The recipe nobody can cook by hand#
Picture a recipe with 49 ingredients, each of which can be added at one of five strengths. It sounds like a fussy soup. In fact it describes a design space of more than 10^34 possible recipes, a figure given in a 2024 review of machine learning for cultivated meat (Todhunter et al., 2024). No laboratory will ever test more than a vanishingly small fraction of them.
That soup is the growth medium, the liquid that feeds animal cells in a tank until they become something you could, in principle, eat. It matters because the medium is where much of the cost of lab-grown meat piles up. The Good Food Institute, a non-profit that supports alternative proteins, calls cell culture medium the largest cost and environmental driver in cultivated meat production (GFI).
So the question for this post is narrow and practical. Can machine learning find cheaper recipes faster than people can, and does the published evidence say it has already done so? The short answer is that the methods work in the lab, the headline cost claims are mostly modelled or company-reported, and the regulators are still some way from a final verdict.
What lab-grown meat is, and why feeding it is so expensive#
Lab-grown meat goes by several names, including cultivated meat and cell-cultured meat. The idea is to take a small sample of cells from an animal, then grow muscle and fat cells in a stainless-steel vessel called a bioreactor rather than raising the whole animal (Todhunter et al., 2024).
Cells in a bioreactor need feeding. A growth medium typically supplies sugars, amino acids (the building blocks of protein), vitamins and salts. It also supplies signalling proteins called growth factors, which tell cells to divide. Insulin, transferrin and fibroblast growth factor (FGF) are examples used in one widely cited cost model (Humbird, 2021). Older recipes leaned on foetal bovine serum, an animal-derived additive, which is expensive, variable from batch to batch and at odds with the point of making meat without slaughter (Nikkhah et al., 2023).
Three technical terms will come up repeatedly, so here they are in plain English.
A techno-economic analysis is a spreadsheet-style model that estimates what it would cost to run a factory, given assumptions about equipment, ingredients and yields. It is a forecast, not a measurement.
Bayesian optimisation is a way of running experiments in rounds. After each round, a statistical model estimates which untested recipes look most promising and which ones the model is least sure about, then picks the next batch accordingly. Think of a chef who tastes after every tweak and uses what they learn, rather than baking a hundred loaves on a fixed grid.
Active learning is the same idea seen from the machine-learning side: the model chooses which experiments would teach it the most.
Why the growth medium is the real bottleneck#
The best-known attempt to price cultivated meat at food scale is a 2021 model by the process engineer David Humbird. In his fed-batch scenario, the cost of wet cell mass came to $37 per kilogram, falling to about $22 per kilogram if amino acids could be bought as a cheap plant-protein hydrolysate at $2 per kilogram. A perfusion system, which continuously swaps old medium for fresh, was costlier at $51 per kilogram. He set an affordability target of about $25 per kilogram of bulk cell mass (Humbird, 2021).
In that model, amino acids dominated the medium bill at roughly $19 per kilogram, while growth factors added only $3 to $4 per kilogram at scale (Humbird, 2021). Other estimates differ sharply. The 2024 review notes figures putting growth factors at more than 95% of total production cost (Todhunter et al., 2024). The gap comes down to assumptions about prices and volumes, which is a useful reminder of how much these forecasts depend on what is fed in.
Why not simply optimise by trial and error? Because the search space defeats it. The same review points out that most work so far uses well-characterised model cell lines, such as mouse muscle cells, rather than cattle, pig or chicken cells, and that much of it is at laboratory scale (Todhunter et al., 2024). Every new cell type may want a different recipe.
Bayesian optimisation: an experimenter that learns from every flask#
The strongest recent demonstration of the approach did not use meat cells at all, and it is worth being clear about that. A team led by J. Christopher Love at MIT published a Bayesian optimisation framework for culture media in Nature Communications in July 2025 (Narayanan et al., 2025). It used a Gaussian process, a statistical model that predicts an outcome along with how uncertain the prediction is, to choose each new round of experiments.
The team tested it on human blood immune cells and on a yeast widely used to manufacture proteins. For the immune cells, four iterations (24 experiments) found a blend of commercial media that kept 75% to 80% of cells alive, against about 60% for the best single medium. For the yeast, 90 experiments over seven iterations lifted output of one test protein 2.5-fold over two benchmark recipes, and the authors estimate this took roughly 2.5 to 3 times fewer experiments than a standard statistical design-of-experiments approach. Reusing earlier knowledge, a technique called transfer learning, doubled the productivity of another protein from 6 to 13 units, and across the paper the authors claim 3 to 30 times fewer experiments than conventional methods (Narayanan et al., 2025).
The caveats are the authors' own. The conventional-method experiment counts are estimates rather than measured runs, the gains varied by protein (one comparison was not statistically significant), and practical factors such as ingredient cost and stability are not yet part of what the algorithm optimises (Narayanan et al., 2025). That last point matters most for meat. A medium that wins on growth but costs a fortune is no use here.
