What happened
Microsoft’s November 12, 2026 announcement describes TerraNova as a model built by combining its Aurora foundation model for weather and environmental prediction with Terradot’s scientific field data. Terradot works with farmers to spread powdered rock on agricultural fields. The technique, enhanced rock weathering, accelerates natural chemical reactions in which certain rocks weather in rainwater and consume carbon dioxide. The companies say the model is intended to help assess both which fields are promising and how much rock is likely to react in their soils.
The prediction problem is highly local. Carbon removal depends on rock type, soil chemistry, rainfall, humidity and temperature, and conditions can differ from one field to another. Terradot says its team has collected tens of thousands of soil samples from fields in Brazil and the United States over several years. Its measurements include rock weathering rates, water movement assessed through drilled wells, and geochemical changes. Such work requires specialized laboratory equipment, and the announcement says measuring and verifying carbon removal can represent a third to more than half of enhanced rock weathering’s cost.
TerraNova uses information about a field’s soil composition, minerals in crushed rock from nearby quarries, local humidity, rainfall and temperatures, and broader weather patterns including La Niña. Microsoft says that training the model on data from Brazilian fields enabled it to make predictions for fields in Mississippi. Rather than replacing physical measurement, the approach combines model predictions with a smaller number of soil samples. In simulations using agricultural-field data, researchers found that TerraNova could reduce the number of samples needed at the largest sites by 75% while maintaining the same level of accuracy.
The work builds on Aurora, which Microsoft introduced in 2024 and says was trained on more than a million hours of atmospheric data. Terradot co-founder Sasankh Munukutla worked with experts from Microsoft’s Climate Innovation Fund, AI for Good Lab and Microsoft Research to adapt Aurora to predict chemical changes in soil. Microsoft says earlier versions were accepted for a climate-focused workshop at NeurIPS 2026 and presented at an AI-for-science workshop at ICLR 2026. The findings are to be described in a forthcoming preprint ahead of peer review; the announcement does not describe a commercial release or price.
Terradot says rock powder already spread on fields in Brazil and the United States is expected to remove up to 100,000 tons of carbon dioxide over time. The company aims to remove one million tons annually by 2035. Microsoft characterizes TerraNova’s results as potential, and Terradot says continued development and extensive validation are needed before AI models like this can be used responsibly.
What the archive adds
New Runtime’s earlier coverage of Ai2 and Google described a different climate-modeling effort: coupled atmosphere-ocean simulations and projections over years to decades (November 4, 2026 note). It is related by climate science, not by the same project or method. TerraNova is a new event, not a continuation of that work.
Trend signal
This does not yet fit a pattern New Runtime tracks. TerraNova is a focused climate-science application, and the announcement does not establish a broader shift in agent systems, developer tooling or AI operating models.
Why now
Microsoft presents the work as addressing a practical constraint on scaling enhanced rock weathering: identifying promising sites and measuring carbon removal require extensive, expensive field sampling. The announced simulations suggest a way to reduce that burden, while the companies emphasize that validation is still needed before the approach can be used responsibly.
Why it matters
For teams building climate or scientific models, TerraNova illustrates a potentially useful division of labor: models can help decide where measurements are most informative, while field samples remain part of the evidence used to estimate outcomes. That is more consequential than a model merely producing a plausible map. If the reported sample reduction holds up outside simulations, it could reduce the cost of evaluating projects and help researchers test more locations with limited measurement budgets.
The result is not yet proof that carbon accounting can be safely automated. Buyers of carbon removal need credible evidence, and a lower sampling burden matters only if estimates remain accurate across soils, rock types and climates beyond the data used to develop the model. Teams considering similar systems should treat prediction as a way to prioritize and supplement measurement, not as a substitute for ground truth. The core operational question is whether the savings persist when independent teams test the model in new fields and verify the resulting carbon estimates.
Limits and uncertainty
Microsoft reports simulation results, not a completed independent validation across operating sites. The announcement does not quantify uncertainty in TerraNova’s predictions, specify the simulation design or provide a peer-reviewed paper; it says a preprint is forthcoming and that extensive validation remains necessary. The 75% sample reduction applies to the largest sites in the simulations and is stated to preserve accuracy there, not necessarily in every location or deployment. Terradot’s figures for expected carbon removal and its 2035 target are company statements, not independently verified outcomes. No product access terms or price are announced.
Skill to build now
If your team uses models to plan field research, define the measurement protocol before optimizing sample counts. Record the baseline sampling plan, the model’s predictions, and the soil and climate conditions for each location; then compare predicted results with samples held out from model development. That makes it possible to test whether fewer measurements preserve accuracy rather than simply producing fewer measurements.
Reader takeaway
TerraNova’s practical promise is not that AI removes the need to measure carbon removal. It is that predictions may help researchers use fewer soil samples without losing accuracy. The reported 75% reduction is promising but remains a simulation result awaiting further validation.