Direct air capture plants
Plants that pull CO2 straight out of the air and bury it are running, but at hundreds to thousands of tonnes a year, not millions.
Open in the interactive tree →Climeworks' Orca (Iceland, 2021) and Mammoth (May 2024, nameplate 36,000 t/yr) store captured CO2 in basalt rock via CarbFix. Costs are still around US$1,000 per tonne. Occidental's Stratos in Texas is designed for 500,000 t/yr.
As of now
Orca has averaged only about 600 tonnes a year against a claimed 4,000, and an Icelandic press analysis in May 2025 found Climeworks' machines capturing only a fraction of their stated capacity. Stratos, designed for up to 500,000 t/yr, was slated to start at the end of 2024 and Occidental now plans to begin operating it at the end of 2026. Worldwide, novel carbon removal is about 0.002 Gt/yr and growing about 40% a year (State of CDR, 3rd edition, June 2026).
Open steps
- ML-screened sorbent discovery High AI leverageUse machine learning on material databases to find or design sorbents with higher CO₂ capacity, faster kinetics and longer lifetime.
- Process optimization to sub-200 kWh/ton Medium AI leverageOptimize DAC system design (temperature, pressure, sorbent reuse) using process simulation and neural surrogates to reduce energy cost.
- Viable CO2 end-uses at scale Low AI leverageIdentify and evaluate CO₂ end-use pathways (EOR, building materials, chemicals, fuels) that can absorb gigatonnes economically.
Where AI could help
Medium AI leverage. AI speeds sorbent and process design, but plants must be built, powered and run for years, and need storage and paying buyers.
- Screen sorbents for uptake in humid air with learned atomic models
- Predict sorbent lifetime and regeneration energy before synthesis
- Optimise plant operation against variable power and weather
- Monitor storage wells and pipelines from sensor and seismic data
Shown so far
- In April 2025, MOFGen, an agentic AI system of language, diffusion and quantum-chemistry models, generated hundreds of thousands of candidate MOFs, and five were synthesised in the lab. source
- In August 2025, Meta and Georgia Tech's ODAC25 dataset covered about 15,000 MOFs with CO2, H2O, N2 and O2 adsorption and trained machine-learned potentials for direct-air-capture sorbent screening. source