In brief
Small teams (and AI tools) each build their own cost model of the same target, the 2036 cost per kg, conditional on one shared set of input parameters. Comparing the models shows how much of the experts' disagreement comes from the input values and how much from how each model is put together.
On this page: Why · Cruxes · Plan · Toy simulator · Our current model · New to this?
Why a modeling hack ¶
Workshop participants disagree a lot about what cultivated meat will cost. On the 2036 cost-per-kg question (CM_01CM_01, as asked at the May workshop: the 2036 production cost per kg of wet cultivated-chicken cell biomass, across large-scale plants. The wording was revised on June 25 and again in September; the link shows the history.), workshop medians run from roughly $1 to $100 per kg of wet cell biomass (US dollars, large-scale plants), and the later independent round goes up to $500/kg. That suggests real disagreement, though some of the spread may have other causesAs of the July 20 public review: 16 human submissions and 14 visible point estimates for CM_01. Thirteen were aligned closely enough to pool; one anonymized $20/kg written estimate used a different output basis and gave no interval, so it is shown but not pooled. The sample is selected, several intervals are missing, and respondents may have read the conditioning and output basis differently..
The hack has two goals:
- Independent models. Small teams of modelers, and AI tools, each build their own model of the 2036 cost per kg. Comparing them shows whether the spread comes mainly from the input values (parameters) or from the way each model is put together (structure).
- Bringing in technical experts. People from GFI, ACIB, Ivy Farm and elsewhere, plus forecasters and informed skeptics, shape the input values and the model structure, mainly by giving numbers with ranges for the underlying drivers and saying which process choices those numbers depend on.
How this fits with the rest of the project
- Workshop summary: qualitative evidence and arguments.
- Public beliefs analysis: consent-checked named and anonymous distributions, subgroup views, and sensitivity to pooling choices.
- Crux map: the main disagreements, with questions for skeptics and optimists.
- The Unjournal team's current cost model: turns explicit process assumptions into a distribution of factory-gate cost per kg of cell mass.
- Cost → demand bridge (set aside for now): an illustrative link from cost to demand. It has had less attention than the cost model, and the hack will not focus on it.
Evidence updates since the May workshop · last updated July 20, 2026
- GFI's 2026 industry report records both progress and headwinds: 20,000-L production was reported in 2025, while disclosed sector funding fell to $73.9m from $144m in 2024. Models should treat scale and finance as separate variables.
- The FDA's February 2026 inventory lists five completed cultured-cell food consultations, including three completed in 2025. That is regulatory progress, but it tells us little about cost or throughput.
- A 2025 porcine-cell study found selected plant hydrolysates reduced serum use by up to 70%, while Pasitka's chicken process used $0.63/L animal-component-free medium. These are useful feasibility points; they don't show that full substitution works across cell lines and large reactors.
- Pablo AMC's July response reframes a central disagreement: a frontier “reasonably achievable” plant is not the same estimand as the average qualifying commercial plant asked about in CM_01. The hack should report these two separately.
Where the estimates diverge ¶
Each of these is a modeling choice that can shift the projected cost several-fold. The crux map covers them in more depth, with evidence and questions for each side.
1. Growth-factor independence
Will cell lines engineered to make their own growth factors work at commercial scale by about 2030–33? If so, the growth-factor cost line largely disappears (some respondents put it at $0/kg by 2036). Performance, stability, safety and the regulatory path are still unproven at that scale.
2. Hydrolysates as base media
Can cheap plant or yeast hydrolysates fully replace purified amino acids, or only partlyFull replacement: FEASTS, Fuchs. Partial: Swartz (~20%), Bomkamp (~40%).? Part of this is definitional: CM_12The workshop belief question on hydrolysates. Its wording mixed up replacing growth factors with replacing the amino-acid base. mixed up two different substitutions.
3. What the model is built on
Grams per litre, cells per mL, or mass productivity (kg per m³ per day)? The choice ties cell density to process mode (fed-batch, perfusion or continuous), so the metric is itself a crux.
4. Medium use per kg
Litres of fresh medium per kg of biomass (the ~20 L/kg industry figure) combine density, perfusion, recycling and losses. Two reviewers argue it depends mainly on nutrient yield rather than density. Any “feed conversion” figure needs a stated basisMass, protein/amino acids, or energy. A 3:1 calorie ratio, kg nutrient per kg biomass, and litres of medium per kg biomass are dimensionally different and can't be compared directly..
5. Capital and the move to dedicated plants
Can companies finance the move from CDMOsContract development and manufacturing organizations: shared facilities that produce for several companies. to their own plants after the post-2024 funding drop? Capital costs are less certain than ingredient costs, and company survival may bind before the technical cost floor does.
