{"id":"07683f42-7d76-4bb4-bb1d-717200a76959","arxiv_id":"2412.06288","paper_version":4,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"U.S. data centers could cause more than $20 billion per year in public health costs by 2028, with low-income counties bearing a disproportionate share.","lead":"This paper estimates the air pollution and health costs caused by U.S. data centers, projecting that the annual public health burden could exceed $20 billion by 2028. It also proposes scheduling computer workloads to avoid times and places where electricity generation does the most health damage.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Scope-2 estimate hinges on attributing every region's average plant mix to data centers and freezing the 2023 geographic load distribution through 2028; either assumption can shift the headline $20.9B and the county map.","rationale":"The paper is a transparent screening-level quantification that correctly identifies a real and growing externality, and its use of EPA COBRA is appropriate for the stated purpose. The reader's conditional verdict is well calibrated: the central quantitative claim, that U.S. data centers will impose $11.7B to $20.9B in annual public health costs by 2028, depends almost entirely on the scope-2 attribution. My stress-test pass finds no internal inconsistency in the arithmetic or in the COBRA application; the issue is that the method's two most consequential assumptions, regional average plant attribution and a static 2023 geographic load distribution, are both reasonable defaults but are neither justified against alternatives nor bounded by sensitivity analysis. Because scope-2 dominates the total, even a moderate error in these assumptions changes the headline more than any other modeling choice, including the minor terminology issue around asthma deaths and the partly tautological HI-GLB evaluation. These are exactly the kinds of uncertainties that a conditional verdict should require the authors to address, but they do not warrant rejection because the direction and rough magnitude of the effect are supported by the underlying EPA model and the paper's own scenario logic. Running the suggested bounding analysis would convert the concern into either a quantitative sensitivity range or a confirmation that the headline is robust.","tokens_in":30348,"tokens_out":5487,"duration_ms":62993,"concrete_test":"Recompute the 2028 high-growth case in Section 5.1.3 under two bounding distributions instead of the fixed EPRI 2023 shares: (i) place all incremental load in the five states with the highest COBRA health cost per MWh, and (ii) place it in the five lowest. For each case, recompute x% per AVERT region and rerun COBRA. If the national total or the top per-household counties move by more than roughly 20% from the paper's $20.9B, the static 2023 distribution is load-bearing and the headline needs a sensitivity range.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The dominant scope-2 component, which accounts for about 92% of the 2028 high-growth estimate, is computed in Appendix A.2 by taking e_DC/e_total in each AVERT region and scaling every power plant's baseline emissions down by that percentage before applying COBRA. This is an average-attribution accounting choice, not an internal error, but the headline numbers inherit two weakly tested premises. First, the geographic driver is the EPRI 2023 state-level data center load distribution, applied to every year from 2019 through 2028 and scaled only by LBNL national totals. The 2028 high scenario adds roughly 400 TWh beyond the 2023 total, so the location of that incremental load, not the historical footprint, determines the regional x% values and thus the health costs. If new capacity concentrates in gas-heavy or coal-adjacent states rather than the 2023 footprint, the $20.9B figure and the 200x per-household disparity can shift materially. Second, average attribution assigns data centers a pro-rata share of every plant in a region, rather than the marginal plants that actually serve the added load. The paper itself cites delayed coal retirements and new gas plants in Section 3.2, which suggests upward bias if marginal generation is dirtier than the fleet average, but PPAs and dedicated renewables could push the opposite way; the paper provides no bounding analysis. The reader's weakest-assumption diagnosis is correct: this is the load-bearing uncertainty, and it is not tested against plausible alternative siting or marginal-dispatch scenarios.