Direct air capture has substantial health and climate opportunity costs

Definitions

In all cases, a given benefit of $X is equal to a cost of -$X and all values are in USD2024.

Throughout, we refer to four types of benefits and/or costs, all expressed as annual levelized costs. These annualized costs represent the levelized financing of capital investments made at the time of deployment and do not imply repeated annual construction, cumulative capacity additions, or learning-driven cost reductions over time.

1) Financial costs: the dollar value of a given intervention.

2) Climate benefit/cost: the value of the avoided or emitted CO2 associated with a given intervention after applying a social cost of carbon of $342 USD202421.

3) Health benefit/cost: the value of the avoided or emitted local air pollutants associated with a given intervention after applying the Estimating Air Pollution Social Impacts Using Regression (EASIUR) health impacts model22 (see below for details).

4) Opportunity cost of DAC: the difference in climate and health benefit/cost between deploying a given amount of capital to build DAC and deploying the same capital to effect another grid intervention. Herein we specifically looked at utility-scale PV and onshore wind, though this framework could be applied more generally.

Overall modeling approach

The overview of our modeling approach is outlined below in Fig. 3. The climate and health benefits and costs of the interventions we model herein come through the increases and decreases in the emissions of CO 2 and health-damaging air pollutants. To model the opportunity costs of deploying DAC, we use two different model frameworks, CoBE Projection15 and avertr, a version of the US EPA’s AVoided Emissions and geneRation Tool (AVERT)23 adapted to the R programming language. We use CoBE Projection to model how the opportunity costs of DAC evolve over time as more renewables are deployed. We also use CoBE Projection to model the near-term opportunity costs and compare its results to avertr. We detail each model below.

Fig. 3: Overview of modeling framework. The alternative text for this image may have been generated using AI. Full size image Diagram depicting the overview of our model.

DAC input parameters and assumptions

Numerous DAC life cycle assessments (LCAs) have found that the net GHG and environmental health impacts of DAC are largely driven by the emissions associated with producing electricity for the systems (i.e., Scope 2 emissions)5,6,8,9,13,24. For this reason, we restrict this analysis to energy-related impacts. Meanwhile, levelized cost is driven largely by capital and non-energy costs5,7,8. We thus treat the levelized cost of electricity and levelized cost of DAC as independent in our analysis. Accordingly, each scenario assumes fixed technology costs and performance over its lifetime, and temporal variation in results arises solely from changes in grid emissions intensity and damages under exogenous AEO projections. We modeled four scenarios representing a range of different values for DAC efficiency and levelized cost of DAC. In each case we assume that our DAC system relies on electricity for 100% of its energy requirements, for instance using electric motors and a high temperature heat pump, as opposed to using waste heat. For efficiency of CO 2 capture, compression, and storage, we selected a range spanning current reported values for climeworks’ Mammoth plant at one extreme (5500 kWh/ton, “Stagnation”)25,26 to values lower than most anticipated scaled-up values (800 kWh/ton, “Breakthrough”) at the other5,7,8,27. We modeled levelized costs spanning $100–$1000/ton, representing the full range of estimated costs7,8,27, with the upper end informed by the price of carbon offsets currently offered by climeworks (see Table 1)28. Our “Stagnation” scenario is our best estimate of current efficiency and cost, “Efficiency Improvement” represents a major jump in efficiency and an incremental drop in cost, “Advanced Efficiency Improvement” represents further incremental improvements in efficiency and cost, and “Breakthrough” represents efficiency and cost values at the extreme low end found in the literature (see Table 1). We model each of the four scenarios in each of the U.S. EPA’s 22 eGRID regions in the contiguous U.S. (see Supplementary Fig. 1). We assume a 100% capacity factor, given the high capital-to-energy cost ratio. The power consumption we model varies by scenario because we normalize to a levelized cost of $100 million/yr; values fall between 34–91 MW. We chose a $100 million annual expenditure to fall within upper range of existing US solar and wind installations while not greatly exceeding the planned capacity of climeworks’ Mammoth DAC plant. We model these four DAC scenarios over the U.S. EIA’s 2019 AEO Reference Scenario as well as the EIA’s seven other scenarios (see below)29.

