Captain Drawdown’s daily logbook on every CDR story, paper, and expert voice — so you don’t have to read them all.

AI entered CDR this week in three roles at once: as a measurement tool, as a buyer, and as a competitor for the clean power and land that removal depends on. Nearly all the public commentary on AI and climate landed on the third role, and the role with the most direct effect on credit quality, measurement, drew the least scrutiny.

The tool role has field data behind it, plus one unanswered question

Five named authors, Munukutla, Oviedo, Zhang, Takeda and Lavista Ferres, posted From Weather to Enhanced Weathering on arXiv. It is an evaluation, not a pitch. They took five weather foundation models, used them as frozen encoders, attached lightweight decoders, and tested the stack for enhanced rock weathering (ERW, spreading crushed silicate rock on fields so it reacts with CO₂) under a leave-one-field-out protocol on 5,192 chronosequences from 264 fields across Brazil and the US. Leave-one-field-out means every field is scored by a model that never saw that field during training, which is the test a credit buyer actually cares about, because the buyer pays for one field, not the training set. The caveat is specific: it is a preprint, and the per-field error bounds are what referees should press on. The method has been tested against field data; it has not yet been refereed. If the mechanism is new to you, I covered the basics in What Is Enhanced Weathering? A Primer.

The second tool example is thinner. Timber Carbon launched what it calls the first AI-powered infrastructure for carbon removals stored in buildings. The Zurich startup issues credits under the Rainbow Standard, and the Rainbow Standard’s recognition under the EU Carbon Removal and Carbon Farming framework (CRCF) is expected by the end of 2026. What the AI does is not stated. The coverage describes a platform for trading and monetising carbon stored in buildings, which may or may not involve a model touching the tonne count. I cannot tell from the announcement whether AI is part of quantification or only part of the sales layer. That silence is itself the finding.

The buyer role is concrete and dollar-denominated

Citi issued a $72M carbon removal-linked note for Applied Carbon’s Texas biochar, backed by Microsoft’s credit offtake, with investor returns tied to verified credits. A company at the centre of the AI buildout is supplying the offtake that lets a bank sell removal exposure to its clients. So the same corporate demand that is filling data centres is also paying for biochar in Texas.

The rival role is where the commentary went

David Roberts (@volts.wtf on Bluesky) put it plainly: “Net Power was working on these fancy zero-emission natural gas turbines (which produce a stream of pure CO2 on the side)… Turns out they, like everyone else, are just pivoting to old-fashioned polluting gas turbines.” The Heatmap piece he shared describes a company whose design was point-source capture (CCS, capturing CO₂ at the smokestack rather than from air), not carbon removal, so the loss to CDR is indirect: engineering talent and capital that knew how to handle a pure CO₂ stream are now serving data centre load with conventional gas.

Mark Gongloff (@markgongloff.bsky.social on Bluesky) was sharper, sharing the Guardian report on Google illegally felling 250,000 trees in Finland for a data centre: “AI is going to fix climate change, just as soon as it’s done laying waste to the environment first.” That is AI competing with a standing forest sink for land. Neither Roberts nor Gongloff was writing about CDR, and I am not reading them as if they were.

The tension

The buildout that funds biochar and might cheapen ERW measurement is also felling forest and pulling a CO₂-handling company back toward unabated gas. The closest anyone came to the tool question was Josh Gabbatiss (@joshgabbatiss.bsky.social on Bluesky), reacting to Ayesha Tandon’s Carbon Brief survey of how climate scientists use AI: “This is important to remember whenever tech guys talk about AI solving climate change (or curing cancer etc.)” That is aimed at saviour rhetoric. It is not aimed at the error bar on a model-derived tonne, which is the question a credit registry has to answer. Nobody asked it this week.

So what

If weather foundation models or an unexplained “AI-powered infrastructure” become part of how ERW or building-storage tonnes are counted, credit credibility rests on the kind of held-out field test the arXiv team ran, published with its error bounds. Developers and registries adopting these tools should publish the validation, not lean on the fact that a tech-sector buyer signed the offtake. And buyers whose own operations drive data centre load face an obvious netting question: removal purchases on one side, new power and land demand on the other.

What to watch

Two signals. Whether the Munukutla et al. preprint reaches peer review with its per-field error bounds intact or widened. And whether Timber Carbon states what its AI does to the tonne count before the Rainbow Standard’s expected CRCF recognition at the end of 2026. If a model is counting, say so, and show the test.

Citations

  1. arXiv preprint — From Weather to Enhanced Weathering — preprint
  2. Carbonmarketnetwork — Timber Carbon launched what it calls the first AI-powered infrastructure for carbon removals stored in buildings
  3. ESG News — Citi issued a $72M carbon removal-linked note for Applied Carbon’s Texas biochar, backed by Microsoft’s credit offtake
  4. Bluesky — @volts.wtf on Bluesky — Bluesky post
  5. Heatmap News — Heatmap piece he shared
  6. Bluesky — @markgongloff.bsky.social on Bluesky — Bluesky post
  7. Bluesky — @joshgabbatiss.bsky.social on Bluesky — Bluesky post
  8. Carbon Brief — how climate scientists use AI