For cultured meat to move the needle on climate, a sequence of as-yet-unforeseen breakthroughs will still be necessary. We’ll need to train cells to behave in ways that no cells have behaved before. We’ll need to engineer bioreactors that defy widely accepted principles of chemistry and physics. We’ll need to build an entirely new nutrient supply chain using sustainable agricultural practices, inventing forms of bulk amino acid production that are cheap, precise, and safe. Investors will need to care less about money. Germs will have to more or less behave. It will be work worthy of many Nobel prizes—certainly for science, possibly for peace. And this expensive, fragile, infinitely complex puzzle will need to come together in the next 10 years.
On the other hand, none of that could happen.
That is the takeaway from a new article by Joe Fassler (2021) in The Counter that draws heavily on two techno-economic analyses (TEA) of cultured meat (CE Delft 2021 & Humbird 2020). For full disclosure, we at Rethink Priorities were independently reviewing these TEAs (plus a third by Risner, et al. 2020) and in the process of writing a summary and comparison of them with our main takeaways.
The article addresses many of the issues we also noticed, and supplements them with interviews from industry experts. Here we want to acknowledge that they beat us to the punch somewhat, add a few relevant things we think the article left out from the comparison, and what the next steps are in our project.
The main cruxes of disagreement across the TEAs are:
Approach to the research question
Investor payback timelines
Food grade versus pharmaceutical grade bioreactors
The costs of media (growth factors and amino acids) at scale
The limits of cell-engineering needed to reduce media consumption needs
First though, we provide our quick summaries of the three TEAs so readers have a background before diving into the comparisons.
As we are investigating a scientific question that sometimes hinges on deep technical expertise (which neither Neil nor I have), we will likely have some errors in the summaries and (especially) personal takeaways. In addition, this report is less thoroughly checked than usual for Rethink Priorities reports. It should best be viewed as our current tentative understanding of the existing literature, rather than a final definitive summary of the existing literature.
tl;dr: We reviewed 3 TEAs on cultured meat. Our summary is that Humbird is very high quality and suggests cultured meat cost-competitiveness is hard and needs everything to go right. CE Delft outlines some of what will need to go right, but doesn’t provide much evidence that any of it is possible, has internal validity errors, and arguably has too much motivated reasoning. Risner, et al. is decent, within the narrow limits it sets itself, but too many details are underespecified for it to reflect the full costs and challenges of scaling up cultured meat. Reading the TEAs and doing surrounding research has turned Linch from a cultured meat optimist to being broadly pessimistic. Neil wants to be more agnostic until further research from Rethink Priorities and others.
Our TEA Summaries
As part of a project forecasting the potential for cost-competitive cultured meat to displace conventional meat, my colleague Neil and I investigated three techno-economic analyses (TEAs) estimating the economic feasibility of cost-competitive cultured meat ($2.50-$8/kg, akin to different estimates for existing wholesale meat prices).
A quick note on terms. The studies we looked at (and us) are only investigating “cultured meat”, that is, animal cell based meat of target conventional meats (usually beef). They do not investigate other technological alternatives to conventional meats such as plant*-based meats (like Impossible Foods or Beyond Meat), fungi-based meats (like Quorn), or brain-dead/genetically happier chickens (no known products).
In addition, the studies only look at the production side of one intermediate product: that of undifferentiated animal cell slurry (70-77% water). They do not investigate the question of scaffolding (converting cell slurry into structured meat products, eg. steak or chicken wings), nor the marketing and consumer acceptance questions of whether consumers will wish to buy cultured meat if it’s widely available and cheap.
To give a quick sense of the difference between the TEAs, here are the range of prices estimated in the TEAs given various assumptions, though note the lowest estimates from Humbird (2020) and Risner, et al (2020) are not regarded as realistic by the authors.
Humbird (2020) presents a study (funded by Open Philanthropy) which argues that a number of very difficult scientific and engineering problems need to be solved to get cultured meat to less than $25/kg (the reference point of the more expensive end of plant-based meats), at least with production processes anywhere similar to known processes.
Humbird models an ecosystem producing 100 kilotons/year of warm-blooded mammal cell mass. He believes that after a number of difficult engineering problems are solved to get to viable cultured meat at scale, the cost of cultured meat is set by expensive capital expenditures and media costs, particularly amino acids. Ultimately, Humbird believes that significant scientific and engineering progress is necessary but insufficient for cultured meat to displace conventional meat.
