This San Francisco Startup Says Plants Have Been Talking All Along

California's orchards are awash with sensors that measure everything but the trees. There are soil probes for moisture, weather stations to record heat and wind, and satellites in orbit to see how green a field is. What they do not do is ask the plant how it is faring. Electric Plant, a San Francisco startup, believes the plants have been answering all along through tiny electrical signals in their stems, and it has put together an inexpensive sensor and an AI model to hear them out.

Electric Plant emerged from stealth on Tuesday with $11.2 million in seed money and a model of its own making called Bombadil, which it touts as the first foundation model for plant electrophysiology. The proposition is simple: put a small probe in a tree, send its electrical activity to the cloud, and tell a grower the tree is stressed or thirsty before anyone could see it by eye.

What Electric Plant Actually Built

The hardware side is a sensor named Bud, "a low-cost probe that attaches to any tree in minutes and streams its electrical signals continuously," as the company puts it. Stainless steel screws fix it to the woody stem down to the cambium, the living tissue under the bark, where it registers changes measured in microvolts. Those readings make their way over a low-power wireless link to a base station and then to the cloud for the model to make sense of. At tens of dollars apiece, the company says, the sensors are a bargain next to the thousands some monitoring tools cost, and a couple per acre will do.

The software is where the ambition lies. According to the technical white paper released with the launch, Bombadil was pretrained on more than 4 million electrode hours of recordings from over 1,400 plants of 100 species at ten locations, from labs and greenhouses to commercial orchards and natural forests, between November 2024 and July 2026. Much like the models behind chatbots, it did its learning without labels, hiding portions of the signal and teaching itself to fill in the blanks. A further layer trained on labeled data then turns the electrical activity into seven physiological stress markers, among them photosynthetic efficiency, leaf temperature relative to the air and the Crop Water Stress Index, a standard measure of how thirsty a crop is.

The company says the data set has since swelled to 7 million hours, and it intends to add new readouts such as disease flags or bloom timing onto the existing model instead of building a fresh one for every task.

Why Read the Plant Instead of the Soil

The notion of plants running on electricity is hardly an outlier view. For over a hundred years botanists have studied electrical signaling in plants, and research bears out that a stem or leaf produces a measurable change in voltage when it is injured, deprived of water or exposed to heat. What has been missing is a way to make those faint, noisy signals useful to a farm manager.

Electric Plant's argument is that the instruments growers already use gauge the environment around a plant, not how the plant is reacting to it. A soil probe might show that the ground is dry, but it cannot say whether a given tree is in distress. Satellite images do not pick up stress until it has changed how the canopy looks. Put a sensor in the stem, the company says, and you get a real time read on the plant's internal condition; it claims the system can boost yields while cutting water use by 20 percent or more.

"Bombadil makes that kind of intimate knowledge possible at scale, using data derived from inside the plant itself," co-founder and chief executive Justin Mast said in the company's announcement, comparing the model to a master gardener who spends a lifetime learning a single plot.

The Team and the Money

Mast previously founded the online plant business Bloomscape. His co-founder and chief scientist, Lee von Kraus, is a neuroscientist who has worked with DARPA, the Environmental Protection Agency and NASA, and who co-founded Halo Neuroscience. The pair set up Electric Plant in 2023, and their background shows in the way they go about things: the company handles a plant's electrical activity in much the same way brain researchers handle neural recordings. It even claims its data set is now bigger than all publicly available human brain recordings put together.

Onto Ventures and Climactic led the seed round, with Ponderosa, Endeavor Catalyst, Ludlow, Michigan Rise, The Fund, Schox VC and Dug Song also taking part. Electric Plant says early deployments are already running at commercial sites with hundreds of sensors in the field, although it has not said who the customers are. Rather than selling hardware directly to farmers, it licenses its technology to existing plant care operators and is recruiting what it calls founding operators as it expands crop by crop.

