{"id":"eb83ed1b-9b26-45f4-9580-d72997dc1672","arxiv_id":"2507.06416","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GPU frequency scaling in AI data centers can act as a distributed voltage-regulation resource, cutting voltage deviations in distribution systems during LLM workloads.","lead":"Data centers hosting AI workloads can help stabilize grid voltage by rapidly adjusting GPU clock speed and reactive power. The paper tests this on a 123-bus distribution feeder with real LLM inference power traces, showing smaller voltage swings than inverter-only control.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5) active-power droop as printed has the wrong sign for a load bus, and Eq. (6)'s S_t is not the cumulative term the text claims; unless the simulator implements a different law, the reported voltage gain cannot follow from the printed controller.","rationale":"I read this as a likely typo/notation problem rather than evidence of bad faith, but the load-bearing point is that the central claim cannot be evaluated from the manuscript as written. The reader's DVFS-latency concern is plausible and secondary; even perfect actuator tracking does not help if the printed droop law has the wrong sign or if the energy-balance integrator is missing. Because the code is open source, the conflict between Eq. (5) and the claimed results is directly checkable. I would keep the verdict CONDITIONAL but add an explicit requirement: the published equations must be shown to match the simulated controller. If variant (i) already reproduces Table I, my concern is retired and the reader's original conditions stand; if it does not, the paper needs a corrected Eq. (5) and Eq. (6) before the central claim can be assessed.","tokens_in":8556,"tokens_out":10006,"duration_ms":125651,"concrete_test":"Run the open-source simulator in https://github.com/chennnnnyize/voltage-regulation-with-data-centers on the same IEEE 123-bus/LLM scenario under three variants: (i) exactly Eq. (5) with -k_p(V-1) and S_t as the one-step mismatch in Eq. (6); (ii) the sign flipped to +k_p(V-1); (iii) Eq. (6) made cumulative, S_{t+1}=S_t+P^ref-uP. Compare Table I: if variant (i) does not reproduce the reported ~0.024 p.u. deviation and 12.8% gain while (ii) or (iii) does, then the published equations do not describe the validated controller.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"As printed, the local active-power law in Eq. (5) is internally inconsistent with the stated role of a data center as a consuming load. Under the standard LinDistFlow convention for positive load consumption, voltage falls when a bus draws more active power. Eq. (5) sets u^P = P^ref - k_p(V-1) + alpha*S, so when V>1 the data center reduces consumption (which pushes voltage up) and when V<1 it increases consumption (which pushes voltage down). That is positive feedback, not regulation. The law would only stabilize if p in Eq. (1) means net injection and data-center consumption enters as -uP, but Eqs. (3b), (3g), and the text explicitly call P 'active power consumption.' This is not a tuning or robustness issue; it is a mismatch at the exact point where the mechanism acts. A second inconsistency is Eq. (6): S_{i,t+1}=P^ref - uP is the one-step tracking error, while Section III says S_t keeps track of 'the power not served up to time t.' Enforcing the QoS/energy-balance constraint (3g) requires a cumulative integrator S_{t+1}=S_t+P^ref-uP. Without that, the printed law does not explain the delays reported in Table I. Since the simulator is open source, the most likely explanation is a sign/indexing typo, but the central quantitative claim is currently attached to equations that, taken literally, would not produce the claimed voltage improvement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a distributed, communication-free voltage regulation scheme for distribution systems with AI data center loads. Data centers adjust active power via GPU DVFS and reactive power via rectifiers, using local voltage measurements and simple droop laws. A cumulative energy-balance term is introduced to enforce that total data center energy consumption remains unchanged, and a reactive droop loop is used for inverter buses. The scheme is evaluated on an IEEE 123-bus system with real LLM inference power traces, reporting that more than 97% of buses stay within 0.05 p.u. voltage deviation and that the approach outperforms inverter-only droop control by more than 12.8%. The LLM power data and simulation code are open sourced.","tokens_in":8905,"tokens_out":6018,"duration_ms":66876,"significance":"If the results hold, the paper offers a timely and practical mechanism for using AI data centers as fast voltage-support resources, addressing a growing concern given the rapid deployment of large data centers. The use of real LLM inference traces and the open-sourcing of both data and code are notable strengths that aid reproducibility. The contribution is, however, primarily simulation-based, and the paper's central quantitative claims depend on resolving several technical inconsistencies in the controller equations and on clarifying assumptions about actuation speed and QoS modeling. With those clarifications, the work could be a