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper quantifies the public health burden of criteria air pollutant emissions from U.S. data centers by decomposing emissions into scope 1 (onsite backup generators), scope 2 (electricity generation), and scope 3 (supply chain), and applying the EPA's COBRA reduced-complexity model to estimate health outcomes and monetized costs. Using LBNL/EPRI electricity projections, the authors estimate a total annual public health cost of $6.7 billion in 2023, rising to $11.7–$20.9 billion in 2028 under low/high growth scenarios, with scope 2 contributing roughly 92% of the high-growth total. They further report large county-level disparities, per-household burdens up to roughly 200 times higher in some low-income counties, location-dependent costs for training a Llama-3.1-scale model, and a health-informed geographical load balancing (HI-GLB) framework that achieves a claimed ~26% health cost reduction relative to a carbon-aware baseline.","tokens_in":30577,"tokens_out":4290,"duration_ms":47342,"significance":"If the central estimates are accepted, this is a timely and policy-relevant quantification of an overlooked externality of AI infrastructure. The paper's strengths include the use of EPA's validated COBRA tool, transparent reporting of low/high ranges, explicit acknowledgement of excluded components (cooling towers, scope 3), and a clear methodological pipeline from emissions to health costs. The proposed health-informed computing framework is a useful conceptual contribution, and the comparison with carbon-aware scheduling underscores an important distinction. The significance is tempered, however, by the fact that the headline national and county-level numbers rest on attribution assumptions that are plausible but not yet stress-tested.","major_comments":[{"comment":"The dominant scope-2 estimate is computed by scaling every power plant's baseline emissions in each AVERT region by x% = e_DC/e_total and applying COBRA linearly. This is an average-attribution accounting choice, but the resulting headline figures inherit two weakly tested premises: (a) the EPRI 2023 state-level data center load distribution, scaled to LBNL national totals, is fixed for all years 2019–2028; and (b) data centers are served by the regional average plant mix rather than the marginal plants that actually respond to new load. For the 2028 high-growth scenario, roughly 400 TWh of incremental demand is added beyond 2023, so the location and marginal generation mix of that incremental load, not the historical footprint, largely determine the regional x% values and thus the $20.9 billion total and the county-level map. The paper itself cites delayed coal retirements and new gas plants in Section 3.2 as consequences of data center growth, which suggests that marginal generation may be dirtier than the fleet average, while PPAs and dedicated renewables could push in the opposite direction. No sensitivity analysis or bounding scenario is provided for either premise. Because scope 2 accounts for about 92% of the 2028 high-growth estimate, this is a load-bearing uncertainty that needs to be addressed before the headline claim can be considered robust.","section":"Appendix A.2, Section 5.1.3"},{"comment":"The HI-GLB evaluation uses the same health price metric as the optimization objective, so the reported ~26% reduction in health cost is partly a property of the objective rather than an independent validation of the framework. The health price from WattTime is a marginal, mortality-only signal, while the baseline national estimates use COBRA's average attribution with multiple health endpoints. To support the claim that health-informed scheduling 'can effectively mitigate health impacts,' the authors should validate the optimized schedules against an independent health metric (for example, COBRA-based average-attribution health costs computed from the resulting load distribution, or a different exposure-response function), or explicitly reframe the result as an in-sample optimized bound. As written, the comparison to carbon-aware GLB conflates objective-aligned cost reduction with demonstrated real-world health improvement.","section":"Section 6.2, Table 4"},{"comment":"The national scope-1 estimate extrapolates the emission rate (tons/MWh) derived from Virginia's backup generator permits to all other states using data center electricity consumption. Generator emissions depend on permit emission limits, generator Tier composition, operating hours, and demand-response activation patterns, all of which vary substantially by state and utility. Virginia's unusually large and concentrated generator fleet may not be representative. The paper notes that scope-2 costs dominate, but scope-1 still contributes about $1.6 billion in the 2028 high scenario (Table 1), which is not negligible. A sensitivity analysis using alternative utilization assumptions or per-state generator data, or at least a clear statement of the resulting uncertainty, would strengthen the estimate.","section":"Appendix A.1, Section 5.1"}],"minor_comments":[{"comment":"The abstract and introduction emphasize 'lifecycle pollutant emissions' and scope 3, but the main empirical analysis excludes scope-3 impacts except for one illustrative facility in Appendix A.3. This should be stated more prominently in the abstract to avoid overstatement of the analyzed scope.","section":"Abstract and Section 1"},{"comment":"The row labeled '2019 to 2023' reports cumulative five-year totals but appears under the 'Year' column alongside annual rows. It should be explicitly labeled as a cumulative period, e.g., '2019–2023 (cumulative)', to avoid confusion.","section":"Table 1"},{"comment":"The legend entries 'Data Center (Low)' and 'Data Center (High)' refer to LBNL growth scenarios, while the paper also uses 'low' and 