"Bombadil turns previously unreadable noise into a foundation model," Tom McQuillen, a partner at Onto Ventures, said in the announcement.

What the White Paper Shows, and What It Does Not

In an unusual move for a seed stage launch, Electric Plant put its figures out there, and they make for less dramatic headlines. In drought testing on six heavily labeled species, how closely the model's estimates followed the measured stress markers ranged from a correlation of 0.27 for the Crop Water Stress Index to 0.71 for linear electron flow, with most in between. Pretraining roughly doubled agreement with ground truth compared with an untrained model, which suggests the method is sound, but as early research rather than evidence of how it performs in the field.

The paper does not shy away from its limitations. One species performed poorly. Depending on the marker, some 9 to 25 percent of the sensors tracked in the wrong direction. The labels covered only normal watering and drought, no plant was recorded for longer than three months, and no test site was left out of training, so the authors have no way to tell site effects apart from species effects. The headline claim that Bombadil can read a species it has never seen holds most clearly for the underlying representation; reading a new species well still took some labeled calibration data.

James Kirk, head of AI, made the company's case in the announcement: "Bombadil learns what is common across species, which is why it can read a plant it never saw in training." The 20 percent water saving, for its part, is a company claim that does not appear in the white paper's results.

What Happens When Every Orchard Can Report Its Own Thirst?

The near term opportunity lies where water is at a premium and trees are valuable. In California's Central Valley, as well as in Arizona and Spain, orchards, vineyards and nurseries already spend money on precision irrigation. A sensor that costs tens of dollars and can tell a grower which block is actually stressed could pay for itself in a single dry season. Foresters and cities could use the same probes to spot drought stress in a watershed or a street tree before it dies off or becomes wildfire fuel. For Silicon Valley, Electric Plant is riding a wider current: foundation models trained on biological signals rather than text or images, from heartbeats and brain waves to the stems of trees. If the trend holds, the firms with the largest and most varied signal libraries will have the kind of edge that web scale text gave the early chatbot builders.

The concerns start with evidence. Correlations in the 0.3 to 0.7 range are promising for a research model, but they are not numbers a grower will bet a harvest on, and the company has yet to publish independent field trials or name customers who can vouch for the water savings. Then there is the Bud sensor having to be screwed into living tissue; that is no trouble with trees, but it is harder to make work on an annual row crop like lettuce or strawberries. On the business side, the licensing model leaves growers dependent on operators and on Electric Plant's cloud, and the data from their orchards becomes training material for a model they do not own.

The controversy is over language. The company and its founders talk about one day letting people speak with plants, which touches a nerve in a long running dispute among botanists. Some researchers want plant signaling studied as a kind of perception or communication; others feel that terms like intelligence and language overstate what is happening and breed hype. Electric Plant's supporters can say with some justice that a farmer does not care about the debate if the sensor reliably flags a thirsty tree three days early. The doubters can answer that such reliability has not yet been demonstrated.

The signals to watch are concrete. Will Electric Plant publish peer reviewed results or third party field trials to back up the 20 percent water saving? Will it name commercial customers and report renewals once a full season is done? Can Bud, or a successor, work on non woody crops? And will the big agriculture and irrigation companies license the model or build their own? Those answers will show whether listening to plants is a lab novelty or the next layer of farm infrastructure.

Below the Soil, Inside the Stem

At Sparknify we have made the case that the farm technology that matters most is not always the most eye catching. Our piece The Silent Crisis Beneath Every Farm examined the pathogens living in the soil and why tending that underground ecosystem matters more than any autonomous tractor or drone. Electric Plant takes the same instinct a step further: rather than deducing a plant's condition from what is around it, it puts the question to the plant itself. Healthy soil and a stem that can be read are two halves of one idea, that food security depends on understanding living systems and not just measuring their surroundings.

For most of history, a tree had to be seen wilting before anyone knew it needed water. Electric Plant is betting that soon the tree will send word first.

Subscribe to Sparknify for Silicon Valley tech news and the people behind it.