Overestimating data center public health costs
Energy & carbon aren’t the only environmental impacts of computing, but are the public health impacts overstated?
Discussion about the environmental impact of computing tends to focus on energy and carbon. These are easy to understand because the goal is to get to (net) zero carbon, improve energy efficiency, and transition energy generation to clean sources like nuclear and renewables.
But energy & carbon aren’t the only environmental impacts of computing.
Water consumption is more complex because it’s not necessarily about getting to zero. The location context matters. Efficiency is always good, but there are many locations with abundant supply so this doesn’t need to be top priority. However, there are also regions of water stress (sometimes within the same country) where it really is important.
Siddik, Shehabi, & Marston (2021) discusses this in the context of the US:
we calculate spatially-detailed carbon and water footprints of data centers operating within the United States, which is home to around one-quarter of all data center servers globally. Our bottom-up approach reveals one-fifth of data center servers direct water footprint comes from moderately to highly water stressed watersheds, while nearly half of servers are fully or partially powered by power plants located within water stressed regions.
Then there are the other impacts - pollution, mining, recycling, chemicals…these are the subject of much less scrutiny.
Big tech! AI! Pollution! Billions of dollars!
I was interested to come across an FT article “Pollution from Big Tech’s data centre boom costs US public health $5.4bn” which reports on a new claiming:
Air pollution derived from the huge amounts of energy needed to run data centres has been linked to treating cancers, asthma and other related issues, according to research from UC Riverside and Caltech.
The academics estimated that the cost of treating illnesses connected to this pollution was valued at $1.5bn in 2023, up 20 per cent from a year earlier. They found that the overall cost was $5.4bn since 2019.
This is certainly worth looking into, but I found the headline quite alarmist, so I decided to look into the paper behind the reporting.
Unfortunately, the paper is not actually linked from the article. After some digging, I found it as a Dec 2024 preprint titled: “The Unpaid Toll: Quantifying the Public Health Impact of AI”.
…AI's lifecycle, from chip manufacturing to data center operation, significantly degrades air quality through emissions of criteria air pollutants such as fine particulate matter, substantially impacting public health. This paper introduces a methodology to model pollutant emissions across AI's lifecycle, quantifying the public health impacts. Our findings reveal that training an AI model of the Llama3.1 scale can produce air pollutants equivalent to more than 10,000 round trips by car between Los Angeles and New York City. The total public health burden of U.S. data centers in 2030 is valued at up to more than $20 billion per year, double that of U.S. coal-based steelmaking and comparable to that of on-road emissions of California.
The authors’ names rang a bell - the same who published another preprint1 overestimating AI’s water footprint back in 2023 - even so, data center pollution is an important topic that needs more research, so I looked through the paper.
Pollution from diesel generators
The paper covers several different environmental impacts across scope 1 (emissions from on-site backup generators), scope 2 (emissions from power generation on the grid), scope 3 (manufacturing and other third party sources).
I focused on Scope 1 because this is rarely discussed, yet is an important part of data center operations. As I discussed previously in relation to demand response, the purpose of a data center is to provide a reliable environment for competing equipment. This means guaranteeing power availability.
In the US (the focus of this paper)2, the grid is pretty reliable, but there are outages (which may be increasing). Data centers mitigate this by having rapid response batteries which provide interim power before long-term generators spin up (usually within a few minutes). These generators are usually fueled with diesel so have high emissions.
The paper appendix states the assumption is “the actual emissions are 10% of the permitted level” based on an analysis of “a dataset of the air quality permits: permits issued before January 1, 2023 and permits issued between January 1, 2023 and December 1, 2024”. The actual calculations aren’t provided, which would help clarify exactly what basis the conclusions are derived from.
Backup generators are tested regularly, but only for short periods (1-2 hours). If this were performed monthly, that’s only 12-24 hours per year. Further, most diesel generators in the US are limited to a maximum of 100 hours per year (4 days) due to the pollution levels.
If we take the preprint assumptions to mean 10% annual capacity (this assumption has been clarified in the comments on this post), equivalent to roughly 36 days of continuous use, then this starkly contrasts with industry norms and represents a 30-fold overestimation. Even in an extreme scenario, such as a 15-day grid outage (360 hours), generator usage would amount to just 4% of the year—still less than half of the preprint’s estimate. This inflated assumption artificially boosts the calculated Scope 1 emissions, despite their minor role in the overall emissions profile.
To clarify this point, consider this 2023 emissions data (note this is global, not US specific):
Scope 1 = 79,400 tCO2e
Scope 2 = 3,423,400 tCO2e
Scope 3 = 10,812,000 tCO2e
Scope 1 = 48,925 mtCO2e
Scope 2 = 1,658 mtCO2e
Scope 3 = 7,445,621 mtCO2e
Scope 1 = 144,960 mtCO2e
Scope 2 = 8,077,403 mtCO2e
Scope 3 = 16,624,000 mtCO2e
These figures reveal that Scope 1 emissions, which encompass generator contributions, are a tiny fraction of Scope 2 and Scope 3 emissions.
The authors could have leveraged such publicly available data (actual emissions vs estimated), been more specific about what their calculations are based on, and performed a sensitivity analysis to align their estimates with reality.
Other problems with the paper?
The 2024 US Data Center Energy Report (one of only two credible sources of data center energy estimates I discussed here) projects a range of 325 to 580 TWh (6.7% to 12% of total electricity consumption) in the US by 2028.
