Don't Stop at the Combine, Y-Combinator
YC is calling for AI for Low-Pesticide Agriculture. But the vision shouldn't stop at the combine harvester.
YC: Bring on AI for Low Pesticide Ag
Y Combinator, the renowned startup acceleration program, publishes a quarterly Request for Startups — a list of the things its partners most want founders to build. It is the closest thing the technology industry has to a weather forecast, and this quarter it carried an unusual amount of signal. The Summer 2026 edition runs to sixteen categories, roughly half of them hardware- or capital-intensive: counter-drone defense, inference chips, semiconductor supply chains, industrial manufacturing on the Moon. The list opened, though, with something closer to home.
Garry Tan, YC’s president, led with agriculture. Not the agriculture Silicon Valley has historically funded — dashboards and farm-management software — but the chemical core of how the world grows food. His thesis rests on a single sentence that anyone who has worked in this industry will recognize immediately:
Farmers are stuck, he wrote, in a bad loop: “Use more chemicals → get diminishing results → pay more → take on more risk.”
That is the in-field pesticide loop, and Tan’s argument for breaking it is that three things changed at once. AI can now identify individual weeds and pests in real time. Sensors and cameras became cheap enough to put everywhere. Robotics can treat one plant instead of blanketing a field. He goes further, into biology — microbes, peptides, RNA-based controls that can replace whole classes of synthetic chemistry. The prize, in his framing, is a company that cuts pesticide use by ninety percent while raising yield.
He is right that it is a generational opportunity.
He is also describing only half of the value chain.
The loop does not end at harvest
Everything Tan says about the field is true again, and arguably worse, in the warehouse. The crop comes in clean and then sits in store for months, and the moment it does, a second chemical loop begins. Stored grain has to be protected from insects, and the protection of choice for half a century has been phosphine fumigation. Phosphine is cheap, it is simple, but doing it right requires expert knowledge and the right tools.
Insects evolve in store exactly as weeds evolve in the field. Strong resistance to phosphine is now documented across the major grain economies, and in Australia — the most exposed and best-studied market in the world — it is severe enough that standard label rates no longer reliably achieve a kill. The operator’s response is the same one Tan diagnosed: dose harder, hold longer, fumigate more often. And the bill for doing so is as ugly as the field equivalent.
YC mapped the field. Post-harvest is the missing half of the same loop — structurally identical, and in several respects worse.
This is the part of the chemical-dependency problem that the Request for Startups did not address – but maybe it will do so in the future. Roughly a quarter to a third of the industry’s chemical-input pain sits after the harvest, in storage and handling, and almost nobody is building for it. The in-field robotics race is loud, crowded, and enormously capital-hungry. The warehouse and the shipping network, by contrast, are quiet — and it is where a great deal of the value is being lost.
Where the value actually sits
Centaur is a storage-side company, and it is worth being plain about that rather than dressing it up. We do not build field robots and we are not trying to.
The interesting territory for us is not upstream in the field; it is in the months between harvest and sale, where the crop is most valuable, most vulnerable, and least instrumented. That is where operators are bleeding dry matter, energy, and chemistry, and it is where the same forces Tan is betting on — cheap sensing, real-time intelligence, biology displacing chemistry — have barely been deployed.
The point here is not to claim a place inside YC’s category. It is to notice that the thesis YC has just validated for the field is, if anything, more acute downstream of it.
Nitrogen is the “AI meets atoms” sentence
Tan’s category quietly rewards a particular shape of solution: residue-free, precise, biologically grounded. The post-harvest world already has its analogue of precision treatment, and it is controlled-atmosphere storage — displacing oxygen with nitrogen (or CO₂) until the store itself becomes hostile to insects. It leaves no residue. It is compatible with organic certification. It does not breed resistance. It can both supplement and complement phosphine.
The reason it hasn’t taken over is not biology. It is uncertainty and gas economics. Did I actually pull oxygen low enough, for long enough, to kill every life stage? Where is the seal leaking? How long do I have to hold, and how much nitrogen will that burn? For decades those questions had no good answer, so operators defaulted to the chemistry they knew. That uncertainty is precisely what real-time CO₂ and O₂ sensing, paired with a CFD digital twin of the store, is built to remove — confirmed lethal exposure, leak localization, dose-and-hold optimized against the actual physics of the structure rather than a rule of thumb.
We have been working on the confidence layer that makes residue-free nitrogen control pencil out — confirmed lethal exposure, leak localization, dose-and-hold optimization — and makes the fumigant you can’t avoid precise enough to use a fraction as much.
That sentence does deliberate double duty. It serves the organic operator who wants nitrogen to work and has been burned by operational uncertainty. It serves the conventional operator who still needs phosphine but wants to manage resistance by dosing accurately instead of dosing more. We are not asking the market to pick a side. We are the intelligence layer under both — the same way Tan’s field thesis sits under both the robot and the biological. Software is the substrate. The atoms are the point.
The people closest to the problem already agree
None of this is merely a forecast. The grain industry’s largest handlers are already moving on it: some of the biggest, operating in regions where phosphine resistance is most acute, have been investing in the approach.
And the research backbone is public — Centaur’s IoT for precision fumigation has been carried out alongside the USDA’s Agricultural Research Service, with co-authored papers out of its Kansas program. When the operators living with the resistance crisis and the public scientists studying it are independently validating the same thesis, the loop is not theoretical. It becomes the daily economics of the warehouse.
The substrate, not the moat
YC’s own framing across this Request for Startups is that, in an era when anyone can generate code, software is the substrate and not the moat. That is exactly right, and it is the reason the warehouse is a better business than the field. The defensibility was never going to be a dashboard. It is the physics-grounded confidence layer over atoms — gas, biology, respiration, the behavior of a sealed structure full of living grain — that is hard to build and hard to copy.
The field robots will get funded loudly, and they should. But the other half of the loop, the half that begins the moment the crop is safely in store, is still open. YC mapped the field. We are building the AI warehouse.
Like this? Give me a holler at sotiris at centaur dot ag.