The sustainability numbers around vertical farming sound like marketing until you trace where each one comes from. Then something surprising happens: almost every one of them turns out to be a software claim.

I spent two years volunteering on Interfarm, a closed vertical-farming venture funded under TÜBİTAK's 1501 industrial R&D programme. TÜBİTAK is Türkiye's national research council, the counterpart of UK Research and Innovation. The funded proposal commits to targets rather than adjectives: 95% less water than traditional agriculture, eight times the output of conventional growing from the same footprint, 90% less fertiliser, zero pesticide, and a 25% energy saving that the document explicitly attributes to the AI control layer, which was my part of the project. Its business model is written against 11 of the 17 UN Sustainable Development Goals by name.

Those targets are the proposal's commitments, not measured outcomes, and everything below keeps them in that register. What interests me is their structure. Each of them is a control problem wearing an agriculture costume.

Where the water goes

Agriculture takes roughly seventy percent of the world's fresh water (the proposal cites the UN Environment Programme's assessment), and a field wastes most of what it is given through runoff, evaporation and drainage past the root zone. A closed hydroponic loop wastes almost none, in principle. In practice the loop only stays closed if something holds it closed. Nutrient solution recirculates. As plants drink, the reservoir concentrates, electrical conductivity climbs and pH drifts. An operator without fine control has one blunt tool, which is to dump the reservoir and remix, and every dump is the water saving leaking out of the promise.

So "95% less water" is a property of the control loop rather than of the pipes. The loop has to keep conductivity and pH inside ranges tight enough that the solution stays usable. The plumbing makes the saving possible, and the software makes it real.

The number that kills vertical farms

The industry's problem has never been growing lettuce indoors. It is paying for the photons and the climate. Energy is the line item that has humbled some of the best-funded ventures in the field. The proposal's own state-of-the-art review measures itself against Plenty of San Francisco, whose SoftBank-led round was, as the document notes, the largest agricultural-technology investment made to that point. When the American market prices a problem in billions, it is telling you where the hard part is.

That is why the proposal's most load-bearing target is the least glamorous one: 25% energy saving, assigned to the AI layer. The reasoning is mundane and correct. Fixed-speed actuators such as pumps, misters and dehumidifiers run at one intensity for every job and burn the same energy for every state of the room. A controller that knows the room's actual state can run variable actuators at the level the moment requires and no higher. Multiply small restraints across every actuator, every hour, all year, and restraint becomes the economics. A vertical farm whose control layer cannot do this is a very expensive salad.

Zero pesticide is an architecture

The zero-pesticide target works the same way. You do not achieve it by deciding not to spray. You achieve it with a sealed environment that pests cannot enter and a climate held far from the conditions in which pathogens bloom, which makes humidity control a form of plant medicine. Break the climate control for a week and the policy collapses. Once again the sustainability property lives in the loop rather than in the intention.

This is the general shape I took away from the project: in controlled-environment agriculture, the environmental claims and the control claims are the same claims. The UN SDG language on the proposal (Zero Hunger, Clean Water and Sanitation, Climate Action) cashes out, unglamorously, as set points held within tolerances by software reading sensors several times a minute.

Why this belongs in a national programme

The call Interfarm was written for, 1501's Digital Agriculture priority area, required a formal European Green Deal alignment section, and the proposal carries one. By the company's report, it ranked first among roughly 300 applications in its call. I did not see the scoring, so I state it as their figure, not mine. A five-person team in Ankara, under a consultancy agreement with a university agricultural-machinery department, worked evenings on the same problem San Francisco was pricing in billions. That is what a healthy national R&D programme is for. Being the person named against its AI work is among the things I am proudest of, because the stakes were bigger than the module.

The honest paragraph

I built the control layer and the operator application. I did not measure the farm's water or energy against a conventional baseline, and the project's early figures would not have meant what a reader would assume. The targets above are what independent reviewers judged fundable, not what I verified. If that distinction bores you, we want different portfolios. The whole argument of this essay is that in this field, the difference between a promise and a measurement is exactly the width of the control loop.