6. What product are we costing?
Products will not be made entirely of cultivated cells, and a low-inclusion hybrid and a whole cut are different products. There is no consensus inclusion rate (GOOD Meat 3 was 3% cells). For comparisons the team's model now uses a default of 25% cells plus plant-based ingredients; the estimate itself still stops at factory-gate biomass.
Lower-tension calibration checks: food-grade vs pharma-grade media (CM_17The workshop belief question on how much of the medium can be food-grade rather than pharma-grade. Responses were close to consensus., near consensus) and growth-factor cost-reduction pathways (E6The belief item on the probabilities of different growth-factor cost-reduction pathways. Answers diverge sharply, which makes it a clean target for structured elicitation., sharply divergent).
The plan: a minimum viable hack ¶
Rather than open-ended model building, start with a small pilot: one shared set of inputs and two or three independently built models.
-
Agree exactly what we are estimating. Pin down the target quantity and its units: best available plant or industry average, which dollar year, cost per kg of what (wet biomass, dry cells, protein), and whether the estimate assumes the industry succeeds.
Why: these choices alone can move estimates several-fold, so models can't be compared until they are fixed.
What was decided on October 26
For the team's cost model: unconditional and hypothetical (a large, nth-of-a-kind plant built with the technology available in 2036, whether or not such plants exist), best available plant rather than industry average, 2026 US dollars, no retail-price modeling. Output is per kg of washed cell biomass at a fixed solids content close to chicken meat (placeholder 25–30% dry matter, to be agreed), plus per kg of protein; the dashboard reports cost per kg of dry cell mass, with the cost per kg of a hybrid product (25% cells at that solids content, 75% plant-based ingredients) alongside. CM_01 A1 shares the unconditional framing but asks for a production-weighted average in 2025 dollars; see the estimand note. Still open: recovery yield and geography.
- Publish a reference model. Fix one version of our cost model that every team starts from: a single sheet listing every input value and range, the exact code version, a fixed random seed (so anyone re-running it gets identical numbers), and results saved in a machine-readable file. Record its default result: the cost distribution with every setting at its default, and how much each cost component contributes. Why: when a team's model gives a different answer, we need a fixed point to compare against, and a way to tell a deliberate change from an accidental one. Automated regression testsChecks that re-run the model after every change and flag any shift in its results that wasn't intended. catch the accidental ones. Partly in place: the dashboard's engine has automated tests and an audit script that records default results and the code version (repository).
- Pick two tracks. Focus the pilot on two clusters of cruxes and leave the rest for later. Recommended first track: medium use, nutrient yield and productivity (cruxes 3–4). The second track is still to be chosen*An earlier draft proposed a factory-gate-to-retail value-chain bridge. That is set aside, since our cost model now stops at production cost.. Why: a narrow scope is more likely to produce finished, reviewable models.
- Build independently. Two or three small teams get the same input sheet but don't see each other's model structure until their results are frozen. Process experts own the input sheet; modelers each build a structure that uses it. Why: this is the test of structure versus parameters. With identical inputs, any spread between the teams' results comes from modeling choices: which cost components exist, how they combine, what is bundled or conditional. The published cost studies differ mostly in this way (Humbird ~$21/kg, CE Delft ~$6–8, Risner $44,500+).
- Compare and decide. Overlay the teams' cost distributions on one chart, trace each difference to structure or inputs, check units and basis, and merge into the shared model only changes that have a source and a named maintainer. Why: the output we want is a record of where experts and modelers converged, and where disagreement can't be resolved with current evidence.
Go/no-go: schedule the live event only after steps 1–2 are complete and at least two modeling leads plus one domain lead have agreed to produce reviewable work. Otherwise run it as an asynchronous model-review sprint, which is more likely to produce lasting improvements than a loosely scoped hackathon.
What we are aiming to produce
- An updated, version-controlled cost model in which each contested assumption is an explicit, documented parameter with a credible interval.
- An updated crux map linking each modeling disagreement to belief responses and evidence.
- At least one new sub-model for an under-modeled driver (e.g. nutrient yield, mass productivity, or product inclusion rate).
- A record of where experts and modelers converged, and where the disagreement can't be resolved with current evidence.
Ideas under discussion ¶
Added October 26, 2026. AI-drafted (Claude) from David Reinstein's design notes; not yet reviewed by a person.
Elicit fundamentals, not derived quantities
Overall cost per kg and litres of medium per kg are derived quantities: two people can give very different numbers because of assumptions neither has stated. Instead, elicit beliefs on six to eight primitive drivers on which experts can genuinely differ, possibly conditional on process mode, and compute costs from them:
- process mode (fed-batch, perfusion, continuous) and its feasibility at scale;
- achievable cell density at harvest, on a stated basis (wet weight, dry weight, or cell count);
- nutrient yield: kilograms of cells produced per litre of medium consumed;
- growth-factor requirement per kg of cells;
- input prices: basal medium per litre and growth factors per gram;
- bioreactor cost per cubic metre of working volume;
- optionally, plant scale and utilization, and cost of capital.