useful reference for distribution grid operators and data center designers.","major_comments":[{"comment":"The sign convention in the voltage model and the active-power droop law is internally inconsistent. Eqs. (1) and (2) write v = Rp + Xq + 1 and v_{t+1} = R u^P_t + X u^Q_t + 1, while Section II-B and Eq. (3b) explicitly describe u^P as active power consumption. Under this convention, decreasing u^P when V > 1, as Eq. (5) does, would lower V further, producing positive feedback rather than regulation. If instead u^P is intended to be a net injection (with consumption entered as negative injection), then the notation P^ref as consumption and the energy-balance constraint (3g) must be reconciled. This is load-bearing because the reported voltage improvements are attributed to Eq. (5); the authors must correct either the voltage model or the control law and state the convention used in the simulation.","section":"Section II-B, Eqs. (1), (2), (5)"},{"comment":"The stated role of S_t and its update equation do not match. The text says S_{i,t} keeps track of the power not served up to time t, which requires a cumulative recurrence such as S_{t+1} = S_t + P^ref_t - u^P_{t+1}. Eq. (6), as printed, defines S_{i,t+1} = P^ref_{i,t} - u^P_{i+1,t}, which is only the one-step tracking error and cannot enforce the cumulative energy-balance constraint (3g). This discrepancy undermines the explanation of the query delays in Table I, which are attributed to the integral action. Please provide the exact recurrence used in the simulator and align Eq. (6) with it.","section":"Section III, Eq. (6)"},{"comment":"The adaptive gain alpha_{i,t} is not specified. The text gives only a qualitative description (\"if S_{i,t} is large ... alpha_{i,t} increased\" and conversely), with no update rule, initial value, or numerical values used in the simulations. Since alpha multiplies the cumulative energy term and directly influences the load-delay tradeoff reported in Table I, the results are not reproducible without this information. Please provide the adaptation law and the parameter values for each scenario.","section":"Section III, Eq. (5)"},{"comment":"The controller assumes that P^ref_{i,t+1} is exactly predictable at time t and that the DVFS command is realized within the 1-second control interval with no latency or tracking error. No measurement or model of GPU frequency-ramp time, scheduler delay, or actuation lag is provided. Because the simulated voltage improvements rely on the commanded active power being delivered at the next timestep, this idealization is load-bearing for the quantitative claims; the paper should either add a latency model or explicitly state this assumption as a limitation and discuss its impact.","section":"Section II-B and Section IV-A"},{"comment":"The reported \"Average Delay (s)\" is a key part of the claimed tradeoff between voltage regulation and QoS, but the manuscript does not describe how this delay is computed from the LLM workload and the control action. No model, formula, or simulator description is given, and the code location is not referenced in the text. Please define the delay metric and explain how it is derived, so that the tradeoff can be evaluated independently.","section":"Table I and Section IV-B"}],"minor_comments":[{"comment":"The index on u^P_{i+1,t} appears to be a typo; to match Eq. (3b), the update should be u^P_{i,t+1}.","section":"Section III, Eq. (5)"},{"comment":"The numerical entries in Table I contain formatting artifacts (e.g., \"0 .3\" and spacing before \"3\"), which should be cleaned up.","section":"Section IV-B, Table I"},{"comment":"The word \"signficant\" is misspelled and should be \"significant.\"","section":"Section IV-B"},{"comment":"The sentence \"We control in turn control set the clock frequency as a linear function of the voltage deviation\" is garbled and should be rewritten for clarity.","section":"Section III"},{"comment":"The claim that \"more than 97% of the buses always have a voltage deviation within 0.05 p.u.\" would be easier to verify if a quantile or empirical CDF of the voltage deviations were plotted or tabulated; the current figures do not directly support this specific number.","section":"Section IV-B"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a topical problem and has the virtue of open-sourcing both the LLM power data and the control simulation code. The main technical risk is that the printed controller equations—Eqs. (5) and (6)—are inconsistent with the stated sign convention and cumulative energy-balance description; these are likely fixable typos, but the authors must ensure the text matches the code before the paper can be trusted. The lack of a DVFS latency model is also worth pressing, since it affects the practical interpretation of the quantitative claims. I would encourage the editor to request a version in which the authors confirm the exact equations used in the simulation, ideally with a link to the relevant code lines."