The core argument is that while microbial processes (for yeast, bacteria) have scaled over decades to extremely large production volumes (and create a lower bound for attainable prices given known limits), producing animal cells at industry scale present very large challenges in comparison:
Animal cells have a slower doubling time (24h-48h) than microbes (20min-3h)
Practical bioreactor volume is smaller, as animal cells are less tolerant of both spatial heterogeneity (eg temperature, pH) and fixes to heterogeneity (sparging, agitation)
Oxygen uptake/transfer rate and catabolite inhibition (lack of cell walls means harder to control osmolality) creates a limit to the practical final cell density
Animal cells need to be fed (expensive and globally low supply of) high-quality media components, particularly near-pure amino acids.
There are significant capital costs to maintain very high hygiene (“aseptic”) conditions, as any viral/bacterial contamination in an animal cell bioreactor is fatal.
Significant genetic engineering is necessary to produce animal cells with the desirable properties for bioreactor use (like breaking the Hayflick limit and achieving anchorage independence)
Humbird first models a Techno-Economic Analysis (TEA) where all parameters are similar to baker’s yeast, except for constraints imposed by the first three challenges (doubling time=24 h, maximum fermentor size=20 m3, maximum oxygen uptake rate=25 mol O2/m3-h).
The analysis comes to $3.87/kg of wet (70% water) cell mass for the constrained yeast process. As yeast production at scale is already a highly optimized process over many decades, and the additional constraints mentioned so far are pretty close to the fundamental biological nature of animal cells, it seems unlikely that we can do better than a lower bound of $3.87/kg. Unfortunately, there are other constraints.
For the rest of the study, Humbird then considers what steps are needed for a more realistic model that incorporates all the real constraints of cellular meat, but also at prices and quantities with a large ecosystem of producers and suppliers (commensurate with a production of 100 kilotons/year), and then does sensitivity analysis on the relevant prices.
The general upshot:
Significant metabolic engineering is necessary so enhanced cells have metabolisms at close to what Humbird believes are theoretical limits
Cheap at scale (<$200/kg) cultured meat with “wild-type” cells is simply not feasible
The cost of cultured meat is then bounded by the costs of high-quality media components, particularly near-pure amino acids, which Humbird models to be very expensive even at scale.
Humbird speculatively considers using soy hydrolysate instead of individual amino acids in media, which will be significantly cheaper if viable.
Secondarily, capital costs (especially but not limited to the cost of aseptic animal cell bioreactors) are also a large cost.
In the most optimistic scenario Humbird considered (cells with enhanced metabolisms designed, soy hydrolysate absorption, all other technical problems solved), Humbird estimates a cost of $21/kg wet cell mass.
Note that this does not include flavoring or structuring cell slurry into meat, nor does it include the amortized costs of R&D.
(The most optimistic scenario is column 2)
Humbird gives five key recommendations for future study. Note that in his model, solving these problems are necessary but insufficient for getting cultured meat to $21/kg.
Vergeer, et. al (2021) presents a study, using both public and private data, to argue that while substantial cost reductions are necessary, they are feasible to bring cultured meat to sufficiently low cost such that we can have a large burgeoning industry of cultured (“cultivated” in their words) meat by 2030. This study is conducted from the consultancy CE Delft and funded by the Good Food Institute.
The core argument goes: Existing models of cultured meat via current production technology and costs for inputs will result in very expensive cultured meat. However, based on their analysis of sometimes private data from cultured meat companies, new changes just around the horizon across a wide range of avenues should be enough to cut cultured meat to fairly affordable prices, perhaps close to conventional meat baseline by 2030.
The authors model an ecosystem producing 10 kilotons/year of wet (70-77%) animal cell mass, with some intentional vagueness about which animal types they ultimately refer to. They gathered data from 16 different (cultured meat and other) companies to do this analysis. They assume food bioreactors rather than traditional animal cell bioreactors as the baseline capital cost.
CE Delft first lightly models three “baseline scenarios” for the expected cost of a kilogram of cultured meat, using different ranges of ingredient costs and amounts required (high is upper end, mid is geometric means of all estimates, low is lower end for everything).