Renewable potential associated with avoided DAC capital expenditures

Normalizing all technologies to the same annualized expenditure is mathematically equivalent to comparing lifetime benefits per unit of upfront capital under fixed cost assumptions, while allowing technologies with different lifetimes to be compared on a consistent basis.

We calculated the power consumption of a DAC plant with a $100 million USD2024 annual cost as:

PC = (365 d/y x 24 h/d)-1 x ($100 million USD2024) x EF_MWh_ton/LCODAC(Eq. 1)

Where PC is the power consumption in MW, LCODAC is the levelized cost of DAC in USD2024 per ton CO2 removed, d is days, y is years, h is hours, and EF_MWh_ton is the efficiency expressed as MWh of electricity per ton CO2 removed.

We then calculated the amount of solar or onshore wind one could deploy with the same amount of capital using the 2024 levelized cost of energy (LCOE) published by LAZARD30. While LCOE varies by geography, within the U.S. this is largely a function of different capacity factors as opposed to other factors like labor and materials30. We thus assume a constant levelized cost of nameplate capacity (LCONC) across the country ($13.2/MWh or $115/(kW•yr) for solar and $22.6/MWh or $197/(kW•yr) for onshore wind). We also performed a sensitivity analysis dropping solar costs to $3/MWh of nameplate capacity (equivalent to ~$13.5 MWh of generation) and onshore wind to $10/MWh of nameplate capacity (equivalent to ~22 MWh of generation), in line with the low end of estimates for 205031.

We calculated the nameplate renewable capacity that could be deployed with a $100 million USD2024 expenditure as NC = $100 million/LCONC, where NC is the nameplate capacity that we input into CoBE.

AVERT takes nameplate capacity as an input (see below). CoBE Projection takes year-averaged generation for each eGRID region. To calculate inputs for CoBE Projection, we divided NC by the given eGRID region’s 2018 average capacity factor for the resource in question (solar or wind).

Islanded DAC

We modeled a grid-islanded DAC plant as a DAC plant directly supplied by newbuild, onsite solar + 12 h of utility-scale battery storage. We modeled the annual removal of the islanded DAC plant as:

annual_removal = Power x 8760 h/y / EF_MWh_ton =

($100 million USD2024)/(LCOS_12 h + LCONC_Solar + LCODAC) x (8760 h/y / EF_MWh_ton) (Eq. 2)

where annual_removal is the DAC plant’s annual CO2 removal in tons, LCOS_12 h is the levelized cost of 12 h battery storage (assumed to be $753 USD2024/kW•yr, linearly extrapolated from LAZARD’s32 2025 analyses of 1, 2, and 4-hour storage), LCONC_Solar is the levelized cost of solar for the eGRID region in question as explained above ($115 USD2024/kW•yr divided by the given eGRID region’s capacity factor), and LCODAC and EF_MWh_ton depend on the DAC scenario being modeled, as described above in Table 1.

Emissions impacts and opportunity costs of DAC modeled with CoBE projection

To estimate the projected future health and climate impacts of each scenario, we utilized the CoBE Projection tool. CoBE Projection is described in greater detail in previous work15. Briefly, CoBE Projection provides a prospective analysis through 2050 by modeling future grid conditions using the U.S. EIA Electricity Marketing Module (EMM)33. The EMM models electricity demand across 26 subregions, accounting for plant retirements, new generating units, fuel pricing, and responses to environmental regulations. For electricity, CoBE Projection focuses on CO 2 emissions, which account for 99% of the total CO 2 eq emissions from the electricity sector34.

