Marco D’Angelo | 16 Sep 2026
https://www.startupbusiness.it/en/foundation-models-in-agriculture-ai-that-does-not-own-its-own-data/197897/
There is a sector that generates free data with every pass of a tractor – billions of georeferenced data points a year – yet receives just 0.57 per cent of global venture capital: an all-time low, reached in the first quarter of 2026 whilst capital was flooding into data centres. Agriculture is simultaneously the sector richest in data and poorest in AI capital. Yet something has already made its way onto the fields: in 2025, John Deere’s customers treated over five million acres using selective spraying guided by computer vision, halving herbicide use according to the company’s data. The question for founders and investors is no longer whether artificial intelligence will enter agriculture. It is who will own the data used to train it.
The context: the foundational models already exist, but they say little about fields
A foundation model is the basic model: it is trained once, using vast amounts of data, and applications are built on top of it. ChatGPT is the foundation model for text; here we are discussing their equivalents for fields, crops and satellite imagery. The race to develop agricultural models is proceeding along two tracks. The first is geospatial: Prithvi-EO 2.0, the IBM and NASA model released in late 2024 with 600 million parameters trained on satellite imagery, is open source and is already being used to classify crops; through the Major TOM project, the ESA has made available over 169 million embeddings from Sentinel data, the largest collection of its kind. The second strand is linguistic and agronomic: Bayer has transformed its in-house expertise into E.L.Y., a compact language model available in Microsoft’s Azure catalogue, Syngenta has integrated generative AI into the Cropwise platform, Google In October 2025, it extended its agricultural models developed in India (field boundaries, predominant crops, harvest windows) to four Asian countries. On the regulatory front, Europe has set the timetable: from August 2026, the full requirements of the AI Act will come into force for high-risk systems, which include the safety components of autonomous agricultural machinery, whilst from September 2025, the Data Act will give farmers the right to access the data generated by their own machinery. Meanwhile, capital is moving in the opposite direction to share prices: deep tech accounted for a third of the agrifoodtech deals tracked by AgFunder in 2025, but without any mega-rounds, and no company in the sector raised more than 200 million in a single round that year. Lots of lab work, little fuel. In other words, the models are there: what is lacking is clarity as to who controls the raw material that sustains them.
The map: who builds, who reaps the benefits, who pays
The geographical distribution of the players is unusually concentrated. John Deere (Illinois) is a textbook example of AI embedded in machinery: See & Spray exceeded five million acres treated in 2025, with an average saving of 50 per cent on non-residual herbicide and a yield increase of two staia per acre in trials reported by the company. But Deere’s 2026 also reveals the other side of the story: in July, the settlement with the US Federal Trade Commission (FTC) obliged it to guarantee diagnostic software tools to farmers and independent repairers for ten years, whilst a parallel class action cost it $99 million. Digital Green (a non-profit organisation operating in India and East Africa) has, through Farmer.Chat, developed the most concrete use case for generative AI in agriculture: half a million farmers have been reached, according to the Rockefeller Foundation, with the cost per adopted practice falling, according to the organisation’s data, from $35 to less than one. KissanAI (India) has released Dhenu, one of the first agricultural language models in vernacular languages. Living Models (Paris) is the new name to watch in 2026: it emerged from stealth mode in March with seven million seed investors and is developing Botanic, a genomic foundation model trained on 43 plant species, with open weights. Corteva at the end of 2025, it entered into an agreement with Profluent to design proteins for gene editing using generative AI, and Heritable Agriculture, spun off from Alphabet’s Moonshot Factory in early 2025, applies machine learning to genetic improvement: breeding – the genetic improvement of plants – is the area where foundation models hold the most promise, because the data there is already structured by a century of variety registers. NASA has even announced the first operational geospatial foundation model in orbit, with processing carried out on board the satellite. On the machinery front, Ecorobotix (Switzerland, 150 million raised) and Carbon Robotics (Seattle, USA, a Series D round worth 70 million) are selling AI as ‘walking hardware’. And in Italy xFarm Technologies (a €36 million funding round in October 2024) controls the key element: the platform where data from half a million farms is collected every day.
