For centuries, humans have understood plants from the outside in. We observe leaves, test soil, read the sky. Modern agriculture has evolved with moisture sensors, multispectral imaging satellites, and handheld spectrometers — sophisticated tools that map environmental variables with increasing precision. But they all share a fundamental limitation: none of them read what the plant itself is experiencing. It's like diagnosing a patient by only looking at the room they're in, without ever measuring blood pressure, heart rate, or vital signs.
A startup called Electric Plant Company, emerging from stealth today after three years of development, wants to change that paradigm. Founded by Bloomscape's former CEO Justin Mast and neuroscientist Lee von Kraus, who previously worked with DARPA and the U.S. EPA, the company has launched Bombadil — the first foundation model in history trained specifically to decode plant electrical signals. The result is a system that reads a plant's physiological state directly from its own electrical activity, without relying on soil, weather, or visual observation.
Bombadil is, in AI technical terms, what the field would call a "foundation model" — a model pretrained without labels on a massive quantity of raw data, capable of generalizing to tasks it never saw during training. In this case, more than 4 million hours of electrical recordings from 1,400 plants across 100 species in ten distinct environments — from laboratories and greenhouses to commercial orchards and natural forests. Since the first training run, the corpus has grown to 7 million hours — more than all publicly available recordings of human brain activity combined.
The hardware architecture is deliberately simple and low-tech: sensors the size of a folded iPhone, attached to the plant with stainless steel screws that penetrate the cambium, the layer of living tissue just below the bark. These screws connect to the natural microvolt variations occurring within the plant. Data is sent to a base station via a low-power wireless network. On average, just two sensors per hectare are needed for accurate results, according to the company.
What happens next is purely in the realm of machine learning. The raw electrical signal reaches Bombadil in the cloud, where the model estimates seven physiological stress markers solely from the signal — including photosystem II efficiency, leaf temperature relative to ambient air, and the Crop Water Stress Index (CWSI), an established agricultural standard for measuring water deficit. The model's differentiator is not point accuracy but generalization: it can read the state of species it never encountered during training. "Most earlier models of plant electrical signals worked for a single species in a single environment," explained James Kirk, head of AI/ML at the company. "When you changed the plant or the environment, you had to start over. Bombadil learns what's common across species, and that generalization is what enables scale."
The scientific context that makes this possible has existed for decades. Burdon-Sanderson recorded action potentials in plants in the 19th century. Volkov systematically mapped plant electrophysiology in the decades that followed. Plants generate and propagate four distinct types of electrical potentials: action potentials (which propagate autonomously, as in Mimosa pudica, the "touch-me-not" plant, at 44 mm/s), variation potentials (triggered by environmental stress), system potentials (involving H+-ATPase activation), and local potentials. These are the same electrical signatures that Bombadil decodes.
What Bombadil brings new to the table is scale. Recent academic work, including papers published in arXiv in 2025 and 2026, has already demonstrated that machine learning can detect water stress in tomato plants with 92% to 99.29% accuracy from electrophysiological signals. Other studies, such as the European WatchPlant project, used wearable devices called PhytoNode to monitor Hedera helix electrophysiology in outdoor environments, achieving macro F1-scores of up to 95%. But all these systems were restricted to a single species, a single environment, or a single task. The Electric Plant model is the first to attempt what AI researchers call "cross-species generalization" — learning electrophysiological patterns that apply to any plant.
The startup raised $11.2 million in a seed round led by Onto Ventures and Climactic, with participation from Ponderosa, Endeavor Catalyst, Ludlow, Michigan Rise, The Fund, Schox VC, and Doug Song. Early deployments are already running at commercial sites with hundreds of sensors in the field. The company licenses its technology to plant-care service operators rather than replacing them, and claims the system can boost yields while reducing water use by more than 20%.
But there is a philosophical question that transcends agriculture. If an AI model can decode what a plant "feels" from its electrical signals, are we truly reading nature or simply translating electrochemical patterns into variables familiar to us? The founders of Electric Plant are cautious: "We're listening more than we're speaking," says Mast. The analogy with large language models is almost inevitable — GPT learned the structure of human language by reading trillions of words without ever "understanding" what it read. Bombadil learned the structure of plant electrical signals by reading millions of hours of raw voltage. The difference is that plants have no brain to interpret what their signals mean. They only send. The model decides what that means.
What follows may be the beginning of something radically new in the relationship between humans and the plant kingdom — a bidirectional dialogue where every plant in a field or forest becomes an active node in a biofeedback network — or it may simply be the next chapter in an industry that has already transformed every aspect of nature into data to be collected, analyzed, and optimized. The fact remains that, for the first time in history, we have the technical capability to read a plant from the inside out. What we do with that capability is what matters.
Sources: TechCrunch, Business Wire, arXiv
✓ Independent sources cross-checked and verified before publishing
