourly decoupled ACOPF (OPFLOW). The coupling is what does the work: it forces slow units to pre-ramp and stay at elevated output because feasibility in later intervals depends on earlier dispatch. Two metrics carry the comparison: percent generator loading (P_g/P_max) and marginal cost (b+2aP_g) at the dispatched point, plus an area import difference computed from line flows.

Load-bearing premise

The paper's load-bearing premise is that the non-convergence of all 100 South-area time-coupled cases is caused by local weak-grid infeasibility, not by numerical or modeling limits; if those cases were excluded for the wrong reason, the area-level dispatch and cost results could change.

What would settle it

Rerun the time-coupled ACOPF for the 100 excluded South-area datacenter placements with a tighter solver tolerance, a different optimal-power-flow algorithm, or a small grid modification, and check whether they converge; if they converge without showing must-run loading of expensive slow units, the claim that weak load pockets produce this effect loses support. Alternatively, look at wholesale market data in a real load pocket before and after a large fast-ramping datacenter connects: if off-peak locational marginal prices show no uplift beyond the peak window, the modeled effect is not visibl

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • In load pockets with slow, expensive generation, firm datacenter connections can raise off-peak production costs by about 8%, a cost not captured in peak-only or hourly-decoupled studies.
  • The must-run effect appears only in the peak-demand window; the steep ramp-stress window shows no dispatch difference, so the shape of the datacenter load profile determines whether the effect materializes.
  • Time-coupled, ten-minute ACOPF fails to converge for all 100 South-area placements, indicating that a weak local grid may make some candidate datacenter sites infeasible before economic analysis begins.
  • Because North shows no difference, well-connected low-cost areas can absorb fast ramps without changing dispatch, so interconnection quality is decisive for mitigating the effect.

Where Pith is reading between the lines

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

  • If the 100 non-convergent South cases are a solver or numerical artifact rather than physical infeasibility, excluding them could bias the area-level results; rerunning with a different solver, tolerance, or small grid modification is a direct test.
  • The must-run mechanism likely generalizes beyond datacenters to any large fast-ramping load, such as electrolyzers or electric-vehicle charging, in similarly constrained pockets.
  • In a real market, if the locational marginal price is set by the most expensive loaded unit, the price uplift could exceed the reported average marginal-cost increase, since the paper notes that loading of the most expensive West unit doubled.
  • The paper's suggested mitigation, battery storage in the load pocket, could be tested by adding a 100–500 MW storage resource next to the datacenter and measuring whether the must-run effect and cost increase shrink.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper compares temporally decoupled hourly ACOPF (OPFLOW) with time-coupled 10-minute ACOPF including generator ramping constraints (TCOPFLOW) on a modified, scaled 5000-bus IEEE 118-type system. It places large, fast-ramping datacenter loads at 350 candidate buses and reports that, in the West area, time coupling increases average generator loading by roughly 3.3–4.0 percentage points and average marginal cost by 5.6–6.9 $/MWh, an ~8% increase, which the authors attribute to slow, expensive generators becoming 'must run' to support datacenter ramps. The North area shows no difference, and all South-area coupled cases are excluded because they did not converge. The paper concludes that time-coupled analysis reveals hidden off-peak economic costs of datacenter ramping.

Significance. If the central result holds, the paper would provide a useful caution that standard hourly, temporally decoupled optimal power flow analyses may understate the system-cost and dispatch consequences of fast-ramping datacenter loads in congested, high-cost load pockets. The study covers many candidate locations (350) and presents a transparent scaling methodology with code availability, which are strengths. However, the current comparison conflates temporal resolution and load-profile representation with the presence of ramping constraints, and one results section directly contradicts the dispatch table for the steep window. These issues must be resolved before the economic-conclusion can be considered supported. The paper is a plausible contribution to the datacenter-grid-integration literature, but its headline quantitative claims are not yet uniquely established.