Neural networks, genetic algorithms and the hunt for less serum#
Other groups have aimed more directly at cultivated meat's twin goals of cost and sustainability. A team at Tufts and Virginia Tech used a radial basis function neural network (a pattern-matching model) to predict the cost, global warming potential and growth rate of a reduced-serum medium for zebrafish cells. A genetic algorithm, which mimics evolution by repeatedly mixing and selecting good candidates, then searched across seven ingredients for the best trade-off. The neural networks predicted all three outcomes with a model efficiency of 0.98, though the published abstract does not state how much cheaper the resulting recipe was (Nikkhah et al., 2023).
A Japanese group at the University of Tsukuba took an active-learning route with human HeLa cells, starting from 232 medium combinations covering 29 components. Both modes of their algorithm predicted serum concentrations roughly an order of magnitude below a commercial recipe. The authors also report a wrinkle: the optimised medium raised a metabolic signal used as a proxy for growth without measurably raising the number of living cells, and the work covered a single cell line (Hashizume et al., 2023).
It would be wrong to give AI all the credit. A Tufts team led by David Kaplan reached a serum-free recipe for bovine muscle stem cells, nicknamed Beefy-9, by starting from a medium designed for human stem cells and adding recombinant albumin, with no machine learning involved. The estimated bulk ordering could cut medium cost by about 70% against the serum-containing standard, an estimate made before peer review. At that stage, serum was still needed to plate the cells at the first step of passaging (Stout et al., 2022). Human insight and algorithms are better seen as partners than rivals.
What the cost models say, and what has changed#
Cost models do not agree, and they are not measuring the same thing, but placing them side by side is instructive. The chart below uses figures from three published analyses.
Figure: Modelled costs from Humbird (2021), Negulescu et al. and Pasitka et al. (2024). These are projections for hypothetical factories, not prices anyone is paying. The Pasitka figure is the authors' $6.20 per pound converted at rate of 2.2046 pounds per kilogram.
Negulescu and colleagues modelled a plant making 100 million kilograms of cultured beef a year, about 0.14% of global beef production, and found costs of $30.4 per kilogram in a 42,000-litre stirred tank, $20.8 in a 210,000-litre tank and $13.0 in a 260,000-litre airlift reactor (Negulescu et al.). They still concluded that bigger reactors and cheaper media were both needed before commercialisation (Negulescu et al.).
The most striking medium-cost claim comes from Believer Meats and the Hebrew University of Jerusalem. In Nature Food in 2024, the team reported an animal-component-free medium costing $0.63 per litre, cell densities of up to 130 million cells per millilitre, and a modelled cost of about $6.20 per pound for chicken from a theoretical 50,000-litre perfusion facility (Pasitka et al., 2024). The authors disclose that the corresponding author is a director and shareholder in the company and that five co-authors were employees, and the economics are a projection extrapolated from a 300-litre pilot (Pasitka et al., 2024).
Industry voices go further. A venture capital firm, Lever VC, reports that several companies now run media at $1 per litre or less and reach 60 to 90 grams per litre of cells, which would beat Humbird's projected limits; Humbird disputes the comparison, and the underlying company data is under non-disclosure agreements and unpublished (AgFunderNews, 2025). That is reporting on an investor's claims, so treat it as a hypothesis awaiting independent data rather than a finding.
Missing data, metabolic maps and the commercial reality#
Machine learning is only as good as the data it learns from, and cultivated meat has a data problem. A search of the public GEO database in early 2024 returned 71,327 human datasets for "stem cell" but only 61 for cattle (Todhunter et al., 2024). The review adds that public data relevant to meat species is scarce and often poorly annotated, and that few peer-reviewed machine-learning studies of cultivated meat exist (Todhunter et al., 2024).
One proposed fix is to model the cells' chemistry directly. Genome-scale metabolic models are computer maps of the chemical reactions inside a cell, which can show where a nutrient is wasted or a bottleneck forms. Draft maps exist for salmon, cattle, shrimp and chicken, but they still need experimental validation (GFI).
The commercial picture is mixed. As of February 2026, Singapore had authorised three cultivated products and the United States had cleared five through a joint FDA and USDA review; the EU had granted none (FoodNavigator, 2026). The same report notes that Believer Meats, whose process is described above, shut down in December 2025 (FoodNavigator, 2026). A good cost model, in other words, is not the same as a viable company.
In the UK, the Food Standards Agency has approved cultivated pet food from Meatly, but there is no human-food approval yet (FoodNavigator, 2026). The FSA's September 2026 sandbox report says the two-year programme, which runs from February 2025 to February 2027, has handled 14 applications from nine companies, and that no safety assessments have yet been completed (FSA, 2026). The target is to finish two by the end of the programme, and the time to issue a first request for more information has fallen from 195 days to as few as 46 (FSA, 2026).
Three things are worth watching. One is whether independent labs reproduce the low-cost medium claims. Another is whether AI-driven media design starts to include ingredient price directly in the objective, which the MIT team lists as future work (Narayanan et al., 2025). The third is the FSA's final evaluation, due in March 2027 (FSA, 2026).