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things. First, this is a legitimate new application: using GPU DVFS as a local droop for active-power voltage support, validated with real LLM inference traces on a 123-bus feeder and open-sourced code. Second, the controller as printed is wrong. The stress-test note is correct. Eq. (5) sets u^P = P^ref - k_p(V-1) + ..., and since P is explicitly called \"active power consumption,\" the law reduces consumption when voltage is high and increases it when voltage is low. That is positive feedback. Eq. (6) also doesn't do what the text claims: S_{t+1} = P^ref - u^P is a one-step tracking error, not the cumulative \"power not served\" integrator needed to enforce the energy-balance constraint (3g). Taken literally, the printed controller would worsen voltage excursions.\n\nWhat's genuinely good: the problem is real (subsecond voltage transients from data center ramps), the DVFS lever is underexplored for grid support, and the paper uses real A100 inference measurements with nonlinear DistFlow simulation rather than a toy model. The open source release is a real plus.\n\nThe soft spots are mostly fixable but serious. The sign and indexing issues dominate; if the simulator uses net-injection sign convention, the authors need to correct the equations and define the notation carefully. Alpha tuning is unspecified — \"tuned\" with no update rule, no values, no sensitivity. Simulation reproducibility is weak: no seeds, no trial counts, no description of the load-scaling or snippet-selection method. The 1-second control loop assumes perfect next-step load prediction and immediate DVFS realization, with no latency or scheduling delay, which is optimistic but not disqualifying for a primary-control proposal.\n\nThe reader's conditional verdict is fair. The central idea is sound, the simulation is suggestive, but the paper cannot be accepted as written because the controller does not match its own equations. Everything points toward a presentation typo rather than a fundamental flaw, but \"probably a typo\" is not something a referee should have to guess. Send it to peer review. A good referee can sort out the sign convention, ask for the missing simulation details, and get it into shape. I'd want to see the revision before citing it in my own work.","headline":"Useful idea with real data, but the printed control law has a sign error that would make it positive feedback rather than regulation; fix the equations and it deserves a serious look.","tokens_in":9405,"tokens_out":3681,"would_cite":false,"duration_ms":41916,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that data centers can regulate distribution voltage by adjusting GPU clock frequency (DVFS) for active power and rectifiers for reactive power through a local, communication-free droop controller, keeping more than 97% of…","keywords":["voltage regulation","data centers","DVFS","distributed control","droop control","distribution systems","LLM inference load","reactive power"],"falsifier":"Measure DVFS command-to-power-response time on a GPU cluster and compare with the one-second control interval; if the realized power lags the commanded setpoint by more than a second, the simulated voltage deviations and the reported 12.8% gain over inverter-only control would not materialize in hardware.","tokens_in":8358,"feed_emoji":"⚡","tokens_out":4767,"duration_ms":48823,"temperature":0.7,"pith_summary":"The paper claims that the volatile power draw of AI data centers—GPU ramps that can swing within seconds—can be turned from a voltage problem into a voltage solution. By running a simple linear droop controller that reads local voltage and adjusts active power via dynamic voltage and frequency scaling (DVFS) and reactive power via rectifiers, each data center can counteract its own disturbances without real-time communication. In simulations on the IEEE 123-bus system with real LLM inference loads, more than 97% of buses stay within 0.05 p.u. of nominal voltage, and the scheme beats inverter-only droop control by at least 12.8% in average voltage deviation. If the hardware can execute the commanded frequency changes within the one-second control interval, data centers become a practical fast voltage-regulation resource at the distribution level.","feed_headline":"Data center DVFS keeps distribution voltages in bounds","feed_subtitle":"A local droop controller on GPU clocks and rectifiers beats inverter-only control by 12.8% in simulations.","key_machinery":"The controller is a linear droop law on active power, $u^P_{i,t+1} = [P^{\\mathrm{ref}}_{i,t} - k_{p,i}(V_{i,t}-1) + \\alpha_{i,t} S_{i,t}]_{\\underline{P}_i}^{\\overline{P}_i}$, where $S_{i,t}$ accumulates the unserved compute load so that total energy over the horizon stays invariant, and a standard reactive droop $u^Q_{i,t+1} = [u^Q_{i,t} - k_{q,i}(V_{i,t}-1)]_{\\underline{Q}_i}^{\\overline{Q}_i}$. The active-power knob is GPU clock frequency, which the paper treats as linearly related to power draw ($\\tilde{P}_i \\sim f_{\\mathrm{clk},i}$), so frequency can be set as a function of voltage deviation. The grid model is the LinDistFlow approximation $v = Rp + Xq + \\mathbf{1}$, with simulation validated against nonlinear branch flow equations.","core_discovery":"The central claim is that a communication-free, local droop controller—active power proportional to voltage deviation via DVFS, reactive power via rectifier injection—can keep distribution voltages