major comments (5)
  1. [§II.A and Table I (comparison protocol)] The central attribution of the West effect to ramping constraints is confounded. OPFLOW uses hourly average demand while TCOPFLOW uses the actual 10-minute datacenter profile with peaks up to 500 MW. The observed loading and marginal-cost increases could simply reflect the higher peak load seen by the coupled simulation. A control with effectively infinite ramping limits in TCOPFLOW, or an OPFLOW run at 10-minute resolution with the same load profile, is needed to isolate the ramp-constraint mechanism. Without such a control, the 'must-run' interpretation and the headline 8% cost increase are not uniquely supported.
  2. [§III.B vs Table I and Fig. 5] The steep/ramp-stress window results are internally inconsistent. Table I reports a 3.32 percentage-point loading increase and a 5.56 $/MWh marginal-cost increase for West in the steep window, yet §III.B states 'there is no difference between coupled and decoupled simulations, indicating that datacenters do not affect dispatch.' If this sentence refers only to the import difference, it should say so explicitly; as written it directly contradicts the dispatch results in Table I and undercuts the claim that both windows show the effect.
  3. [§III.A and abstract] The statement that 'loading of the most expensive West unit doubled' is not supported by any table, figure, or statistic in the manuscript. Since the abstract's 'up to 100% loading' depends on this unit-level claim, please provide unit-level dispatch results or remove the unsupported claim. The current Table I reports only area averages, which are a lower bound and do not justify the 'most expensive unit doubled' assertion.
  4. [§III, first paragraph (South exclusion)] All 100 South-area time-coupled cases are excluded because they 'did not converge on ramping scenarios due to the local weak grid.' Non-convergence is attributed to physical infeasibility without supporting evidence. The paper should justify this attribution, for example by reporting solver exit status, attempting a feasibility restoration, or showing that relaxed formulations also fail. The exclusion removes roughly 29% of candidate placements and should be stated as a limitation in the abstract and conclusions. The West-area central result is not directly affected, but the scope of the study's claims is.
  5. [Title, abstract, and §III.A] The paper claims 'economic consequences' and 'unexpectedly high system costs' but never reports total system production cost. Average marginal cost across generators is a proxy and does not directly yield system cost. Please report aggregate production cost for OPFLOW versus TCOPFLOW for each window and datacenter placement, or revise the economic-consequence language to match the proxy metrics actually computed.
minor comments (5)
  1. [§II.A, Eqs. (3)–(5)] The terminology 'module' versus 'area' is inconsistent, and the definition of D_m in Eq. (5) does not state units or explicitly clarify that the max is over hourly aggregated values. Please add units and define terms precisely.
  2. [§III.A] The phrase 'up to four-digit precision' is unclear; it should say 'to four decimal places' or similar.
  3. [Fig. 5] The r values are reported without a clear description of what is being correlated (import values? deviations from the diagonal?) and with no formal definition. Add axis labels, a legend, and define r in the text.
  4. [Fig. 3] The 'average total cost per MW at Pmax' used as a cost indicator is not defined. Please specify the formula and whether it is the same as total cost divided by Pmax.
  5. [References and code availability] The manuscript states that the full scaling code is in an online appendix, but no URL or repository identifier is given beyond the ExaGO software DOI. Please provide a direct link or DOI for the scaling code.

Circularity Check

0 steps flagged

No significant circularity; the claimed 8% cost increase emerges from a simulation comparison and is not equivalent to the model inputs.

full rationale

I walked the derivation chain: the paper constructs a modified IEEE 118-bus system, scales it to 5,000 buses, defines generator costs and ramping limits, places datacenter loads at candidate buses, and then compares decoupled OPFLOW with temporally coupled TCOPFLOW solutions. The reported outputs — generator loading percentages (Eq. 1), marginal costs (Eq. 2), and import differences (Eqs. 3–5) — are computed from the optimized dispatch decisions, not fitted to reproduce the headline 8% marginal-cost increase. No parameter is calibrated to the target result, and no equation reduces to another by construction. The only overlap with the authors' prior work is the ExaGO solver (Ref. [13]), which is used as a computational tool rather than as an argument for the conclusion; its use does not import a uniqueness theorem, an ansatz, or a precomputed result. The exclusion of the 100 non-converging South-area cases and the fact that the decoupled baseline uses hourly average load while the coupled simulation uses the 10-minute profile are legitimate methodological or confound concerns, but they are not circularity: they concern whether the comparison isolates ramping constraints, not whether the output is defined into the input. Thus the central claim has independent content and the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The paper's result is entirely produced by a simulation model. It introduces no new physical entities, but depends on many unvalidated modeling choices: the modified IEEE 118-bus grid as a stand-in for real RTOs, specific cost/ramp parameters, the choice of load windows, and the interpretation of solver non-convergence as physical infeasibility. These are domain assumptions, not fitted parameters, but they determine the 8% figure.

