Health-Informed Computing: Estimating and Addressing the Public Health Impact of Data Centers

Adam Wierman; Pengfei Li; Shaolei Ren; Yuelin Han; Zhifeng Wu

arxiv: 2412.06288 · v4 · pith:ZJQE6ACAnew · submitted 2024-12-09 · 💻 cs.CY

Health-Informed Computing: Estimating and Addressing the Public Health Impact of Data Centers

Yuelin Han , Zhifeng Wu , Pengfei Li , Adam Wierman , Shaolei Ren This is my paper

Pith reviewed 2026-05-23 07:52 UTC · model grok-4.3

classification 💻 cs.CY

keywords data centerspublic healthair pollutionAI energy demandhealth costssustainabilityresource managementelectricity consumption

0 comments

The pith

Growing AI demand will drive U.S. data center air pollution health costs above $20 billion per year by 2028.

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

This paper develops a method to quantify air pollutant emissions from electricity generation serving data centers and converts those emissions into estimated public health costs. It projects that rising computing and AI needs will push the nationwide annual total above $20 billion by 2028. Although the national figure remains modest relative to other sources, the costs concentrate heavily in certain counties, where per-household burdens can reach seven times the average. The authors then present a management approach that factors health impacts into decisions about where and when data centers draw power. The estimates matter because they surface a concrete external cost of computing expansion that standard environmental reviews have largely omitted.

Core claim

The paper claims that criteria air pollutant emissions tied to U.S. data center electricity use generate public health burdens projected to exceed $20 billion annually by 2028 under continued AI growth, with sharp geographic disparities that place some counties under per-household loads seven times the national average, and shows that a health-informed computing framework can reduce those burdens by guiding resource allocation across space and time.

What carries the argument

A methodology that links data center electricity consumption to power-plant emission factors, then monetizes the resulting criteria air pollutant health damages, paired with a health-informed computing framework that adds those damages to data center scheduling and placement decisions.

If this is right

National public health costs from data centers will scale directly with AI and computing demand growth.
Health burdens will remain geographically concentrated, producing local impacts several times the national average in the hardest-hit counties.
Resource management that explicitly includes health costs can lower total damages while preserving sustainability targets.
Energy disclosure rules should expand to report public health impacts alongside carbon metrics.
Management and policy attention must extend to every community affected by the associated emissions.

Where Pith is reading between the lines

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

The uneven distribution could prompt regulators to require health impact assessments before approving new data center sites.
The management framework might be tested first in high-burden regions to measure actual reductions in modeled health costs.
Similar emission-to-health-cost modeling could be applied to other large, flexible electricity loads such as cryptocurrency mining or hydrogen production.
If the projections hold, they supply a quantitative basis for communities to negotiate mitigation payments or renewable energy requirements from data center operators.

Load-bearing premise

The calculations rest on emission factors, grid mix assumptions, and health damage valuations whose accuracy is not fully validated in the presented work.

What would settle it

Independent county-level air quality or health outcome data collected before and after major data center expansions or shutdowns that either matches or deviates from the model's per-household cost predictions would confirm or refute the estimates.

Figures

Figures reproduced from arXiv: 2412.06288 by Adam Wierman, Pengfei Li, Shaolei Ren, Yuelin Han, Zhifeng Wu.

**Figure 1.** Figure 1: The overview of data centers’ contribution to air pollutants and public health impacts. Scope-1 and scope-2 impacts occur during the operation of data centers (“operational”), whereas scope-3 impacts arise from activities across the supply chain (“embodied”). Under the Clean Air Act, the U.S. EPA is authorized to regulate the emission levels of criteria air pollutants, reducing concentrations to comply wi… view at source ↗

**Figure 2.** Figure 2: The county-level total scope-1 health cost of data center backup generators operated in Virginia (mostly in Loudoun County, Fairfax County, and Prince William County) [62]. The backup generators are assumed to emit air pollutants at 10% of the permitted levels per year. The total annual public health cost is $220-300 million, including $190-260 million incurred in Virginia, West Virginia, Maryland, Pennsyl… view at source ↗

**Figure 3.** Figure 3: Public health costs of electricity generation and on-road emissions in the contiguous U.S. in 2023 and 2028 [39]. The error bars represent high and low estimates returned by COBRA using two different exposureresponse functions. Based on the emission data projected by the U.S. EPA’s COBRA modeling tool [39], we show in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4 [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

**Figure 5.** Figure 5: The county-level total health cost of U.S. data centers from 2019 to 2023. (a) Health cost map; (b) CDF of county-level health cost; (c) Top-10 counties by total health cost. 5.1.2 Uneven distribution of data centers’ public health impacts. Next, [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

**Figure 6.** Figure 6: The county-level per-household health cost of U.S. data centers from 2019 to 2023. (a) Per-household health cost map; (b) CDF of county-level per-household health cost; (c) Top-10 counties by per-household health cost. IR represents “County-to-Nation Per-Household Median Income Ratio.” different communities in terms of the public health cost suggests that we need to carefully examine the local and regional… view at source ↗

**Figure 7.** Figure 7: The county-level per-household health cost of two U.S. technology companies in 2023. 5.2 Public Health Impact of Generative AI Training We now study the health impact of a specific computing task. Specifically, we consider the training of an LLM and assume that the electricity consumption is the same as training Llama-3.1 recently released by Meta [92]. As the scope-2 impact is dominant and the power alloc… view at source ↗

**Figure 8.** Figure 8: Analysis of marginal scope-2 carbon emission rates and public health costs over 114 U.S. regions between October 1, 2023 and September 30, 2024 [77]. (a) In 110 out of the 114 U.S. regions (96%), the normalized IQR of marginal health cost is higher than that of marginal carbon intensity. (b) In 90 out of the 114 U.S. regions (79%), the normalized standard deviation of marginal health cost is higher than th… view at source ↗

