The cost of not thinking. The bill will come as less breathable air, not euros

Delegating has a price too. Each outsourced decision spends energy on a server that physically lives somewhere —Ashburn, Virginia; Cedar Rapids, Iowa; Dublin, Ireland. The generic name "the cloud" hides a postal address, and the postal address has neighbors who breathe what the nearby power plants emit to feed the data center. The free service for the individual user is the appearance that holds up the model. The bill will come as less breathable air, not euros. And the problem with less breathable air is that no one signs a bill for it: it's paid by whoever breathes, not whoever delegated.

The physical geography

"The cloud" is a comfortable metaphor. The cloud isn't ethereal. It's a geographic concentration of data centers in specific, identifiable places.

• Loudoun County, Virginia. The region known as "Data Center Alley" concentrates, according to the figure put out by the county's own economic development agency (Loudoun County Economic Development), around 70% of the world's internet traffic —an estimate from the agency itself, methodologically debatable. The concentration is due to the combination of dense fiber, electricity supply, favorable tax regulation and historical proximity to the origin of the internet in the U.S.
• Central Iowa. Microsoft, Google and Meta operate large campuses in the state (Google in Council Bluffs, Meta in Altoona, Microsoft in West Des Moines), in zones more exurban than strictly rural, taking advantage of cheap land, available water and relatively accessible energy.
• Dublin, Ireland. The concentration in the metropolitan area has reached levels that strain the national electrical grid. EirGrid, the operator of the Irish electrical system, applied from 2021-2022 an effective moratorium on new data center connections in the Dublin area for insufficient capacity, lifted at the end of 2025 in exchange for requiring each new facility to have its own generation or storage able to cover all of its demand.
• Singapore. After reaching saturation, the government introduced in 2019 a moratorium on new data centers that was partially lifted in 2022 with mandatory efficiency criteria.

The geographic concentration has consequences for nearby communities. Loudoun County published in 2024 and 2025 reports from its Department of Economic Development acknowledging growing tensions between the data center sector and the residential community over land use, electricity consumption and noise. The specialized press has been documenting the citizen mobilization around the expansion.

The chain of cost

The chain between the user's click and the physical impact on concrete communities has several links worth unfolding.

• User and query. A person writes a question to the chatbot in any city in the world.
• Data center. The query is routed to the nearest data center with available capacity. For European users, frequently in Ireland, the Netherlands or Frankfurt. For American users, frequently in Virginia, Iowa, Oregon.
• Regional electrical grid. The data center consumes electricity from the local grid. The local grid must respond by generating or importing.
• Power plants. The additional generation, in an electrical mix partly dependent on fossil fuels (in the U.S., approximately 60% of the national mix in 2023 according to the EIA), produces emissions of CO2, nitrogen oxides, sulfur oxides and fine particulates (PM2.5).
• Local and global atmosphere. The emissions impact air quality locally (the power plant's neighbors) and the carbon balance globally (everyone).

The user sees the first link. The chain backwards is invisible to them. The invisibility isn't a design accident; it's a property of the business model that separates the origin of cost from the destination of cost.

The classic externality reformulated

What neoclassical economics described as an externality —Ronald Coase in The Problem of Social Cost (Journal of Law and Economics 3, 1960) and Arthur Pigou earlier in The Economics of Welfare (Macmillan, 1920)— operates again in an updated form.

The classic externality was industrial pollution: the factory produced a good the consumer bought, and simultaneously emitted waste that third parties suffered without compensation. The cost of the waste didn't appear in the price of the good. The consumer bought more of the good than they would have if the price reflected the full cost.

The digital externality has the same structure with two modifications.

• Immaterial product. The classic factory produced a physical good (cement, steel). The digital factory produces inference (text, decision, recommendation). The immateriality of the product makes it easier to conceal the real physical cost.
• Geographic distance between consumption and production. The classic factory tended to be relatively close to the consumer. The digital factory is in a rural county in Virginia the European consumer will never visit. The geographic distance reinforces the cognitive dissociation between the click and its consequences.

The underlying structure is identical: third parties pay part of the production cost not internalized in the price. The name of the product changes; the mechanism doesn't.

The strategic free service

The free service for the individual user isn't an accident. It's a strategic decision that produces two operational effects.

• Mass adoption. The free product is adopted at a scale the paid product wouldn't reach. Mass adoption allows economies of scale, training data, a dominant position.
• Internalizing the benefit without internalizing the cost. The user who receives the benefit without paying a price has no cognitive mechanism to assess whether the utility received offsets the aggregate cost. They decide to consume without information about the real cost because the cost isn't presented to them.

If ChatGPT's page showed, next to the send button, the estimated CO2 emissions figure of the query and the neighborhood of the data center where it will be processed, the aggregate behavior of the user would change. The exact figure isn't published; the choice not to publish it is a design decision that operates coherently with the mass-adoption strategy.

The legitimate objection to the above is that the figures would be hard to calculate precisely per individual query. The objection is partly true —attributing emissions to an individual query requires assumptions about marginal electrical mix, data center efficiency, the nature of the inference— and is no excuse for not providing reasonable default estimates. The physical-product industry lives with approximate footprint estimates; the digital industry could do the same if it decided to.

What emerging regulation is attempting

The U.S. EPA published in December 2025 Clean Air Act Resources for Data Centers, the first explicit regulatory hub on the sector. The content acknowledges that data centers, although they aren't classic industrial sources, contribute to air quality locally via the electricity generation they demand, and proposes an initial framework of coordination with state and local authorities.

A report by the think tank Next 10 co-signed by Shaolei Ren (UC Riverside), published in 2025, analyzed health impacts attributable to electricity generation for data centers in California. The report reported a tripling of the estimated health impacts (measured in external costs from attributable respiratory morbidity) between 2019 and 2023, with geographic concentration in lower-income communities near the power plants that cover the demand of Northern California's data centers.