free parameters (4)
  • Datacenter load size and ramp rate = 0-500 MW, up to 5%/min
    Load profile taken from Frontier dataset and industry projections; fixed by authors, not fitted, but determines the stress applied.
  • Four-hour peak and steep windows = two intervals from Frontier data
    Selected by maximum demand and maximum swing; choice affects which periods show coupled/decoupled divergence.
  • West-area generator ramping capability = 10-20 MW/min
    Assigned in the modified 118-bus model; slower ramping is the mechanism behind the must-run effect.
  • Generator cost coefficients = not listed
    Quadratic for fossil, linear for others, drawn from the IEEE 118-bus operating manual but the actual coefficients are not reported in the paper, directly determining marginal cost values.
axioms (5)
  • domain assumption The modified IEEE 118-bus model with added congestion, cost, and ramp constraints is representative of a real RTO (e.g., PJM/MISO) for studying datacenter impacts.
    Section II.B; the paper builds the 5000-bus system from this model and extrapolates conclusions to real load pockets.
  • domain assumption Non-convergence of TCOPFLOW for South cases indicates genuine infeasibility due to weak grid, not solver numerical failures.
    Section III: 'all 100 time-coupled cases associated with the South area did not converge on ramping scenarios due to the local weak grid.'
  • domain assumption Replicating the 118-bus module 50 times yields a 5000-bus system whose congestion and ramping behavior is representative of an RTO.
    Section II.B: 'extend... to 5000 buses... by replicating it 50 times and interconnecting through transmission corridors.'
  • domain assumption The Frontier supercomputer load profile is representative of datacenter load ramping in general.
    Section I Fig. 1(b) and Section II.A windows; the simulation uses two four-hour intervals from this dataset.
  • domain assumption Ten-minute temporal discretization is sufficient to capture datacenter ramping constraints.
    Section II.A: 'ten minutes was selected so that our ramp values used would be consistent with [12].'

pith-pipeline@v1.3.0-alltime-deepseek · 6436 in / 13229 out tokens · 136867 ms · 2026-08-03T01:32:17.560939+00:00 · methodology

0 comments
read the original abstract

Artificial intelligence workloads are driving the rapid expansion of datacenter infrastructure, which imposes substantial stress on the US energy system. While high peak electricity prices are an anticipated outcome, measurable under peak hour simulations, the high off peak prices are a significantly underestimated threat. We simulate different ramping conditions on a congestible 5000-bus system, based on a modified IEEE 118-bus grid, to show that, in the presence of fast ramping loads and slow ramping generation, datacenters can aggravate latent load pockets. This results in unexpectedly high system costs during periods outside of datacenter peak. We test the datacenter effects using two distinct load conditions. We find that in the system coincident peak, coupled simulations result in up to 100% loading of slow expensive units, with an average marginal cost increase of 8%. These findings are of extreme importance as they reveal the hidden costs of preventively ramping slow generation in anticipation of datacenter load changes.

Figures

Figures reproduced from arXiv: 2607.28833 by Charles Foltz, Eve Tsybina, Nicholson Koukpaizan, Slaven Peles, Willa Gutowski.

Figure 1
Figure 1. Figure 1: (a) Distribution of nominal capacity, MW, for data centers reported [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Generation cost and ramping by area The distribution of ramping and cost characteristics is shown in [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: 5,000-bus extended test system. The dark blue box denotes the original [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of fully decoupled and fully coupled module import [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

17 extracted references · 1 linked inside Pith

  1. [1]

    FERC technical conference regarding the challenge of resource ade- quacy in RTO and ISO regions,

    “FERC technical conference regarding the challenge of resource ade- quacy in RTO and ISO regions,” Federal Energy Regulatory Commis- sion, Jun. 2025

    2025

  2. [2]

    2026 long- term reliability assessment,

    North American Electric Reliability Corporation, “2026 long- term reliability assessment,” NERC, Tech. Rep., January

    2026

  3. [3]

    Processing power: The effect of data centers on wholesale electricity markets,

    K. Owen, R. Reaser, and R. Taylor, “Processing power: The effect of data centers on wholesale electricity markets,” Mar. 2026. [Online]. Available: https://www.dallasfed.org/research/papers/2026/wp2606

    2026

  4. [4]