near 1 p.u. even when data center loads swing violently. The paper demonstrates this in the IEEE 123-bus system with real LLM inference traces: with five data center buses, average voltage deviation stays at 0.024 p.u. versus 0.034 p.u. for inverter-only control, and bus 13, which leaves the [0.95, 1.05] p.u. band under inverter-only control, is pulled back inside when data centers participate. The control respects a total energy constraint designed to preserve quality of service, and it requires no communication: each bus acts on its own voltage measurement.","pith_inferences":["If DVFS response time is slower than the one-second interval, the effective droop gain is lower than designed, so a hardware-aware gain tuning or faster control interval would be needed in practice.","The same local droop logic could be applied to data center UPS/battery systems or cooling loads to provide transient voltage support during dips.","A testable extension is to run the controller on hardware-in-the-loop with actual GPU frequency ramps and compare the measured voltage profile against the simulated one."],"forward_implications":["Data centers can serve as primary voltage regulators without needing real-time communication, acting on local measurements only.","The scheme shifts LLM query load in time, so voltage regulation comes at the cost of mild inference delay; the gain $k_p$ controls the tradeoff.","Incorporating data center active-power control reduces the need for fast inverter-based reactive reserves in distribution feeders.","The approach extends to other controllable computing loads beyond GPUs, as long as power draw scales with a controllable clock frequency."],"supporting_citations":[{"why":"Provides the LinDistFlow approximation used for the grid voltage model in equation (1).","marker":"[28]"},{"why":"Supplies the nonlinear branch-flow equations used to validate the simulation results.","marker":"[33]"},{"why":"Gives the inverter-based reactive droop control baseline that the proposed scheme is compared against.","marker":"[30]"},{"why":"Documents the volatility of LLM power draws that motivates the need for fast voltage regulation.","marker":"[17]"},{"why":"Supplies the DVFS mechanism used as the active-power actuator for GPU frequency control.","marker":"[18]"},{"why":"Supports the assumption that GPU power scales linearly with clock frequency, making DVFS usable as a droop actuator.","marker":"[21]"}],"fun_headline_variants":["Local droop control keeps grid voltages in bounds under LLM load","Data center DVFS: no-communication voltage regulation","Droop controller on GPU clocks stabilizes distribution voltages","Voltage swings from data centers tamed by local DVFS droop","LLM load voltage: local DVFS beats inverter-only control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a data center can predict its next-step compute demand and that GPU frequency changes actually realize the commanded power within the one-second control interval, without added latency or hardware delay.","fun_headline_variants_meta":{"raw":{"variants":["Local droop control keeps grid voltages in bounds under LLM load","Data center DVFS: no-communication voltage regulation","Droop controller on GPU clocks stabilizes distribution voltages","Voltage swings from data centers tamed by local DVFS droop","LLM load voltage: local DVFS beats inverter-only control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001078,"raw_usage":{"total_tokens":4456,"prompt_tokens":833,"completion_tokens":3623,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":449,"completion_tokens_details":{"reasoning_tokens":3547}},"tokens_in":449,"tokens_out":3623,"duration_ms":24976,"temperature":1.0,"reasoning_tokens":3547,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:05:06.036779+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure DVFS command-to-power-response time on a GPU cluster and compare with the one-second control interval; if the realized power lags the commanded setpoint by more than a second, the simulated voltage deviations and the reported 12.8% gain over inverter-only control would not materialize in hardware.","supporting_citations":[{"cited_title":"A note on branch flow models with line shunts,","cited_arxiv_id":null,"evidence_quote":"Provides the LinDistFlow approximation used for the grid voltage model in equation (1)."},{"cited_title":"Optimal sizing of capacitors placed on a radial distribution system,","cited_arxiv_id":null,"evidence_quote":"Supplies the nonlinear branch-flow equations used to validate the simulation results."},{"cited_title":"Decentralized safe reinforcement learning for inverter-based voltage control,","cited_arxiv_id":null,"evidence_quote":"Gives the inverter-based reactive droop control baseline that the proposed scheme is compared against."},{"cited_title":"Static and dynamic frequency scal- ing on multicore cpus,","cited_arxiv_id":null,"evidence_quote":"Supplies the DVFS mechanism used as the active-power actuator for GPU frequency control."},{"cited_title":"Agarwal and J","cited_arxiv_id":null,"evidence_quote":"Supports the assumption that GPU power scales linearly with clock frequency, making DVFS usable as a droop actuator."}],"review_version":1}