Center for Practical AI
AI and the Environment · Guide 4 of 5Questions 4 & 5

Somebody breathes it. Somebody pays for it. Usually the same somebody.

The health and cost burden of the AI buildout doesn’t spread evenly — and the algorithms designed to make the system more efficient overall appear to make the distribution worse.

12 min read · Answers the fourth and fifth of the seven questions

Question 4

What the power plants do to the air.

The least discussed of the seven questions, and among the better substantiated. Every figure here is scenario-dependent, and this page states the scenario every time it states a number.

$11.7–20.9B

modeled US public health cost of data centers in 2028 — low-growth to high-growth scenario, against a 2023 baseline of $6.67B

Han, Wu, Li, Wierman & Ren (2026), modeled

Airshed
~12:1

how much grid electricity (Scope 2) outweighs backup diesel generators (Scope 1) in that burden

Han, Wu, Li, Wierman & Ren (2026)

Airshed
~7×

per-household burden in the worst-affected counties against the national average

Han, Wu, Li, Wierman & Ren (2026)

County / parcel
Pending

whether North Carolina's large-load customers pay the full cost of the generation they require — an open question with a September 2026 decision date

NCUC Docket E-100 Sub 208

Utility territory

In the high-growth 2028 scenario the same modeling estimates roughly 1,300 premature deaths and 600,000 asthma symptom cases. The $20.9B figure is the high case; the low case is $11.7B. If you see the larger number quoted without its scenario, you are watching a figure begin to drift — which is how nearly every misused number in this literature started.

Against the coverage

It isn't the diesel generators.

Local reporting on data centers is dominated by backup generator permits. The research points somewhere else, and the difference changes what a community should actually ask for.

In the health burden from data centers, emissions from grid electricity generation (Scope 2) outweigh emissions from onsite backup diesel (Scope 1) by roughly 12 to 1.

Cleaner electricity matters far more than cleaner generators.

That is the paper’s own conclusion, and it cuts directly against the framing that dominates local coverage. Backup generators are visible, permitted locally, and easy to photograph. They are also the smaller share of the problem by an order of magnitude. A community that wins a hard fight over generator emissions standards and never asks what generation is being built to serve the facility has won the wrong argument.

Distribution

The burden is concentrated.

Per-household health burden in the worst-affected counties runs about seven times the national average. Averages are the wrong instrument here: a national per-household figure describes almost nobody, because the whole phenomenon is that some counties carry far more than their share.

A correction, while we are here. A 200× figure circulates for this comparison. It came from a highest-versus-lowest county comparison — a very different quantity — and it has been superseded. The defensible figure is the roughly 7× ratio against the national average, and it is quite bad enough to make the point.

Health, at this scale, is an Airshed question: downwind, regional, and driven by generation rather than by the building. The cost question is a Utility territoryquestion: a defined, nameable list of people on one utility’s bill. Neither is global. Neither is answerable with a national average. And they fall on the same people for the same reason — both are downstream of a siting decision those people did not make.

The counterintuitive one

Efficiency is not fairness.

The most surprising finding in this series, and the one with the longest reach beyond data centers.

Algorithms that route workloads between data centers to minimize aggregate energy cost, carbon, or water actively amplify regional inequity compared with simply routing to the nearest facility. The mechanism is not mysterious: an optimizer exploits whichever region is cheapest, aggressively — for instance shifting load toward Texas despite worse carbon intensity and worse water-use effectiveness — and cheap frequently means a place with less capacity to absorb the burden.

Read this one with its caveat attached. The finding comes from a trace-based simulation across ten hypothetical facilities, not from measurement of a real deployment. It is a strong result about how these objective functions behave, not an audit of what any operator is currently doing.

Efficiency and equity are not the same objective, and a system optimized for one will trade away the other unless someone writes the other into the objective function.

The researchers’ proposed fix is exactly that: add a fairness term, and most of the effect goes away. Which is the encouraging half of the finding, and the half worth carrying to every other domain where something is being optimized on your behalf.

Is routing workloads to the cheapest region better?

The version that goes too far

Treats every optimization decision as deliberate exploitation of poorer regions, as though the objective function were chosen to do harm rather than chosen without anyone asking this question.

The version that waves it away

Treats aggregate efficiency as self-evidently good — less energy, less cost, less water in total, so what is the complaint? It is a real gain, and totals say nothing about who absorbs what.

What the evidence supports

Optimizing for aggregate cost, carbon, or water measurably amplifies regional inequity versus routing to the nearest facility, because the optimizer exploits whichever region is cheapest — and cheap often means less able to absorb the burden. Adding a fairness constraint largely fixes it. Shown on simulated facilities, not measured deployments.

Grid region

Sources for this split: equitableAI — full citations below.

Question 5

The bill: real, contested, and being decided now.

Across PJM’s territory, the 2025/2026 capacity auction cleared at $269.92 per megawatt-day for much of the region, with data-center load growth cited as one contributor. Average US residential retail electricity prices rose about 6% in nominal terms in 2024, more than twice inflation, after roughly tracking inflation from 2019 to 2024. Both figures come from secondary reporting that we have not verified against the auction results or the underlying federal price series.

And here is the counterweight, which deserves to be quoted by its title rather than paraphrased, because the title is the whole argument:

“Are Data Centers Raising Your Electric Bill? Mostly Not. Yet.”

New Jersey State Policy Lab, Rutgers University

Attribution of specific retail increases to data centers is genuinely contested, and this analysis says so from a position sympathetic to the concern. Retail rates move on fuel costs, storm recovery, transmission investment, and rate-case outcomes all at once, and picking one cause out of that is harder than most coverage admits. Note what the last word is doing, though. “Yet” is a statement about timing, not about whether.

Data centers are why my electric bill went up.

The version that goes too far

Attributes a broad retail increase to a single cause, when rates move on fuel costs, storm recovery, transmission investment, and rate-case outcomes simultaneously. The New Jersey State Policy Lab's answer is “mostly not — yet.”

The version that waves it away

“There's no evidence data centers raise rates” ignores that the question is being actively litigated precisely because the cost allocation is unresolved — and treats “not yet” as though it meant “not ever.”

What the evidence supports

Whether large-load customers pay the full cost of the generation and interconnection they require is an open regulatory question with live dockets in multiple states. The answer determines your bill. It is being decided by people who accept public comment and are almost never asked.

Utility territory

Sources for this split: njPolicyLab · largeLoadDocket · largeLoadFilings — full citations below.

The venue

Where this is actually decided.

Not in the news, and not in the comment section. In utility commission dockets, in tariff design, and in tax-incentive statutes.

Large-load tariffs determine minimum bills, contract length, exit penalties, and — most consequentially — whether interconnection upgrades are assigned to the customer who caused them or socialized across residential ratepayers. That is the single largest lever on whether the buildout shows up on your bill, and it is set in proceedings that are open to the public, accept written comment, and are attended almost exclusively by parties with a financial interest.

It is the most consequential and least-watched venue in this entire debate.

North Carolina is where CPAI has followed this most closely, and it is specific enough to be worth reading as a case rather than a summary — a repealed carbon target, 37 unpublished tax determinations, and a large-load queue the state’s own Commissioner describes as 70% data centers.

What you can do

Action for every level of influence.

1

For yourself

  • Find out who your state utilities commission is and whether it has an open large-load or rate proceeding. In most states this is one search, the filings are public, and almost nobody outside the industry reads them.
  • Learn the difference between a capacity charge and an energy charge on your own bill. It is the vocabulary the entire cost fight is conducted in.
2

For a community

  • When a facility is proposed, ask three questions in writing: what generation is being built to serve it, who pays for the interconnection upgrades, and what the minimum-bill and exit terms are if the project changes.
  • Borrow North Carolina's Public Staff formulation, which is the clearest statement of the risk anyone has put on a record: "If the [Large Load Customer's] plans change, the [Large Load Customer] cannot simply walk away from the system and leave the remaining customer base with the bill."
  • Ask about generation policy, not just generator policy. Because grid electricity outweighs backup diesel by roughly 12 to 1 in the health burden, what serves the facility matters far more than what sits behind it.
3

For an organization

  • If you are a large electricity customer of any kind, large-load tariff design affects your rates too. Read the docket — you are an interested party whether or not you have filed as one.
  • If you procure compute, ask where workloads are routed and on what objective. "Cheapest region" is a defensible business answer and it is not a neutral one.
4

For policy

  • Separate rate classes for large load, so cost causation can be assigned rather than argued about after the fact.
  • Cost-causation-based assignment of interconnection upgrades, rather than socializing them across residential ratepayers by default.
  • The underappreciated one: because Scope 2 dominates Scope 1 by about 12 to 1, generation policy is health policy. Emissions standards on backup generators are the visible lever and the much smaller one.
  • Write distributional constraints into efficiency mandates. A system optimized only for aggregate efficiency will trade away equity unless someone puts equity in the objective.

Where this leads

Reading is one thing. Practicing it is another.

The Applied AI Certification builds practical AI fluency across all six domains — the working competence that advances toward proficiency, with structured practice, feedback, and a cohort on the same problems.

Sources

Research & further reading.

Peer-reviewed studyModeled estimate · not metered measurementAirshed
Han, Wu, Li, Wierman & Ren (2026)Health-Informed ComputingCommunications of the ACM. $20.9B in public health costs is the high-growth 2028 scenario; low-growth is $11.7B and the 2023 baseline is $6.67B — always state which. About 1,300 premature deaths and 600,000 asthma symptom cases in the 2028 high scenario. Scope 2 (generation) dominates Scope 1 (onsite backup diesel) by roughly 12 to 1. The worst-affected counties carry about 7× the national average per household.
Peer-reviewed studyTrace-based simulation · hypothetical facilitiesGrid region
Li, Yang, Wierman & Ren (2024)Towards Environmentally Equitable AI via Geographical Load BalancingACM e-Energy. Algorithms that optimize for aggregate cost, carbon, or water actively amplify inequity compared with simple nearest-datacenter routing — efficiency at the aggregate concentrates harm at the margin. Trace-based simulation across 10 hypothetical facilities, not a real deployment.
Independent policy analysisUtility territory
New Jersey State Policy Lab, Rutgers UniversityAre Data Centers Raising Your Electric Bill? Mostly Not. Yet.The honest counterweight on retail rates. Attribution of specific retail increases to data centers is contested, and this is the analysis that says so from a position sympathetic to the concern.
Regulatory filing / docketUtility territory
NCUC Docket E-100 Sub 208 and DEC rate case E-7 Sub 1329In the Matter of Large Electric Load Additions — the proceedingsThe docket facts, which are confirmable from the Commission's own record: opened June 6, 2025; technical conference October 2025; filing-requirements order March 2026; Commission decision due September 20, 2026; expedited large-load tariff filing due end of September 2026.
Journalism · secondary reportingUtility territory
Filings in E-100 Sub 208 / E-7 Sub 1329, via journalism and advocacy releasesIn the Matter of Large Electric Load Additions — the contentsDuke's large-load tariff terms (≥50 MW at ≥80% load factor, 10–15 year contracts, minimum bill at 75% of contract demand, 25% exit penalty, and no separate data center rate class); Public Staff's request that about $200M of Duke's $247M in requested grid upgrades be assigned to large-load customers; Attorney General Jackson's push for a separate data center rate class and a 7.4% return on equity against Duke's 10.95% ask; the July 17, 2026 settlement at 3.7% average annual increases over two years and 9.8% ROE, which Jackson declined to sign; and Commissioner Tommy Tucker's remark that Duke's large-load queue is 70% data centers.Citation still being verified against our research files.
Journalism · secondary reportingWatershed
Data center community impacts cluster (2026)Assorted secondary reporting on siting, consumption, and disclosureWhere the widely repeated "about two-thirds of data centers built or in development since 2022 are in water-stressed areas" figure circulates, along with roughly 17.4 billion gallons directly consumed in 2023 rising to 38–73 billion by 2028 (attributed to EPA), and the finding that California has no comprehensive disclosure requirement. All secondary. The vault's instruction is to treat this as a map of where to look, not as a citable base — which is why the siting argument on the water guide is built on aqueduct40, im3Projected, and computeAtlas instead.Citation still being verified against our research files.
Last reviewed: August 2026We review this page quarterly. Statistics in this category change rapidly.Health cost figures are modeled and scenario-dependent; the scenario is stated wherever a number appears. The geographical load balancing finding is a simulation on hypothetical facilities and is labeled as such in the body text, not only here. The PJM clearing price and the 2024 retail price rise are secondary reporting and are marked in the sentence. The North Carolina docket outcome was pending at the time of review.

Want CPAI to teach this in your community?

We deliver this material as workshops and sessions for schools, libraries, local government, and community organizations — including a version built for people preparing to comment in a utility proceeding.