The Plant Is the Experiment: Learning and Failure in Industrial Biology

John F. Diener
Vertical Oceans
Singapore
john@verticaloceans.blue

Position paper
12 June 2026

Abstract

Industrial biology has a scale-up problem, and it is not the biology. Across mycelium leather, gas fermentation, cultivated meat, and others, two mistake patterns recur. The first mistake is applying scale to a learning problem: committing capital to an architecture whose exposure to unresolved biological uncertainty grows beyond its ability to resolve it. The second mistake is misaligned technoeconomics where productivity, cost, captured value and scalability are not aligned to the same reality. I set out both, using public failures and counterpoints against them, and argue for an architecture in which learning is by design: a fleet of independently addressable production units under shared intelligence, where production pays for experiments and the consequences of being wrong stay contained. Scale is a stock. Learning is a rate. The companies that survive scale-up will be the ones that built the rate before they bought the stock.

Keywords: industrial biology · scale-up · technoeconomic assessment · experimental design · aquaculture · fermentation · economies of intelligence

1. The position

Biological systems are inherently complex, and do not follow the deterministic relationship between input and output typical of most industrial processes. This reality becomes acute when scaling novel biological production systems. The opportunity in industrial biology is is measured in trillions of dollars. But, the track record in novel systems is mixed, and not because the biology does not work. It is mixed because of two structural mistakes that recur across sectors. Here I discuss each in turn.

The first mistake is applying scale to a learning problem. A company commits to a production architecture whose exposure to unresolved biological uncertainty grows faster than its ability to resolve that uncertainty. In the terms I used in Economies of Intelligence, it buys stock when the business needs rate [diener2026eoi]. The second mistake is a technoeconomic model that does not describe one business. Productivity, cost, captured value and scalability are each defended on their own terms, under assumptions that cannot all be true at once. The model assumes the stock exists before the rate has produced it. The two mistakes occur separately. The worst outcomes are where they occur at the same time.

Underneath both is a cost the industry has not fully priced. A large plant run at a fixed set point is a local optimum defended with capital. Every batch repeated at that set point is an experiment not run. Until recently that opportunity cost was zero, because experiments were expensive to run. In the current era of low cost intelligence, the opportunity cost is potentially existential.

2. Scale applied to a learning problem

2.1. A result is not an operating range

A single successful batch is evidence of possibility. An operating model needs evidence about repeatability, variance, constraints and recovery when conditions leave the intended range. The unit of understanding is the biological system in its total operating environment, not an organism or a metric in isolation. Bhagwat et al modelled 32 fermentation configurations and showed that the economic effect of a change in titer, rate and yield (TRY) depends on the feedstock, separations and facility around it. Optimizing a single TRY metric proved to be largely zero sum in determining the minimum product selling price[bhagwat2026]. The question before adding capacity is what must be true for the result, how consistently those conditions can be held, and what happens when they slip.

2.2. The scale fallacy in industrial biology

Large production units exist for good reasons. Capital cost per unit of volume falls with vessel size, setup costs amortise over long runs, labour per unit falls, and when experiments are expensive the rational design is to find a set point that works and build as much capacity around it as you can. That is minimum efficient scale, the equilibrium of a cost structure. Long runs at a fixed set point are optimal when there is nothing better to do than repeat that recipe. Repetition is optimal when learning is impossible.

In industrial biology learning is the whole point. Biological processes are complex and non-deterministic. A given recipe could be far from the optimum and thus poorly characterised. In that setting, pursuing batch scale with a fixed topology is setting up for failure. For novel systems, larger batch capacity may perform dramatically different than the smaller batch capacity in a pilot plant.

2.3. Production has to be informative, not just observable

In one facility, testing a change means altering a large connected system, waiting out a long cycle, and comparing the result with a previous batch grown under different conditions. In another, the operator runs the change on a few units beside a contemporaneous baseline, repeats it, and then rolls it out. Both produce. Only one produces decision-quality evidence.

More sensors do not close that gap. Randomisation, replication and blocking are what separate a treatment effect from everything else that varied [nist2012], and they still have to be provided when the plant is automated. Control engineers know a sharper version of the problem: the better a plant is controlled, the less its data reveals, because a system held at a set point generates almost no information about what happens away from it [ljung1999]. To learn, you have to excite the system deliberately, and that costs a little production. A plant designed never to deviate is a plant designed never to learn.