European regulation goes another way. The revised Energy Efficiency Directive (EU 2023/1791) introduced from 2024 reporting obligations on consumption and efficiency for data centers above a certain size, with a first reporting date in September 2024 and publication of the aggregate national registries starting in 2025.

The two regulatory paths are distinct. The American one targets local health externalities; the European one targets efficiency and aggregate consumption. Neither obliges the producer to internalize the full cost. Neither reaches the individual user.

The political question

This is personal opinion but it documents what the literature on digital externalities and digital sustainability has been showing. Each free cognitive delegation is an externality paid by third parties. Acknowledging it doesn't switch off the chat or invalidate the product. It changes the perceived price and, eventually, the pattern of use.

The response options operate at different scales.

• Information to the user. Default labeling of the estimated footprint of each interaction, similar to the nutritional or energy labeling in other sectors. The technology exists; the regulatory will hasn't been articulated.
• Internalizing costs via price. Pricing that reflects the external cost. It clashes with the dominant freemium model and with the social expectation of free service inherited from classic software.
• Siting with lower-impact criteria. Data centers in regions with a predominantly renewable electrical mix. Some companies (Google, Microsoft) declare that criterion; effective implementation is partial and independent verification is limited.

The concrete figure: the study by Shaolei Ren et al. (UC Riverside, Making AI Less "Thirsty", first published in arXiv:2304.03271 in 2023 and expanded in 2025) estimated the water footprint of LLM inference at approximately 0.5 liters of water per 20-50 responses in typically cooled data centers. For ChatGPT with hundreds of millions of weekly users (figure declared by OpenAI in August 2025), the aggregate sits at millions of liters a day just in the fraction allocated to direct cooling, before adding the indirect water footprint of electricity generation (thermoelectric plants consume water for cooling). The bill comes; the air and the water charge it. Whoever signs the bill, if anyone signs it, will probably be someone who wasn't in the conversation when the product was designed.

Definitions

• Externality: a cost or benefit of an economic activity that affects third parties not participating in the transaction and isn't reflected in the price. A concept formalized by Pigou (1920) and reformulated by Coase (1960).
• PUE (Power Usage Effectiveness): a data center efficiency metric. Typical current values between 1.2 and 2.0.
• Marginal electrical mix: the composition of electricity sources that covers the additional demand when consumption rises. Distinct from the average mix, frequently dirtier.
• PM2.5: fine particulates suspended in the air with a diameter of 2.5 micrometers or less. Associated with respiratory morbidity and mortality documented by the WHO and the EPA.

References

• Crawford, K. — Atlas of AI (Yale UP, 2021).
• Hu, T.-H. — A Prehistory of the Cloud (MIT Press, 2015).
• Coase, R. — The Problem of Social Cost. Journal of Law and Economics 3 (1960).
• Pigou, A. — The Economics of Welfare (Macmillan, 1920).
• de Vries, A. — The growing energy footprint of artificial intelligence. Joule 7.10 (2023).
• Ren, S. et al. — Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models. arXiv:2304.03271 (2023, expanded 2025).
• EPA — Clean Air Act Resources for Data Centers (December 2025).
• EirGrid — suspension of new data center connections in the Dublin area due to grid capacity restrictions (from late 2021/2022), reported by Data Center Dynamics, "EirGrid says no new applications for data centers in Dublin until 2028." https://www.datacenterdynamics.com/en/news/eirgrid-says-no-new-applications-for-data-centers-in-dublin-till-2028/ The lifting of the moratorium at the end of 2025, conditioned on own generation or storage, was reported by Energy Connects, "Ireland Ends Moratorium on New Power Links to Data Centers" (December 2025). https://www.energyconnects.com/news/utilities/2025/december/ireland-ends-moratorium-on-new-power-links-to-data-centers/
• Data Center Dynamics — "Singapore lifts data center moratorium - but sets conditions" (2022): the 2019 moratorium was partially lifted with mandatory efficiency criteria (PUE ≤ 1.3, among others). https://www.datacenterdynamics.com/en/news/singapore-lifts-data-center-moratorium-but-sets-conditions/
• OpenAI / CNBC — "OpenAI's ChatGPT to hit 700 million weekly users" (August 4, 2025): the weekly users figure declared by the company. https://www.cnbc.com/2025/08/04/openai-chatgpt-700-million-users.html
• EIA (U.S. Energy Information Administration) — "What is U.S. electricity generation by energy source?": in 2023 around 60% of U.S. electricity generation came from fossil fuels. https://www.eia.gov/tools/faqs/faq.php?id=427
• Loudoun County Economic Development — figure put out on the proportion of world internet traffic that crosses "Data Center Alley." https://loudounpossible.com/data-centers
• Next 10 (report co-signed by Shaolei Ren, UC Riverside) — analysis of the health impact of electricity generation for data centers in California, 2019-2023 (2025). https://news.ucr.edu/articles/2025/11/21/california-data-center-health-impacts-tripled-4-years
• Directive (EU) 2023/1791 — Energy Efficiency Directive.

Going deeper

• Hardin, G. — The Tragedy of the Commons. Science 162 (1968).
• Bratton, B. — The Stack: On Software and Sovereignty (MIT Press, 2015).
• IEA — Energy and AI (Special Report, April 2025).
• Pellow, D. — What is Critical Environmental Justice? (Polity, 2017).

You might also like

• Scaling conversations, scaling consumption
• The energy cost of a conversation
• The energy footprint of outsourced thought
• The sustainability of artificial intelligence

Elsewhere

• Ember — Electricity Data Explorer
• IEA — Energy and AI, executive summary