They then consider six scenarios for future cost reduction:
Note that scenario 4 onwards builds from scenario 3 (low end of all estimates) rather than scenario 2 (mid-range of estimates). This modeling choice is not explained.
Each scenario builds on top of the previous scenario.
Scenario 4: Lower prices for growth factors
“As in other literature, we adopt the scenario that a price reduction of a factor > 1,000 may be possible. “
Scenario 5: Cheaper or lower usage of recombinant proteins
“We have credible industry sources that point towards a feasible cost reduction of a factor 100.”
Scenario 6: Capital expenditures can be repaid at a slower rate than normal
While traditional investors require a payback time of <4 years for commercial investment projects, CE Delft argues that socially motivated investors can instead opt for a payback time of 30 years.
Linch (For judgment calls that one of us looked into in some detail, and the other broadly agrees with but is not vetted, we specify author names in Bold so it’s clearer that we’re referring to just one author’s opinion): It is unclear to me whether a 30 year payback time, even if true, is correctly pricing in what may presumably be higher maintenance costs, for 30+ year old equipment using novel technology.
Scenario 7: Increase in maximum cell density during proliferation
“50*106 cells/ml is the baseline cell density modeled. According to experts in the field, it may be feasible to increase cell densities by a factor 4 (to 20*107 cells/ml)”
“Higher cell densities would mean that more meat cells can be grown in a reactor of the same volume”
Though note that a factor of 4 increase in cell density means that the percentage of bioreactor volume that is now cells would increase from 17.5% to 70%, very close to sphere packing density limits.
Scenario 8: Shortening of production run time
“It is estimated that a reduction in production run time of 25% is feasible.”
Scenario 9: Increase in cell volume
“As the companies involved in this study produce a range of species and cell types, we used an average cell volume for the baseline scenario and determined potential larger cell volume (5,000 µm3 ) on primary data collected and literature.”
However, note that combining scenario 7 with scenario 9 translates to a pure cell density (% of bioreactor volume that is cells) of 100%.
Risner, et al. (2020)
Executive Summary Risner, et al. (2020) present a study (funded by the Innovation Institute for Food and Health at UC Davis) on the costs of a 20m3 (20,000 L) food-grade core bioreactor based-system producing 121 kilotons of bovine cell wet mass (1% of current US beef production). The authors argue that “technological performance will need to approach [or surpass] known technical limits for animal cell based meat to achieve profitably as a commodity”. They think cultured meat “may not be economically viable as a commodity for some time” (undefined how long), but “may enter the market place sooner as a minor ingredient which lends desirable organoleptic qualities to an otherwise plant-based products” or become more like luxury food items; almas beluga caviar (US$10,000/kg), Atlantic bluefin tuna (US$6,500/kg), and foie gras (US$1,232/kg).
The study employs six global sensitivity analysis algorithms of 67 parameters, chosen using cellular biology and chemical/process engineering conventions. The list is narrowed down to the nine factors that most influenced capital and annual operating expenses by consolidating the top five parameters across all six algorithms:
Average single cell volume
Average single cell density
Fibroblast growth factor-2 (FGF-2) concentration
FGF-2 cost
Glucose concentration
Glucose consumption rate per cell
Maturation time
Transforming growth factor beta TGF-b (a growth factor) concentration
Oxygen consumption per cell
In the paper itself the authors don’t spend much time discussing TGF-b or oxygen consumption per cell, despite TGF-b being a major cost variable and oxygen uptake rates being a major obstacle in Humbird’s TEA. Their interactive cost calculator illustrates that media/recombinant protein costs also substantially affect cost (especially transferrin, L-Ascorbic acid 2-phosphate, and insulin). Furthermore, they also cite bioreactor capital costs as a significant factor in costs, even after assuming food-grade bioreactors.
These nine factors were then clustered into technological components. Four technology development scenarios are then developed, based on altering the values of some of these nine variables (see table below). The four technology development scenarios are:
Scenario 1 US$437,000 per kg: baseline scenario based on existing animal cell based meat production, in a food-grade bioreactor, including 2019 pricing (obtained from a GFI’s Specht (2019) for animal serum-free media and growth factors, and reported human embryonic stem cells’ glucose uptake rates.