In mathematical terms, a directed experiment can be described as λ > 0 where λ is a learning term multiplied by a variance. In a constant recipe λ = 0, there is no variance. That is the model for large scale, large batch production. Causal models massively expand the response surface of a complex biological system. A fixed recipe may be an optimum, but an expanded response surface opens the possibility of a new global optimum. The only way to get there is through λ > 0 experimentation, probing new points on that response surface. Moving from one peak to another may involve experiments that go down before they go up. That negative variance (“regret”), is a learning cost. The process to find a new optimum is directed by an optimizer that directs learning to test specific points that provide maximum learning with minimum regret. A fleet of smaller batch units creates more learning opportunities per unit time while containing regret.

However, a collection of small units doesn’t guaranty a collection of independent experiments. Units sharing water, feed, genetics or air are correlated (“loop coupled”) and thus all participate in the same experiment to some degree. At Vertical Oceans some parameters are independent at the unit level and others are coupled through a shared water loop, but a given facility contains multiple independent water loops.

2.4. Blast Radius: the consequences of being wrong

I use blast radius for the extent of production, biomass, capital and recovery time exposed to a given failure. It is not the probability of the failure. Two designs can have the same probability of a contamination event and very different consequences. One design contains it, the other lets it propagate through an entire batch in a shared room, a shared water system or a centrally deployed instruction. Smaller units help only when the boundaries between units are real, and modularity can pay the cost it adds. This is not a case against large units; where the performance envelope is understood, they win on cost. The mistake is using scale to answer an uncertainty the organisation has not yet found an economical way to resolve. At the extreme the plant itself becomes the experiment: every lesson exposes a large share of production and every correction costs another long cycle.

2.5. Scale laws of Intelligence

Biological systems are dynamic over time. Genetic selection and modification change biological outcomes, sometimes significantly. Breeding programs are constantly moving the needle on organism performance. Genetics alone introduces a dynamic to the most fundamental input in biological production - the organism itself. That argues strongly against large batch capacity designed around a set-point, and demands a different scaling model.

Today’s environment of low cost intelligence fundamentally changes scale laws of many economic models, and especially industrial biology. Today’s intelligent alogorithms can model causality in large complex systems that would have been impossible even just a few years ago. The intersection of biological dynamics with algorithms that explain the causality in the system creates a new scale paradigm: Economies of Intelligence.

Economies of intelligence accrue to the system that is able to learn as it produces with a contained learning cost. A fleet of small production units creates more opportunities for learning, finding new optima. Economies of intelligence facilitate economies of scale when the new recipe is propagated across the fleet. Economies of intelligence is the best model for most biological systems because it enables the constant pursuit of optimal recipes against a changing landscape driven by genetics, nutrients, and other parameters.

3. The technoeconomic model has to describe one business

A system that learns well can still be a bad business if the technoeconomics don’t work. I test technoeconomics against four components. They are separate questions with highly interrelated answers.

Productivity is saleable output per unit of time: growth, survival, feed conversion and cycle length in aquaculture; cell density, titer, yield, rate and recovery in fermentation. A new technology does not have to beat the incumbent on every physical measure, and a dramatic gain in output per square meter contributes little if land is not the binding constraint. What matters is a demonstrable advantage in the integrated model, on repeatable performance rather than a best batch.

Cost uses credible operating performance and a like-for-like comparison with the alternative that serves the customer; a farm-gate operating cost set against an incumbent’s delivered price is not a comparison. Operating cost and capital requirement should be visible separately and then reconciled, because attractive cash margins can still earn a poor return on the capital they need. Higher cost is not the failure. The failure is a cost the rest of the model cannot carry.

Value is what reaches the producer. Quality, freshness, consistency and provenance are worth something only if there is a mechanism for capturing them: price, mix, reduced losses, better terms. The model needs the price actually realised after channel costs and discounts, and evidence of repeat demand at that price. Taking on distribution or a brand adds functions that have to be funded; the gain is what is left after those costs.

Scalability is the ability to add economically viable capacity on a relevant timescale. It does not require a bigger unit; more units, more sites or manufacturing partners all count. Technically, new capacity has to be built, commissioned and run without exceptional intervention. Financially, construction, ramp losses, working capital and time to cash generation have to earn a return external capital will accept.