'high' for COBRA's health-cost bounds. The figure and Table 1 would benefit from a consistent naming convention, such as 'growth-low' and 'growth-high' for scenarios and 'estimate-low/estimate-high' for COBRA bounds.","section":"Figure 4"},{"comment":"The claim that health costs exhibit greater temporal variation than carbon intensity in 110 out of 114 regions is based on WattTime's marginal health price, which considers only PM2.5 mortality. This important limitation is relegated to a footnote; it should be stated in the main text where the the comparison is made.","section":"Section 6.1, footnote 7"},{"comment":"The text says 'when estimating the electricity cost for data centers in 2023 and 2038,' but the paper analyzes 2028. This appears to be a typo.","section":"Appendix A, electricity price paragraph"},{"comment":"Minor typo: 'wile AI and data centers offer many societal benefits' should read 'while AI and data centers offer many societal benefits.'","section":"Introduction, last paragraph of Section 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses a genuinely important and timely topic, and the use of EPA's COBRA tool with clearly stated low/high ranges is a strength. However, the headline $20.9 billion figure and the county-level equity results are dominated by a single attribution assumption—average regional plant mix with a frozen 2023 geographic load distribution—that is not tested against alternatives. This is exactly the kind of load-bearing uncertainty that should be resolved or bounded before publication. The HI-GLB evaluation also needs an independent validation metric. I therefore recommend major revision rather than rejection, because the issues are fixable within the manuscript's scope and the core methodology is sound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, this is the first paper I know of that runs EPA COBRA on U.S. data center electricity and backup-generator emissions and maps the health burden to counties. The $6.7B (2023) and $11.7–20.9B (2028) numbers are screening-grade, but they come with low/high ranges and a clear enumeration of exclusions (scope 3, cooling towers). Second, the marginal-health-price vs. marginal-carbon analysis is the real contribution: across 114 regions, normalized IQR of health cost exceeds carbon in 96% of regions, yearly average correlation is only 0.292, and HI-GLB cuts health cost ~26% while carbon-aware GLB can increase it. That decoupling is robust and policy-relevant.\n\nWhere it is soft. The scope-2 attribution in Appendix A.2 scales every plant in an AVERT region by e_DC/e_total and applies COBRA linearly. That is average accounting, defensible, but the 2028 projection applies the 2023 EPRI state-level load distribution to the incremental ~400 TWh. The location of new capacity, not the historical footprint, drives the 2028 x% values. The paper never tests alternative siting or a marginal-dispatch scenario; the stress-test note is right that this is the load-bearing uncertainty. It could be addressed with a simple sensitivity analysis, so it is not fatal.\n\nThe '1/3 of asthma deaths' sentence in Section 5 compares 600,000 asthma symptom cases to asthma deaths—units don't match. That should be fixed.\n\nThe HI-GLB section is honest but the 26% reduction is partly a property of optimizing the same metric you evaluate. The non-tautological part is the contrast with carbon-aware GLB, which is the actual claim, and that stands. No code or data is shipped, which limits reproducibility; for a screening analysis built on COBRA inputs, releasing the COBRA runs and the EPRI/WattTime derived arrays would be easy and would materially raise the value.\n\nCitation pattern looked fair—EPRI, LBNL, WattTime, EPA, and prior GLB work are all credited. I don't see self-citation abuse.\n\nBottom line: this deserves a serious referee. It is a screening estimate, not a definitive accounting, and the authors mostly say so. The right outcome is peer review with a required sensitivity analysis on siting and marginal attribution, not desk rejection.","headline":"First credible COBRA-based screen of U.S. data center health burden; headline $20.9B rests on an untested average-attribution assumption, but the core environmental-justice and carbon-health decoupling results hold up.","tokens_in":31184,"tokens_out":1566,"would_cite":true,"duration_ms":16390,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"U.S. data centers could impose $20.9 billion in public health costs by 2028.","keywords":["data centers","public health","criteria air pollutants","AI sustainability","COBRA model","health-informed computing","geographical load balancing","environmental justice"],"falsifier":"One decisive test is a bottom-up marginal-accounting comparison: take the same 2023 and 2028 electricity totals for U.S. data centers, attribute them to the actual marginal plants using plant-level dispatch data, and run COBRA on those emission changes; if the resulting national annual health cost differs from $6.7 billion or $20.9 billion by more than about 30%, the average-attribution method fails. A second observable check is whether region-level power plant emissions actually track data