To quantify the health impacts, CoBE Projection estimates the air pollution emissions of PM 2.5 , NO x , and SO 2 . Premature mortality is estimated using three reduced complexity models (RCMs): Estimating Air pollution Social Impact Using Regression (EASIUR), AP2, and Intervention Model for Air Pollution (InMAP)35. These models provide county-level estimates of premature mortality impacts associated with exposure to PM2.5 and its precursors (NOx and SO2). The RCMs are built from more complex chemistry, fate, and transport models (CTMs) that incorporate background atmospheric chemistry, population distribution downwind, meteorology, and other factors to estimate the health impacts associated with air pollution emissions. For this analysis, we used the mortality estimates from EASIUR. EASIUR is a statistical model derived by CAMx, a full computation air pollution chemistry and transport model with more explicit representation of chemistry than the other RCMS, which closely reproduces the modeled health impacts of the air pollution by CAMx36,37.

Electricity grid level health impact factors for each RCM were developed using the eGRID 2018 power plant level emissions data38. These impacts were then monetized using a value of statistical life (VSL) of $14.76 million39. The original CoBE Projection tool utilized the 2019 AEO reference case scenario to estimate future changes to the U.S. electrical grid. In this analysis, we expanded on this by incorporating the other 7 scenarios in the 2019 AEO to act as a sensitivity analysis for variations in the future outlook of the grid16. These scenarios encompass a variety of possible futures, including high and low economic growth, high and low oil prices, and electrical grid projections without the clean power plan.

CoBE limitations

The CoBE Projection model likely underestimates the health impacts of electricity generation for several reasons. First, CoBE calculates impacts based on the grid “average” MWh of electricity over a year. Since DAC plants are capital-intensive and thus intended to operate at essentially constant capacity over the year, they may use more electricity during off-peak times than is reflected by the annual average, thus underestimating DAC’s costs (an hour-by-hour model, AVERT, shows greater grid impacts than CoBE, see Supplementary Figs. 21–23). Second, CoBE is based on EIA projections, which, along with similar projections by the International Energy Agency (IEA), have substantially under-predicted renewables deployment40 and cost decreases and which model a narrow range of future grid compositions (Supplementary Fig. 20)41,42. Moreover, CoBE uses 2019 eGRID regions and 2019 AEOs because this was the last year where the regions spatially overlapped. Since adding renewables to a renewables-saturated grid has a smaller impact on emissions and public health than adding them to a fossil-heavy grid, CoBE’s reliance on EIA projections may result in an overestimate of the benefits of additional renewables. However, since an under-estimate of renewable deployment (which errs in favor of renewables in our comparison) likely coincides with an over-estimate of renewable costs (which errs in favor of DAC in our comparison), it is unclear how the renewables-bullishness of a given scenario would impact our comparison. Regardless, future work should extend this analysis to newer AEOs and to other projections that include greater renewables penetration. Third, CoBE only includes the health impacts of fine particulate matter (PM 2.5 ) and its precursors, so the health impacts of air pollution emissions that lead to population exposure to ozone and NO 2 are not included here, thus underestimating the costs of DAC. Fourth, other impacts of electricity generation and DAC across their life cycles are not incorporated in CoBE, most notably the health impacts of oil & gas extraction, methane leaks across the natural gas supply chain, and health impacts in communities near coal generators. Moreover, most DAC systems require pipelines, and construction of pipelines results in additional air pollutants and GHG emissions directly and due to the additional energy required to transport CO 2 through the pipelines9. This underestimates DAC’s costs and the embodied climate and health impacts of DAC and renewable energy-related materials. Finally, this analysis does not include ancillary activities related to renewables or DAC deployment, such as company-related travel. Recent disclosures indicate that when these factors are included, the current leading DAC company may not cover its own emissions26.