In the preprint they assume 519 TWh from assessments in a McKinsey white paper. This is drawn from a “medium” growth rate without justification. The linked white paper does not justify this scenario either, merely describing it as a “medium scenario” and citing “Global Energy Perspective 2023, McKinsey, October 18, 2023; McKinsey analysis”.
Unfortunately, this 2023 Perspective is not available publicly (a 26 page executive summary does not explain the various scenarios) and there is no further information about the McKinsey analysis. So why did they pick this number? This is important for understanding the basis of their analysis of Scope 2 and 3 emissions.
We see something similar with the analysis of GPU power consumption - they make the common mistake of using the maximum thermal design power (TDP) when estimating total energy consumption of AI training:
We consider Llama-3.1 as an example generative AI model. According to the model card, the training process of Llama-3.1 (including 8B, 70B, and 405B) utilizes a cumulative of 39.3 million GPU hours of computation on H100-80GB hardware, and each GPU has a thermal design power of 700 watts. Considering Meta’s 2023 PUE of 1.08 and excluding the non-GPU overhead for servers, we estimate the total training energy consumption as approximately 30 GWh.
As I discussed previously, this is too simplistic an approach. Workload configuration, batch size, and number of nodes has a major impact on power demand and manufacturer TDP is never reached.
Without digging into the paper any further, these assumptions suggest a pattern of overestimation which will compound errors in the final conclusions.
Conclusions
Computing’s environmental footprint is complex, involving energy consumption, water usage, pollution, resource extraction, and beyond. Preprints like "The Unpaid Toll" play a role in spotlighting overlooked issues, such as data center pollution.
However, flawed assumptions - such as overestimating generator usage by a factor of 30 or relying on maximum thermal design power (TDP) for energy estimates - can skew results and mislead readers. Here, the preprint likely overstates the public health impact by a wide margin, undermining its credibility.
For meaningful progress, the field demands rigorous, standardized approaches to evaluate computing’s full environmental impact, tailored to regional and global contexts. This means grounding analyses in real-world data - such as corporate emissions reports - and factoring in variables like local grid reliability or water scarcity.
Both researchers and journalists must approach preprints with caution, verifying claims before they shape public discourse or policy, especially on high-stakes issues like health and sustainability.
I don’t blame the authors for trying to gain publicity for their preprints per se. The state of academic publishing is terrible - it can take years to get through peer review, by which time your findings are likely out to date. However, anyone can publish a preprint - journalists need to be very careful about reporting on them because nothing has been verified, reviewed, or refined. I wonder how many preprints never actually reach publication.
Power grid reliability is much more of a concern in less developed countries, which also often suffer from other relevant issues e.g. poor water supplies, governance challenges and corruption, and high fossil fuel dependence.
Thank you again for your interest in our research.
I'll add the following to conclude my responses.
Your post contains multiple factual mistakes. Most notably, you repeatedly suggest indirectly or use directly the term "overstate" qualitatively without providing any quantitative assessment, except for your fundamental mistake in claiming a "30-fold overestimation." This factual error arises from your misinterpretation of "10% of permitted emissions," as reflected in your statement:
“If we take the preprint assumptions to mean 10% annual capacity (this assumption has been clarified in the comments on this post), equivalent to roughly 36 days of continuous use, then this starkly contrasts with industry norms and represents a 30-fold overestimation.” (Note: The sentence "(this assumption has been clarified in the comments on this post)" seems to have been added recently after you realized your factual mistake, but you still keep your misleading and irresponsible conclusion "30-fold overestimation".)
Whether this is due to your lack of knowledge in this field or other motivation and however you spin your claims in your later replies, your interpretation is factually wrong. As explicitly stated in our paper, we did not assume 10% of a year’s time, and the numbers we used are transparently disclosed.
Your post containing factually incorrect claims in its current form misrepresents our research, distorts the understanding of the field, and ultimately affects your own credibility.
I appreciate your continued interest in our research. However, I encourage you to engage with the findings accurately rather than spreading misinformation. Constructive discussions are always welcome, but misrepresenting technical details only demonstrates your lack of knowledge in this space and does not contribute to an informed dialogue.
1. “The actual emissions are 10% of the permitted level” refers to 10% of the emission limits set in backup generator permits—not 10% of a year’s time. While the actual emissions vary case by case, this figure serves as a reference and aligns with publicly disclosed government reports from Washington and Virginia.
2. The Berkeley Lab report was published after our study. Nonetheless, even our highest 2030 projection (519 TWh) remains within the 2028 range projected by Berkeley Lab. Thus, our estimates are on the conservative side. If you're interested in our updated estimates (to be included in our forthcoming update) based on Berkeley Lab's projection, please read: https://www.linkedin.com/posts/shaolei-ren-68557415_estimates-of-the-public-health-cost-caused-activity-7298237825433448449--pVD
3. We consider the training energy consumption, not inference. Even when incorporating TDP and accounting for PUE, our estimates remain conservative since they do not include server energy overheads, which typically add 20–50% to GPU energy consumption. For a more detailed discussion on training energy estimates, I recommend you review the literature (e.g., arXiv:2104.10350 and arXiv:2211.02001).
4. I also encourage you to read our updated paper on AI’s water footprint (arXiv:2304.03271), which has been accepted for publication in Communications of the ACM.