center load hour-by-hour; if they do not, the $x\\% = e_{DC}/e_{\\mathrm{total}}$ scaling assumption is unsupported.","tokens_in":30094,"feed_emoji":"🌫️","tokens_out":7721,"duration_ms":70813,"temperature":0.7,"pith_summary":"This paper aims to quantify the hidden public health toll of U.S. data centers, which it frames as an overlooked consequence of AI's growth. Using EPA's COBRA model and LBNL electricity projections, it estimates that annual data-center-related health costs will rise from $6.7 billion in 2023 to $11.7–$20.9 billion in 2028, rivaling California's on-road vehicle emissions. The authors further claim that these costs fall unevenly, with per-household burdens in the most affected counties roughly 200 times higher than in the least affected, and that the worst-hit places are often not where data centers sit. They then propose health-informed computing, a load-scheduling framework that treats health cost per megawatt-hour as an objective and demonstrates a ~26% health-cost reduction without sacrificing electricity or carbon savings. The stakes are practical: if the estimates hold, data center growth is creating a multi-billion-dollar public health externality that current carbon-centric sustainability metrics do not capture.","feed_headline":"Data centers could cause $20.9B in U.S. health costs by 2028","feed_subtitle":"New estimate rivals California's on-road pollution toll and hits low-income counties hardest.","key_machinery":"The argument rests on two tools. COBRA, the EPA's reduced-complexity Co-Benefits Risk Assessment model, couples the PCAPS air-dispersion source-receptor matrix with concentration-response functions, converting pollutant changes into county-level counts of deaths, asthma symptoms, and dollar damages; the paper feeds it with emissions scaled from baseline data. The second is the AVERT regionalization: the U.S. grid is split into 14 regions, and the paper assigns data center electricity to each region from EPRI state-level data, then scales every plant's baseline emissions down by the ratio $x\\% = e_{DC}/e_{\\mathrm{total}}$ in that region, exploiting COBRA's near-linearity. For the scheduling part, hourly marginal health price per megawatt-hour (from the data source cited as [77]) becomes the objective coefficient in a linear program that shifts workloads across 13 data center locations over 8,760 hours.","core_discovery":"The central claim is that data centers, through their backup generators (scope 1) and the power plants that supply them (scope 2), will cause approximately 1,300 premature deaths and 600,000 asthma symptom cases per year in the contiguous U.S. by 2028, with a public health cost of $11.7–$20.9 billion depending on AI growth. The paper asserts that this burden is location-dependent and largely decoupled from carbon emissions: the spatial correlation between average health price and carbon intensity across 114 U.S. regions is only 0.292, and the per-household county-level ratio between the highest and lowest affected communities is about 200. It further shows that training a single Llama-3.1-scale model can produce air pollution equivalent to more than 10,000 LA-NYC car round trips and a health cost exceeding 120% of the training electricity cost. Finally, it claims that a health-informed geographical load balancing formulation, using marginal health price signals, cuts health cost by 25.8% relative to the 2023 baseline while simultaneously reducing electricity cost by 3.0% and carbon by 1.4%, and that carbon-aware load balancing can actually increase health costs.","pith_inferences":["If the appendix's scope-3 semiconductor facility health costs are scaled to AI chip demand, the paper's main totals would likely rise beyond $20.9 billion, since the single example facility contributes $26–39 million per year.","The average-attribution assumption may understate the marginal burden: if data centers are served by marginal gas or coal plants, the projected $20.9 billion is more likely a lower bound, and the county-level distribution would shift toward coal-belt states.","Because the spatial correlation between health price and carbon intensity is only 0.292, a health metric should be reported alongside carbon in sustainability disclosures; carbon-only regulation would miss the communities that carry the highest health burden.","One could validate the COBRA-based figures by running a full photochemical grid model on a few representative power-plant configurations; large discrepancies in downwind populated counties would indicate the reduced-complexity model needs recalibration for this application."],"forward_implications":["U.S. data center health costs are projected to roughly triple from $6.7 billion in 2023 to $11.7–$20.9 billion in 2028, making the industry's health burden comparable to California's on-road vehicle emissions.","The county-level per-household burden varies by about 200-fold, and every one of the top-10 most affected counties has a median income below the national median.","Training one Llama-3.1-scale model (about 30 GWh) can generate health costs of $0.23–$2.5 million depending on site, exceeding 120% of the training electricity bill in some