The decisive test is whether the four describe the same business. A model combining pilot-scale productivity, factory-scale cost, the premium earned during scarce early production and the financing terms of mature infrastructure is internally inconsistent and will fail. A persistent gap with no credible route to closure is a signal to change the product, process, market or financing plan, not a reason to commit more capacity. The switch from greens to strawberries is a correction of TEA calculus: the captured value was less than the additional cost of producing greens in that farm. More on this and other examples next.

4. What the record shows

These are not controlled experiments, and a filing does not reveal every decision behind an outcome. I separate reported results from announced capacity and modelled economics, and treat the lessons as interpretations.

4.1. Scale bought before the operating range

Atlantic Sapphire’s 2021 annual report records a March mortality incident of roughly 500 tonnes head-on-gutted, carried at USD 4.6 million, with other independent systems unaffected, and describes splitting the six US grow-out systems into twelve to reduce future exposure [atlantic2021]. That is a company shrinking its blast radius after the fact. In 2025, feeding-capacity limits forced early harvests at lower weights and prices [atlantic2025]. By mid-2026 harvests were up and biomass stood at 3,592 tonnes, alongside a USD 20 million bridge and restructuring package and a mandatory offer from an investor group holding 68.73 percent with the stated intention of delisting [atlantic2026]; press coverage put the half-year net loss at USD 30.9 million [atlantic2026press]. Operations are improving. The point is the price of the cycle: when each lesson exposes a large share of production and each correction costs a long cycle, the capital structure is consumed before the operating model is proven.

MycoWorks shows the same failure in a different organism. Its 136,000 square foot plant in Union, South Carolina started production in September 2023 with a tray-based process, a stated capacity of millions of square feet of mycelium a year, and a USD 125 million Series C earmarked for it [mycoworks2023]. In October 2025 the company announced it would close the plant and stop growing mycelium, sourcing it from cheaper growers and tanning it instead. The chief executive’s letter said the plant had proved too expensive for the current world, its cost of capital too high, and that the team had “solved infections, yield loss and the operational complexity of a brand-new biomanufacturing process” [mycoworks2025]. My reading is that the architecture made those problems expensive to solve. Mycelium grown in large shared rooms has a large blast radius: one contamination reaches a whole batch, and a large room is harder to isolate, disinfect and return to service than a small one. The biology was learnable. The plant was built before the learning was done, and when capital became scarce there was no cheap way to finish it.

Calysta announced plans in 2017 for a FeedKind plant starting at 20,000 tonnes a year and expanding to 200,000 [cargill2017]. Those were plans. Its later Calysseo venture with Adisseo built a 20,000 tonne plant in Chongqing that, by Adisseo’s 2025 annual report, still needed technical modification after a longer-than-expected trial period, with production stopped pending upgrades and financing [adisseo2025ar]. In June 2026 Adisseo disclosed USD 80 million invested, unaudited 2025 revenue of USD 0.71 million against a net loss of USD 15.52 million, greater-than-anticipated scale-up difficulties and significant challenges to economic viability in the feed-protein market, and said it would provide no further financial assistance [adisseo2026]. The plant had been built. It was still a development asset, in a market under pressure. That is both failure modes at once, and it says nothing about whether gas fermentation works in principle.

Believer Meats announced a facility with more than USD 123 million of initial investment and planned capacity of at least 12,000 tonnes of cultivated chicken [ncgov2022,gea2024], completed an FDA pre-market consultation in July 2025, and ceased operations in December 2025 [fda2025,watson2025]. A 2024 paper involving its researchers reported cell densities up to 130 million per millilitre and modelled a theoretical 50,000 litre facility at USD 6.20 per pound [pasitka2024], a modelled cost, not a demonstrated one. The shutdown does not invalidate the science, and the science did not establish that the business was viable. Milestones are not commercial readiness, and the capital was committed as if they were.

4.2. The four components did not describe one business

Plenty opened its Compton farm in 2023 describing capacity of up to 4.5 million pounds of leafy greens a year and yields of up to 350 times conventional output per acre [plenty2023]. It entered Chapter 11 in March 2025 and emerged focused on premium strawberries [plenty2025]. The likely reading is that the cost of the production advantage exceeded the value captured from it. Land productivity does not settle the economics when achieving it requires inputs the selling price cannot cover.