Also ask each respondent for one or two derived numbers (for example, their overall cost per kg) to check against the cost their primitive answers imply. The model now has a yield-based option for medium use, so both views in crux 4 can be run. Details: model documentation.
A devil's-advocate panel by shuttle diplomacy
Instead of starting with a joint session where skeptics and proponents talk past each other, run the exchange in one-on-one rounds:
- Collect each skeptic's exact objections, in their own words and as specifically as possible (which number, which step, what evidence would change their mind).
- Take them to a proponent for evidence-based answers.
- Return the answers to the skeptic and record which objections are resolved, which are narrowed, and which stand.
- Only then convene the group, on what remains unresolved.
This keeps the live discussion on the real cruxes and gives each side time to check evidence. Skeptic objections are already being collected on the team's internal claims page.
Longer term: describe a model, see it rendered back
The aim is an interface where you describe a model however you like (prose, a sketch, a spreadsheet) and see it rendered back as a diagram, equations and a plain-language summary, with AI-generated clarifying questions to pin down what you meant. Each model would end up in a common, comparable form that its author has checked, so independent models can be compared side by side. AI-assisted Squiggle is one option; the describe → represent → check loop matters more than the language.
A quick exercise: toy cost simulator ¶
A fun, quick way to see how uncertain inputs add up. It is a toy, much simpler than The Unjournal team's current cost model, and its presets are older hand-set ranges, so don't read the results as estimates. Enter a low, most likely and high value for each cost line, in $/kg of dry cell mass; the page draws 5,000 random samples and adds them up. It varies numbers inside one fixed structure, so it shows parameter uncertainty only (step 4 is about structure).
Per kg of hybrid product (25% cells)Products will not be made entirely of cultivated cells. This is the ingredient cost of a kg of product with 25% cultivated cells, counted at chicken meat's water content (27.5% dry matter), and 75% plant-based ingredients at $3/kg: dry cost × 0.275 × 0.25 + $2.25. There is no consensus inclusion rate (crux 6); 25% is the convention the team's model uses. Forming, packaging and margins are excluded.: P10 — · P50 — · P90 —
Runs entirely in your browser; nothing is sent anywhere. How the sampling worksEach line is sampled from a triangular(low, mode, high) distribution, clamped so low ≤ mode ≤ high and all ≥ 0, and the draws are summed per run. Triangular is used for transparency, not because it is the right shape; the full models use lognormals and betas with explicit p5/p95.
Where the preset ranges come from
They were set by hand from the older Squiggle model (v0.3) and an earlier version of the dashboard, per kg of wet cell biomass. Here they are converted to dry cell mass at 20% solids (wet values × 5, the conversion the current model uses). They have not been refreshed against the current model; for its numbers, use the dashboard.
Each line is entered directly in $/kg, so the lines add up without the intermediate $/L × L/kg step. In the full model, annualized capital is derived from working volume, cost of capital and plant cost multipliers. Contract manufacturing, bundled media, downstream processing and process modes are left out here.
Our current cost model ¶
The Unjournal team's current model goes well beyond the toy: process-mode mixtures, a latent industry-maturity factor linking technology adoption and financing costs, capital cost derived from required reactor volume, and separate regimes for cheap and expensive growth factors. It reports cost per kg of dry cell mass, with the cost of a hybrid product (25% cells) alongside.
Before relying on it, read its model limits & critique page. An older Squiggle version (v0.3) is no longer maintained.
New to this? ¶
You don't need to know Monte Carlo modeling, model code or forecasting to take part. The hack mixes domain experts who may never have built a model, modelers who may not know the biology, and skeptics who want to test both. You can describe a model in words, a sketch or a spreadsheet and use AI tools to turn it into something runnable, and we will run short onboarding sessions.
Four short primers
Monte Carlo basics
What sampling uncertain inputs thousands of times and adding them up does, and why uncertainty compounds. The simulator above is a hands-on example.
Session: a 20-minute walk-through of the simulator and the dashboard's overview.
Describe a model; let AI build it
Say what the cost components are and how they combine, and an AI tool (Squiggle AI, or a general coding assistant for Python or a spreadsheet) turns it into a model you can run and check.
Session: describe one cost structure aloud and watch it become a model.
Techno-economic cost modeling
How a $/kg figure is built from medium volume, cell density, reactor cost, capital annualization and fixed operating costs, and which assumptions move it most.
Start: the methodology and parameter reference and the comparison of published studies.
Calibrated forecasting
How to give a 90% interval you would actually bet on: wide enough to be right about 90% of the time, narrow enough to be useful. Most people are overconfident, and practice helps.
Start: CM_01 on Metaculus · Quantified Intuitions calibration · Calibrate Your Judgment.
Want a particular tutorial, or to run one? Email contact@unjournal.org.