    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,” Jul. 2026. [Online]. Available: https://arxiv.org/abs/2509.07218

    Pith/arXiv arXiv 2026

  5. [5]

    A guide to reducing carbon emissions through data center geographical load shifting,

    J. Lindberg, Y . Abdennadher, J. Chen, B. C. Lesieutre, and L. Roald, “A guide to reducing carbon emissions through data center geographical load shifting,” inProceedings of the Twelfth ACM International Confer- ence on Future Energy Systems. Association for Computing Machinery, Jun. 2021, pp. 430–436

    2021

  6. [6]

    Technical challenges of AI data center integration into power grids—a survey,

    E. Ginzburg-Ganz, P. Lifshits, R. Machlev, J. Belikov, Z. Krieger, and Y . Levron, “Technical challenges of AI data center integration into power grids—a survey,”Energies, vol. 19, no. 1, p. 137, Jan. 2026

    2026

  7. [7]

    The impact of large-scale data center load integration into electrical power networks,

    M. S. Hossain, M. M. Rahman, M. R. Islam, D. Sutanto, and K. M. Muttaqi, “The impact of large-scale data center load integration into electrical power networks,” in2025 IEEE International Conference on Energy Technologies for Future Grids (ETFG), Dec. 2025, pp. 1–6

    2025

  8. [8]

    Load management of data centers as regulation capacity in denmark,

    A. Clausen, G. Ghatikar, and B. N. Jørgensen, “Load management of data centers as regulation capacity in denmark,” inInternational Green Computing Conference, Nov. 2014, pp. 1–10. [Online]. Available: https://ieeexplore.ieee.org/document/7039161

    arXiv 2014

  9. [9]

    Flexibility from networks of data centers: A market clearing formula- tion with virtual links,

    W. Zhang, L. A. Roald, A. A. Chien, J. R. Birge, and V . M. Zavala, “Flexibility from networks of data centers: A market clearing formula- tion with virtual links,”Electric Power Systems Research, vol. 189, p. 106723, Dec. 2020

    2020

  10. [10]

    Distributed economic dispatch in power networks incorporating data center flexibility,

    A. Tsiligkaridis, P. Andrianesis, A. K. Coskun, M. C. Caramanis, and I. C. Paschalidis, “Distributed economic dispatch in power networks incorporating data center flexibility,”IEEE Transactions on Sustainable Computing, vol. 10, no. 4, pp. 768–783, Jul. 2025

    2025

  11. [11]

    Datacenters and energy report,

    S&P Global Market Intelligence, “Datacenters and energy report,” S&P Global Market Intelligence, Tech. Rep., 2025, subscription based

    2025

  12. [12]

    Energy dataset of frontier supercomputer for waste heat recovery,

    J. Sun, Z. Gao, D. Grant, K. Nawaz, P. Wang, C.-M. Yang, P. Boudreaux, S. Kowalski, and S. Huff, “Energy dataset of frontier supercomputer for waste heat recovery,”Scientific Data, vol. 11, no. 1, p. 1077, Oct. 2024

    2024

  13. [13]

    Exago v2,

    S. Peles, N. Koukpaizan, M. Alam, J. Hambrick, and E. Tsybina, “Exago v2,” [Computer Software] https://doi.org/10.11578/dc.20260223.2, 2026

    doi:10.11578/dc.20260223.2 2026

  14. [14]

    IEEE 118-bus three-area test system,

    “IEEE 118-bus three-area test system,” University of Washington. [Online]. Available: https://labs.ece.uw.edu/pstca/

  15. [15]

    A real-time operations manual for the ieee 118 bus transmission model,

    A. Anderson, S. Kincic, B. Jefferson, B. Mcgary, C. Fallon, D. Ciesiel- ski, J. Wenskovitch, and Y . Chen, “A real-time operations manual for the ieee 118 bus transmission model,” Pacific Northwest National Laboratory, Richland, Washington 99354, Tech. Rep. PNNL-334996, 2022

    2022

  16. [16]

    [Online]

    The natural gas war on coal. [Online]. Available: https://kleinmanenergy. upenn.edu/commentary/blog/the-natural-gas-war-on-coal/

  17. [2026]

    Available: https://www.nerc.com/globalassets/our-work/ assessments/nerc ltra 2025.pdf

    [Online]. Available: https://www.nerc.com/globalassets/our-work/ assessments/nerc ltra 2025.pdf

    2025