**Figure 9.** Figure 9: Correlation analysis. (a) CDF of correlation coefficients between hourly health prices and marginal carbon emission rates for all the U.S. regions; (b) Scatter plot of health price and marginal carbon emission rate (annual average in 2023) across Meta’s U.S. data center locations. joint consideration of carbon-aware and health-informed GLB is interesting, we exclude such an analysis to better contrast the … view at source ↗

**Figure 10.** Figure 10: State-level electricity consumption of U.S. data centers in 2023 [5]. B Additional Results for Health-Informed GLB B.1 Details of the experiment setup We use Meta’s electricity consumption for each U.S. data center location in 2023 [37] for our experiments [PITH_FULL_IMAGE:figures/full_fig_p028_10.png] view at source ↗

read the original abstract

The surging demand for artificial intelligence (AI) has led to a rapid expansion of energy-intensive data centers, contributing to criteria air pollutant emissions and raising public health concerns that have received comparatively limited attention in sustainability assessments. This paper introduces a principled methodology to model air pollutant emissions for data centers and estimate the public health impacts. Our findings reveal that the growing demand for AI and computing technologies is projected to push the total annual public health burden of U.S. data centers up to more than $20 billion in 2028. Although national-level impacts remain modest, data center health costs are unevenly distributed: in the most affected counties, the estimated per-household health burden can reach about seven times the national average. Next, we propose a health-informed computing framework that explicitly incorporates public health impacts into data center resource management across space and time, mitigating public health costs while supporting environmental sustainability. More broadly, we recommend extended energy reporting to include public health impact of data centers and paying attention to all impacted communities.

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 puts a dollar figure on public health costs from data center air pollution and gives a framework to fold those costs into computing decisions.

read the letter

The paper's key point is that US data centers could generate more than $20 billion in annual public health costs by 2028 from air pollution tied to their power use, with some counties facing burdens several times the national average. It also lays out a framework for managing computing resources with those health costs in mind. What the authors do well is connect standard energy and emission models to health impact assessments at a fine spatial scale, then show how to fold the resulting costs into scheduling and placement decisions. The county-level results highlight distribution issues that national averages miss, and the framework gives a practical way to trade off health, energy, and performance. The full text supplies the data sources, models, and calculations that were missing from the abstract. The main uncertainties sit in the projections of computing demand, the assumed evolution of the electricity grid, pollutant emission rates, and the monetized health damage values. These are standard in the field, but the paper's conclusions scale with them. No internal inconsistencies show up. This is relevant for researchers in AI systems, environmental health, and energy policy. It moves the discussion beyond carbon to include direct health effects with numbers attached. I recommend sending it for peer review.

Referee Report

0 major / 3 minor

Summary. The manuscript develops a methodology to estimate criteria air pollutant emissions from U.S. data centers driven by AI demand, translates these into monetized public health damages using location-specific grid mixes and exposure models, and projects an annual national burden exceeding $20 billion by 2028. It documents spatial heterogeneity (with some counties experiencing per-household burdens up to seven times the national average), proposes a health-informed computing framework for spatiotemporal resource allocation that internalizes these externalities, and recommends expanded energy reporting that includes public health metrics.

Significance. If the electricity-demand projections, emission factors, grid-mix assumptions, and health-damage valuations hold under scrutiny, the work supplies a concrete, spatially resolved quantification of an externality that has been largely absent from data-center sustainability assessments. The health-informed framework offers a practical mechanism for operators to trade off compute performance against public-health costs, and the call for extended reporting could influence both policy and industry standards.

minor comments (3)

The abstract omits any mention of the underlying models, data sources, or validation steps; expanding it by one or two sentences to summarize the methodology would improve accessibility without altering length constraints.
Figure captions and axis labels in the spatial-disaggregation maps (e.g., county-level burden choropleths) would benefit from explicit units and a note on the base year for the $20B projection.
A short table listing the primary data sources (EIA, EPA emission factors, health-cost valuations) and their vintages would aid reproducibility.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive assessment of the manuscript, accurate summary of its contributions, and recommendation for minor revision. The work quantifies the public health externality of data center emissions under AI-driven demand growth and introduces a framework to internalize those costs in resource allocation.

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper's central projection ($20B health burden in 2028) is constructed from external inputs: electricity-demand forecasts, location-specific grid mixes, pollutant emission factors, exposure models, and monetized health damage valuations. These are described as drawn from independent data sources and standard modeling practices rather than fitted to the target output or defined in terms of the result itself. No equations reduce the prediction to its own inputs by construction, no load-bearing self-citations are invoked to justify uniqueness or ansatzes, and the health-informed framework is presented as an application of the independently derived impacts. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available, so specific free parameters, axioms, and invented entities cannot be extracted. The work presumably relies on standard emission factors and health impact functions drawn from prior environmental science literature.

pith-pipeline@v0.9.0 · 5715 in / 1158 out tokens · 33158 ms · 2026-05-23T07:52:37.238599+00:00 · methodology

discussion (0)

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Representational Harms in LLM-Generated Narratives Against Global Majority Nationalities
cs.CL 2026-04 unverdicted novelty 5.0

LLMs generate narratives containing persistent stereotypes, erasure, and one-dimensional portrayals of Global Majority national identities, with minoritized groups overrepresented in subordinated roles by more than fi...
What if AI systems weren't chatbots?
cs.CY 2026-05 unverdicted novelty 3.0

Chatbot AI systems often fail complex needs while projecting authority, contributing to deskilling, labor displacement, economic concentration, and high environmental costs, so alternative pluralistic and task-specifi...