What the vertical farming failures taught

Between roughly 2018 and the middle of this decade, vertical farming absorbed an extraordinary amount of investment and then produced an extraordinary number of insolvencies. The usual summary is that the technology failed.

It did not. The plants grew. Yields per square metre were often exactly as projected, food safety was excellent, water use was a fraction of field production, and the growing systems worked as designed. What failed was the arithmetic connecting those results to a price the market would pay.

That distinction is worth insisting on, because the same reasoning error appears across agricultural technology and it is not confined to one sector.

The error

The error was treating a demonstrated technical capability as evidence of an economic one.

A vertical farm can grow lettuce. This was never in question and was demonstrated repeatedly. What follows from it is nothing at all about whether lettuce grown that way can compete with lettuce grown in a field, because that comparison depends on the cost of the photons, the cost of removing the heat those photons produce, the capital cost of the building, and the price of the alternative.

Field lettuce is grown under free light on land that costs a fraction of urban industrial space. Beating it requires either a large premium or a large structural saving, and the sector generally assumed both would materialise: energy costs would fall, automation would remove labour, and consumers would pay for local pesticide-free produce.

Energy costs rose. Automation proved harder and less reliable than projected. And the premium consumers would pay turned out to be real but small - considerably smaller than the cost gap.

Why the error was easy to make

The demonstrations were genuinely impressive, and the pilot facilities genuinely worked. What pilots systematically understate is the cost structure of production at scale.

A pilot has attentive expert staff, which a production floor cannot afford. A pilot reports yield per square metre of growing area; a business pays rent on the whole building, including plant rooms, packing, handling and circulation, which can be a large fraction of the footprint. A pilot buys energy in small quantities on whatever the current tariff is; a business signs long contracts and carries the exposure when they expire.

Every one of these gaps runs in the same direction. That is the tell.

What generalises

The pattern is not specific to vertical farming, and it is worth naming because it recurs.

Technical success is not economic success, and demonstrating the first says nothing about the second. This applies equally to autonomous machinery, on-farm robotics, novel sensing platforms and much else in the sector.

Costs that scale with production are different in kind from costs that do not. A vertical farm’s lighting cost scales with everything it grows, forever. A one-off capital cost is amortised and eventually disappears. Business models that treat a permanent per-unit cost as though it were a startup cost do not become profitable with scale - they lose more money with scale.

Ask what the alternative costs. The relevant comparison for any agricultural technology is not zero, it is the existing practice it replaces, priced properly. This is the same discipline as asking what a trial’s control group was, and it fails in the same way when skipped.

The part that survived

The applications that are working are the ones where the value was never about beating field production on price. Propagation, where uniformity and disease-free status have real downstream worth. Research and breeding, where the product is environmental control rather than food. Herbs and specialist crops in markets that genuinely pay for freshness.

These were the least exciting parts of the original pitch. They are the parts that turned out to have a coherent answer to the question of why anyone would pay for the photons - which is the question the sector should have been answering from the start. The numbers are in vertical farming.