Present-day emissions changes associated with DAC modeled with Avertr

In addition to CoBE, we also modeled the avoided emissions of each scenario using avertr, a version of the U.S. Environmental Protection Agency’s (U.S. EPA’s) AVoided Emissions and geneRation Tool (AVERT)23, recoded in R. avertr reflects AVERT version 4.3 with a 2023 model year, the most recent version and year available. It produces results which differ only trivially from running AVERT directly using its Excel-based Main Module (regionwide annual results differences rarely above 0.001% for any pollutant). AVERT is an intermediate complexity electrical dispatch model. Details are available elsewhere23, but briefly, AVERT allows users to input basic design criteria for renewable energy generation projects, energy storage, and energy efficiency projects, or increases in load on the grid. After designating the type, location, capacity, and other design criteria of a project, AVERT then produces 8760-hour generation profiles of change in electricity generation or demand, as applicable. AVERT is also capable of directly taking a user-specified 8760-hour generation profile. AVERT then produces plant-level estimates of reductions in generation and emissions of CO 2 , nitrogen oxides (NO x ), sulfur dioxide (SO 2 ), ammonia (NH 3 ), and primary fine particulate matter (PM 2.5 ), and volatile organic compounds (VOCs) based on historical data of each power plant on the affected grid for each of its 14 grid regions.

We model the deployment of utility-scale solar and onshore wind using AVERT’s default region-specific capacity factors for these technologies. We model the deployment of a grid-connected DAC plant in each region as a flat increase in load on the grid, scaled up by the region-specific 2023 transmission-and-distribution line loss factor from AVERT. For example, in New England, the 2023 loss factor is 7.23%, so we model a 50 MW DAC plant as drawing a 50 / (1 - 0.0723) = 53.897 MW load on the grid. We separately modeled 63, 38, 34, and 91 MW DAC plants, corresponding with our four DAC scenarios.

Comparison of CoBE projection with Avertr results

For the year 2023, we compared the CO 2 outputs of CoBE Projection against those of avertr, an intermediate-complexity model, with all the same input assumptions. Because avertr’s grid regions are slightly different from CoBE Projection’s, we focused our comparison on grid regions that largely overlap and on the nationwide average (see Supplementary Figs. 21–23). Results are qualitatively the same as for CoBE, but effect sizes are larger.

Modeling of health and climate benefits and costs

We model the health and climate benefits (or costs, in the case of increased load on the grid) of grid-connected DAC, grid-islanded DAC, utility-scale solar, and onshore wind using an approach similar to previous studies43,44,45,46.

While the social benefit of a given quantity of avoided CO 2 emissions is different than the social benefit of the same quantity of captured CO 2 , the asymmetry is much smaller than the spread of social cost of carbon (SCC) estimates47. Thus, in the interest of model simplicity, we model avoided CO 2 emissions and captured CO 2 with the same SCC of $283 USD2020 per metric ton ($342 USD2024 per metric ton)21. Qualitative results were robust to a sensitivity analysis varying SCC from $10 to $500 USD2024 per metric ton of CO 2 (see Supplementary Fig. 24).

To model the health benefits of reduced air pollutant emissions (and, conversely, the health costs of increased air pollutant emissions) now at risk, we link the results from CoBE and avertr simulations to the results of EASIUR22, one of three major RCMs that estimate the health impacts of emissions of PM 2.5 precursors and are designed for rapid policy assessment35,48. EASIUR provides estimates of health impacts of emissions of the PM 2.5 precursors NO x , SO 2 , NH 3 , and PM 2.5 , and on a per-ton and per-source-county basis. EASIUR estimates the mortalities that result from air pollutant emissions and monetizes them using the value of a statistical life (VSL), equal to $14.76 million39 in USD2024. Qualitative results were robust to a sensitivity analysis varying VSL from $1 million to $25 million USD2024 (see Supplementary Fig. 24).

Of the three RCMs, we chose to use EASIUR because estimates from EASIUR are derived from the Comprehensive Air Quality Model with Extensions (CAMx), and of the three RCMs, the estimates most closely match those from the Community Multiscale Air Quality (CMAQ) Model49, which is often used by the U.S. EPA and others for regulatory impact assessment35,50,51.