Frequently asked questions#
Q: What is lab-grown meat? A: Lab-grown meat, also called cultivated or cell-cultured meat, is animal muscle and fat cells grown in a bioreactor instead of inside a live animal (Todhunter et al., 2024).
Q: How is lab-grown meat made? A: Cells from an animal are placed in a bioreactor and fed a nutrient medium that supplies amino acids, sugars, vitamins and growth factors, then harvested as cell mass (Humbird, 2021).
Q: Why is the growth medium so expensive? A: Medium is the largest cost and environmental driver in cultivated meat, and suitably pure amino acids and recombinant growth factors cost more than feed-grade equivalents (GFI; Humbird, 2021).
Q: What is Bayesian optimisation, in plain English? A: It is a method of running experiments in rounds, where a statistical model uses all earlier results to choose the next recipes most worth testing, so fewer experiments are wasted (Narayanan et al., 2025).
Q: Has AI already made lab-grown meat cheaper? A: Not in a way that independent studies have confirmed for meat cells; AI methods have cut experiment counts in non-meat systems and predicted lower-serum recipes, while the lowest cost claims are modelled or company-reported (Narayanan et al., 2025; Hashizume et al., 2023; Pasitka et al., 2024).
Q: Can I buy lab-grown meat in the UK? A: Not for people yet. Cultivated pet food from Meatly has been approved, while the FSA's sandbox has completed no human-food safety assessments so far (FoodNavigator, 2026; FSA, 2026).
Q: Can AI remove animal serum from the recipe entirely? A: Serum-free media for bovine muscle stem cells have been reported, though still used serum at the first plating step, and AI-predicted low-serum recipes so far come from single cell lines (Stout et al., 2022; Hashizume et al., 2023).
References#
- Todhunter ME, Jubair S, Verma R, Saqe R, Shen K, Duffy B. Artificial intelligence and machine learning applications for cultured meat. Frontiers in Artificial Intelligence 7 (2024). Review article. https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2024.1424012/full
- Narayanan H, Hinckley JA, Barry R, Dang B, Wolffe LA, Atari A, Tseng Y-Y, et al., Love JC. Accelerating cell culture media development using Bayesian optimization-based iterative experimental design. Nature Communications 16, 6055 (2025). https://www.nature.com/articles/s41467-025-61113-5
- Nikkhah A, Rohani A, Zarei M, Kulkarni A, Batarseh FA, Blackstone NT, Ovissipour R. Toward sustainable culture media: Using artificial intelligence to optimize reduced-serum formulations for cultivated meat. Science of the Total Environment 894, 164988 (2023). https://doi.org/10.1016/j.scitotenv.2023.164988
- Hashizume T, Ozawa Y, Ying B-W. Employing active learning in the optimization of culture medium for mammalian cells. npj Systems Biology and Applications 9, 20 (2023). https://doi.org/10.1038/s41540-023-00284-7
- Stout AJ, et al. Simple and effective serum-free medium for sustained expansion of bovine satellite cells for cell cultured meat. Communications Biology (2022). https://doi.org/10.1038/s42003-022-03423-8
- Humbird D. Scale-up economics for cultured meat. Biotechnology and Bioengineering 118, 3239 to 3250 (2021). https://pmc.ncbi.nlm.nih.gov/articles/8362201
- Negulescu PG, Risner D, Spang ES, Sumner D, Block D, Nandi S, McDonald KA. Techno-economic modelling and assessment of cultivated meat: impact of production bioreactor scale. Biotechnology and Bioengineering. https://doi.org/10.1002/bit.28324
- FoodNavigator. Where countries stand on cultivated meat regulation in 2026. 18 February 2026. Science journalism, cited because no single primary source covers all jurisdictions. https://www.foodnavigator.com/Article/2026/02/18/cultivated-meat-regulation-where-countries-stand-in-2026/
- Pasitka L, Wissotsky G, Ayyash M, Yarza N, Rosoff G, Kaminker R, Nahmias Y. Empirical economic analysis shows cost-effective continuous manufacturing of cultivated chicken using animal-free medium. Nature Food 5, 693 to 702 (2024). Authors include employees and a shareholder of Believer Meats. https://doi.org/10.1038/s43016-024-01022-w
- Good Food Institute. Mapping animal cell metabolism to optimize media formulation. Non-profit organisation page. https://gfi.org/solutions/mapping-animal-cell-metabolism/
- Watson E. Humbird was spectacularly wrong on cultivated meat economics, says report. AgFunderNews, 10 April 2025. Secondary reporting of a Lever VC report; company data unpublished and not peer reviewed. https://agfundernews.com/humbird-was-spectacularly-wrong-o
- Food Standards Agency. Regulatory sandbox report for cell-cultivated products: FSA Business Committee, September 2026. GOV.UK. https://www.gov.uk/government/publications/food-standards-agency-business-committee-meeting-september-2026/regulatory-sandbox-report-for-cell-cultivated-products-fsa-business-committee