Amyris reported $222.3 million of renewable-products revenue in 2022 against $258.7 million cost of products sold and $493.6 million of selling, general and administrative expense, then entered Chapter 11 in August 2023 and exited its consumer brands [amyris2023]. Owning more of the value chain is not free margin. Each downstream function has to be profitable after its own cost.

Vital Farms shows the combination working and then slipping. For 2025 it reported $759.4 million of revenue and $66.3 million of net income from a network of small farms [vital2026fy]. In the first quarter of 2026 gross margin fell from 38.5% to 28.3% and the company recorded a net loss, which management attributed to higher costs and an oversupply of eggs that pushed sales into breaker and wholesale channels at lower prices; capital expenditure guidance was cut by half [vital2026q1]. The brand did not break. The premium is gated by supply discipline, and when production outran the demand that would pay for it, scalability stopped describing the same business as value.

4.3. Counterpoints

Novonesis reported an adjusted EBITDA margin of 37.1% and EUR 770.4 million of free cash flow before acquisitions for 2025 [novonesis2026]. Its products change what the customer can do, which is a route to value beyond supplying a tonne of a substitutable bulk material, and it stops one failed protein project from becoming an argument against industrial biotechnology. Vow builds its cultured quail into foie gras, parfait and smoked spread rather than a drop-in commodity [vow2025], and received permission to sell it in Australia and New Zealand in June 2025 [fsanz2025]. That is a product strategy and a regulatory milestone, not proof that the four components reconcile, but it is the right experiment: treat the product and the customer as variables, and test whether people will pay for what the method makes possible before demanding cost parity.

5. A production system that gets better as it grows

The design brief for a biological facility should say how the important uncertainties will be investigated during operation: which variables can be changed independently, where comparisons can be made, and how a result is validated before wider deployment. Those capabilities belong beside throughput and capital cost in the evaluation, because they set the cost of improving the process over its life.

At Vertical Oceans this is the proposition behind a fleet of individually addressable production units under a shared intelligence. Most of the fleet runs near the current best-known set points to protect output. A designed fraction runs at set points chosen because we are uncertain about them. The system spends a little production to buy information, and each cycle updates the models that set the next cycle’s protocols across the whole fleet [diener2026eoi]. A result is tested on a few units, repeated, and introduced gradually with a way back, so learning travels further than risk. In a fleet of 50 properly isolated units, a failure confined to one exposes only two% of capacity. In a fleet of 800 it is 0.13%.

It is a design commitment, not an exemption from proving commercial performance. I use economies of intelligence for the relationship capacity measures miss: more units mean more opportunities to generate evidence, and an improvement is worth more when there are more units to apply it to; fleet scale is scale efficient. That relationship is conditional and should be measured, not asserted: time to resolve an uncertainty, production exposed by the test, reproducibility, and realised benefit after deployment. The technoeconomic model should direct the effort. A modest gain in survival, recovery, energy or commissioning time can change the viability of the whole business, and learning priorities should follow those sensitivities rather than whichever metric is easiest to move.

Capital commitments should then be explicit about whether they are buying validated capacity or the evidence needed to validate it. Before approving expansion I want to know which assumptions the commitment depends on, what evidence supports them, and how much capital remains if they are wrong. That changes what investors underwrite. Industries with steep learning curves were valued on the demonstrated slope, not the current unit cost [wright1936], and a biological business with a working loop should be able to show its slope: the improvement per cycle, the mechanism behind it, and why the rate holds as the fleet grows. That evidence is worth more than a capacity announcement.

6. Beyond aquaculture

None of this is specific to shrimp. Calysseo and Believer each built one large vessel or plant before the process was characterised. MycoWorks built large rooms. The fleet-of-small-units architecture answers that failure in fermentation and cell culture as directly as in aquaculture, and it is the direction we are extending our own work. Any bioprocess that is far from its optimum and poorly characterised is a candidate.

The productive response to failure in industrial biology is neither retreat nor another round of optimism. We should be able to recognise a scientific achievement inside a failed business, an economic weakness inside an impressive technology, and a viable business whose advantage lies somewhere other than the headline productivity metric. Inside the facility, the architecture has to make biological uncertainty testable and keep mistakes survivable. Across the business, the evidence has to keep reconciling productivity, cost, captured value and viable capacity. Scale is a stock. Learning is a rate. Build the rate first.