The dark side: some sow data, others reap value
The first issue is ownership. The farmer generates the data, the platform aggregates it, the model capitalises on it: in between lie licence agreements that few people read. The Deere case is instructive precisely because of its limitations: the FTC won on the right to repair, but ownership of machine data remained untouched. In Europe, the Data Act is the first real counterbalance, and those developing systems today would do well to design for compliance rather than simply having to comply. The second issue is liability: incorrect agronomic advice is not merely an inaccurate answer; it is a lost harvest. The AI Act classifies robots as high-risk, not crop recommendation systems, which remain in a grey area where there is as yet no clear party responsible for errors. The third issue is the bias in datasets: models learn from maize and soya in the Midwest, because that is where data abounds. Hillside viticulture, terraced olive groves and Mediterranean mixed farming are statistically invisible, and a model that has never seen a sloping vineyard will offer advice suited to the plains of Iowa. There are those working to bridge the gap – from FAO projects to indigenous Indian models – but the bulk of the funding goes elsewhere.
Where can you build, if you want to?
Italy will not have its own general-purpose foundation model, and pursuing one would be unrealistic. However, it does have two assets that model builders lack: data on crops that the rest of the world does not grow, and the infrastructure to validate them. Protection consortia, cooperatives and testing centres possess time series data on high-quality vines, olive trees and fruit and vegetables that no American dataset contains: organising these into data cooperatives, which sell access to model developers on their own terms, is a business model yet to be explored. The European AgrifoodTEF network, coordinated by the Bruno Kessler Foundation in Trento with a budget of 60 million, offers the testing infrastructure that a start-up alone could not afford. And the first market is right on our doorstep: the twenty-year agronomic monitoring requirements for agrivoltaic installations funded by the PNRR are exactly the kind of client that pays for certified data. There is just one condition: start with proprietary and verifiable data, not with the model. Models become outdated in six months; no one can reconstruct a twenty-year time series on Sangiovese.
Note to the reader: the author works with investment funds active in the agritech sector, which may have backed – or may in future back – some of the start-ups mentioned. This article reflects solely the author’s personal views and is intended for informational purposes only. (Photo by Teo Sticea on Unsplash)
Fonti principali
John Deere, “See & Spray technology across 5 million acres”, 4 novembre 2025: https://www.deere.com/en/news/all-news/see-spray-technology-across-5-million-acres/
FTC, “FTC, states secure settlement with Deere & Company advancing farmers’ right to repair”, 8 luglio 2026: https://www.ftc.gov/news-events/news/press-releases/2026/07/ftc-states-secure-settlement-deere-company-advancing-farmers-right-repair
Microsoft, “Microsoft introduces new adapted AI models for industry” (E.L.Y. Crop Protection di Bayer), 13 novembre 2024: https://blogs.microsoft.com/blog/2024/11/13/microsoft-introduces-new-adapted-ai-models-for-industry/
Syngenta Group, “Syngenta Group adds cutting-edge generative AI to Cropwise”, ottobre 2024: https://www.syngentagroup.com/newsroom/2024/syngenta-group-adds-cutting-edge-generative-artificial-intelligence-genai-cropwise-0
Google, “Sowing seeds of agri AI models: from India to APAC”, 24 ottobre 2025: https://blog.google/intl/en-in/company-news/technology/sowing-seeds-of-agri-ai-models-from-india-to-apac/
IBM Research, “Prithvi-EO-2.0”, 6 dicembre 2024: https://research.ibm.com/blog/prithvi2-geospatial
ESA Φ-lab, “New AI-powered insights with the latest Major TOM embeddings”, 31 gennaio 2025: https://philab.esa.int/new-ai-powered-insights-with-the-latest-major-tom-embeddings/
Rockefeller Foundation, “An AI-powered app fights climate change while revolutionizing farming” (Farmer.Chat): https://www.rockefellerfoundation.org/grantee-impact-stories/an-ai-powered-app-fights-climate-change-while-revolutionizing-farming/
iGrowNews, “Living Models” (seed da 7 milioni di dollari, marzo 2026): https://igrownews.com/living-models-latest-news/
AgFunderNews, “Agrifoodtech funding is flat, the interesting part is where it isn’t”, 18 marzo 2026: https://agfundernews.com/agrifoodtech-funding-is-flat-the-interesting-part-is-where-it-isnt
CropLife, “2026 Q1 agtech venture capital investment and exit round-up”: https://www.croplife.com/smart-tech/2026-q1-agtech-venture-capital-investment-and-exit-round-up/
xFarm Technologies, comunicato sul round da 36 milioni di euro, ottobre 2024: https://www.xfarm.ag/en/blog-posts/xfarm-technologies-leader-europeo-nella-digitalizzazione-del-settore-agroalimentare-chiude-un-round-da-eu36-milioni
AgrifoodTEF (rete europea di test per AI e robotica agricola): https://agrifoodtef.eu/
Commissione europea, quadro regolatorio AI Act: https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai