Microsoft and Terradot Reveal TerraNova Carbon Removal Model

Microsoft and Terradot have described TerraNova, a research-stage artificial intelligence model intended to cut the number of soil samples needed to measure carbon removal from enhanced rock weathering. In a Microsoft Source feature published on 12 November 2026, the company said TerraNova “builds on Aurora, Microsoft’s foundation AI model” for predicting global weather and other environmental conditions, and that it uses field data from Brazil and the United States.

In simulations using data from agricultural fields, Microsoft said, “the model could reduce the number of samples needed at the largest sites by 75% while maintaining the same level of accuracy.” That 75% figure is Microsoft’s simulation result, not a figure already demonstrated in independent operational audits. Microsoft said measuring and verifying carbon removal “can account for a third to more than half the cost of enhanced rock weathering.”

The work is still at the research stage. Microsoft said the research “will be described in a forthcoming preprint ahead of peer review.” A preprint, “From Weather to Enhanced Weathering: Weather Foundation Models for Carbon Dioxide Removal at Operational Scale,” was already posted on arXiv on 20 October 2026 as 2610.28973. Terradot listed the Microsoft feature on its news page the same day, 12 November.

Microsoft’s Climate Innovation Fund is an early investor in Terradot, a fact the Microsoft article itself discloses. That relationship should be kept in view when reading the companies’ claims.

Company claims, a preprint and what the paper actually says

Terradot, according to Microsoft, says rock powder it has spread over fields in Brazil and the US “is expected to remove up to 100,000 tons of carbon dioxide over time,” and that the company is “aiming to remove one million tons annually by 2035.” Those are Terradot’s own claims about expected and targeted removal, not independently verified totals in the Microsoft feature.

Microsoft chief sustainability officer Melanie Nakagawa said: “TerraNova is an exciting example of how AI can accelerate climate progress by addressing a real barrier to scaling carbon removal.” Terradot co-founder and chief executive James Kanoff said: “We’re going to need to remove nearly 10 billion tons of carbon dioxide from the atmosphere every single year” to avoid the worst effects of climate change.

The arXiv abstract reports an evaluation of five weather foundation models as frozen encoders, under a leave-one-field-out protocol, on 5,192 chronosequences from 264 fields across Brazil and the US. The authors say their best model, TerraNova-ERW, built on Aurora 1.5, “positively orders measurements within fields it has never seen.” They also say “zero-shot prediction of a new field’s mean removal is not yet feasible.” With few-shot prediction, blending the model with a few samples per field, they say TerraNova-ERW “roughly halves the samples needed to estimate a field’s mean removal, and can cut sampling on our densest fields by 28–98%.”

Those paper figures are not the same as Microsoft’s 75% simulation result for the largest sites. Both should be kept in their own boxes: one is a company simulation quoted in the 12 November feature; the other is the authors’ account of few-shot sampling in the preprint.

Terradot co-founder Sasankh Munukutla, who the Microsoft article says collaborated with the Climate Innovation Fund, the AI for Good Lab and Microsoft Research, said “continued development and extensive validation are needed to use AI models like it responsibly.” Microsoft presents the work as a complement to physical samples, not a replacement for them.

Erika Basham, director of natural capital investments at the Climate Innovation Fund, said: “We wanted to do more than invest in and purchase from Terradot.” The same article describes the fund as an early investor in the company.

Other climate tools on TSN are useful context, not the same method

TerraNova is a soil-sampling model for one carbon-removal method, enhanced rock weathering. It should not be collapsed into other recent climate-tech stories. TSN has just reported that UNEP says its methane satellites have helped close 46 major leaks, a tally UNEP itself published. That is an emissions-detection claim about methane, not a rock-weathering measurement tool.

A closer neighbour is work on models that stand in for slower climate simulations. TSN covered an MIT method for training climate emulators on far less data. That paper is about emulator design, not about counting carbon removed from crushed rock in farm fields.

Meta’s recent low-carbon supply deals, which TSN explained as book-and-claim certificates, are a different kind of climate claim again: contractual environmental attributes, not a soil-sampling model. The three stories share a climate-accounting problem — how to measure or claim a reduction — but they are not evidence for one another.

The model is a research tool, and the investor is also the messenger

The most important limit is the stage of the work. Microsoft is describing a model that, in its simulations, could cut sampling at the largest sites by 75% if further research confirms those results. The preprint’s authors say zero-shot prediction of a new field’s mean removal is not yet feasible, and that even the few-shot setting still depends on some physical samples. Munukutla has said extensive validation is still required. None of that is a deployed standard for carbon-credit markets, and the paper is ahead of peer review.

A second limit is the 100,000-ton and one-million-ton figures. They come from Terradot, as reported by Microsoft, and they are framed as an expected result of rock already spread and as a 2035 aim. An aim is not a contracted delivery, and an expected total “over time” is not a verified annual removal rate.

A third limit is the corporate relationship. Microsoft’s Climate Innovation Fund is an early investor in Terradot, and people from that fund helped adapt Aurora to the soil-chemistry task. Basham’s comment that the company wanted to do more than invest and purchase is a clear statement of that dual role. It does not make the research unusable. It does mean the 75% simulation result, the Terradot removal claims and the enthusiastic framing should be read as company-reported, and checked against the more cautious language in the preprint and in Munukutla’s own validation warning.

The date tension between Microsoft’s “forthcoming preprint” line on 12 November and the 20 October arXiv posting is best reported plainly, not resolved by guesswork. The paper is on arXiv; Microsoft still described it as forthcoming and ahead of peer review. Both statements can sit together: a preprint can be posted before a company feature, and still not have been peer reviewed.

Sources

1. Microsoft Source, Susanna Ray, 12 November 2026: How a new AI model could help crushed rock fight climate change
2. Terradot news page, listing the Microsoft article, 12 November 2026
3. Munukutla et al., arXiv:2610.28973, submitted 20 October 2026: From Weather to Enhanced Weathering: Weather Foundation Models for Carbon Dioxide Removal at Operational Scale

Related stories

• UNEP Says Methane Satellites Helped Close 46 Major Leaks
• Meta’s Low-Carbon Supply Deals Are Book-and-Claim Certificates
• MIT Method Trains Climate Emulators on Far Less Data