locations.","A health-informed geographical load balancing algorithm cuts health cost by ~26% under modest slack ($\\lambda = 1.5$) while also lowering energy cost and carbon, whereas carbon-aware scheduling alone can increase health cost.","Virginia backup generators at 10% of permitted emissions already cause roughly 14,000 asthma symptom cases and $220–$300 million in annual health costs across the region."],"supporting_citations":[{"why":"Supplies the historical and 2028 U.S. data center electricity consumption totals and growth scenarios that drive the $20.9 billion projection.","marker":"[4]"},{"why":"COBRA tool provides the source-receptor dispersion and health impact valuation used for all county-level cost estimates.","marker":"[39]"},{"why":"COBRA user's manual defines the concentration-response functions and health endpoints, and documents PCAPS validation.","marker":"[22]"},{"why":"EPRI state-level data center electricity distribution is scaled and used to assign loads to AVERT regions.","marker":"[5]"},{"why":"AVERT defines the 14 electricity regions and the state-to-region apportionment used for the x% emission scaling.","marker":"[76]"},{"why":"Marginal health price and carbon intensity data across 114 regions support the health-informed GLB case study and correlation analysis.","marker":"[77]"},{"why":"Virginia air permit data gives backup generator capacity and permitted emission limits for scope-1 estimates.","marker":"[60]"},{"why":"Virginia JLARC report documents that actual backup generator emissions are about 7% of permitted levels, anchoring the 10% assumption.","marker":"[42]"},{"why":"Meta's Llama-3.1 model card provides the 39.3 million GPU-hours used to estimate training electricity (30 GWh).","marker":"[92]"},{"why":"PCAPS validation against photochemical grid models supports COBRA's dispersion accuracy for power and on-road sectors.","marker":"[83]"}],"fun_headline_variants":["Data centers to cause $20B in US health costs by 2028","Health-aware scheduling cuts data center health costs 26%","AI data center pollution: $20B health bill and 1,300 deaths","Data center health costs hit $20B, low-income areas hardest"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The biggest assumption is that data center electricity is drawn from the average mix of all power plants in each region, so the paper scales every plant's emissions down by the data center's share of regional electricity; if data centers actually pull from the dirtiest plants, the $20.9 billion figure and its county distribution would change.","fun_headline_variants_meta":{"raw":{"variants":["Data centers to cause $20B in US health costs by 2028","Health-aware scheduling cuts data center health costs 26%","AI data center pollution: $20B health bill and 1,300 deaths","Data center health costs hit $20B, low-income areas hardest"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000345,"raw_usage":{"total_tokens":1904,"prompt_tokens":969,"completion_tokens":935,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":857}},"tokens_in":585,"tokens_out":935,"duration_ms":8513,"temperature":1.0,"reasoning_tokens":857,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:49:13.775556+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One decisive test is a bottom-up marginal-accounting comparison: take the same 2023 and 2028 electricity totals for U.S. data centers, attribute them to the actual marginal plants using plant-level dispatch data, and run COBRA on those emission changes; if the resulting national annual health cost differs from $6.7 billion or $20.9 billion by more than about 30%, the average-attribution method fails. A second observable check is whether region-level power plant emissions actually track data center load hour-by-hour; if they do not, the $x\\% = e_{DC}/e_{\\mathrm{total}}$ scaling assumption is unsupported.","supporting_citations":[{"cited_title":"Smith, Alex Hubbard, Alex Newkirk, Nuoa Lei, Md Abu Bakar Siddik, Billie Holecek, Jonathan Koomey, Eric Masanet, and Dale Sartor","cited_arxiv_id":null,"evidence_quote":"Supplies the historical and 2028 U.S. data center electricity consumption totals and growth scenarios that drive the $20.9 billion projection."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"COBRA tool provides the source-receptor dispersion and health impact valuation used for all county-level cost estimates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"EPRI state-level data center electricity distribution is scaled and used to assign loads to AVERT regions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"AVERT defines the 14 electricity regions and the state-to-region apportionment used for the x% emission scaling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Marginal health price and carbon intensity data across 114 regions support the health-informed GLB case study and correlation analysis."},{"cited_title":"Introducing Llama 3.1: Our most capable models to date.https://ai.meta.com/blog/ meta-llama-3-1/","cited_arxiv_id":null,"evidence_quote":"Meta's Llama-3.1 model card provides the 39.3 million GPU-hours used to estimate training electricity (30 GWh)."}],"review_version":2}