Scenario 2 US$57,200 per kg:halves the cost of growth factors, reduces cell doubling time from 24hrs to 16hrs, halves glucose consumption rate per cell, increases total cell volume density 5x—based on 2x108 cells/ml average max cell density, which the authors note is a physical limit and the number in reality would be much lower.
Scenario 3 US$44,500 per kg: Same as scenario 2 but eliminates the cost of growth factors.
Scenario 4 US$1.95 per kg: An intentionally pie-in-the-sky scenario was designed to see what it would take to reach ~$2/kg (for conventional beef price parity). Nearly all technical challenges are resolved, including reducing growth factor costs to zero, increasing myoblasts/myosatellite cells (MSC) tolerance to glucose concentrations, glucose/media consumption is reduced by an order of magnitude, halving again myoblast/MSC doubling time and lower maturation time 6x, reducing media/recombinant proteins costs (base media containing over 50 ingredients, TGF-b, transferrin, L-Ascorbic acid 2-phosphate, and Insulin) to zero, and increasing total volume density to 100% (which is biologically implausible but unlike CE Delft (2021), Risner, et al (2020) acknowledge this).
Note that this TEA modelled only the core bioreactor system costs, as if every other part of the process was free, so is intended as a minimum cost.
The key three recommendations of the paper in order to lower costs are:
cell selection/engineering to lower the media consumption rate via a more glucose-efficient metabolism
“Engineering and/or screening for cell lines which shift rapidly from a Warburg metabolism (aerobic glycolysis) to a more glucose-efficient metabolism (oxidative phosphorylation i.e. production of 2 ATP vs. a theoretical 38 ATP per glucose molecule)”: from 4.13*10-13 moh/h cell glucose consumption rate per cell to 4.13*10-14 moh/h cell. Note that their argument is less about glucose specifically, which after all is just a simple sugar, and more about using it as a proxy for total media consumption requirements and thus a measure of cell-engineering.
reducing or eliminating the cost of growth factors (especially Fibroblast growth factor-2)
They suggest one way to eliminate FGF-2 costs is by “leveraging the ability of cancer cells to increase glucose uptake rates and exhibit cell proliferation without the presence of growth factors, or cell lines could be engineered/identified to express oncogenes related to these traits.” However, they note this “could be challenging from both a regulatory perspective as well as for consumer acceptance.”
scaling up of perfusion bioreactors.
Doing so allows for multiple changes of media per batch and higher cell concentrations. However, the TEA’s costs are based on estimates for standard food-grade bioreactors and Risner, et al. (2020) acknowledge more sophisticated bioreactors (i.e. single-use or novel perfusion systems) may substantially increase capital costs.
An implicit fourth recommendation is also to substantially lower the costs of the media/amino acids used.
“The amino acids being a big variable in the cost calculation assumes that there has been a solution for the other expensive components like growth factors” (private correspondence) . . .“Our model did not account for amino acid uptake rates due to glucose being the most consumed nutrient in cell culture, however amino acid (AA) metabolism should be a consideration for commercial scale up” (supplementary material, p2)
Why do the TEAs disagree?
Overall approach of the TEAs
Humbird asks, if cultured meat scaled to a global volume of 100 kilotons per year, what would it cost? CE Delft imagines a world where 10,000 metric tons per year of cost-competitive cultured meat can be produced, and then reverse engineers what would need to change for that to happen (unrelated to how likely those changes are). Risner, et al. (2020) follow a similar model to Humbird and then explore what changes to the major cost factors would be needed to get to price parity (acknowledging that they are unlikely).
Our takeaways on the TEAs overall:
Linch: My general impression is that Humbird’s analysis was very carefully and rigorously done (though I wish his uncertainty was sometimes made more numeric, as from an EA perspective many of these investments are worth doing even if there’s only a 5-20% chance of cultured meat working out in a thirty year time horizon). In contrast, my impression is that both the CE Delft report and its subsequent heavy publicity can best be viewed as a textbook exercise in motivated reasoning (akin to this, but less subtle). Various decision points and errors in the report shade in favor of cultured meat being cheap/quickly viable, likely without corresponding errors in the opposite direction. I have not read Risner et.al as carefully as Neil and feel less qualified to opine on it. My understanding is that Risner concludes cost-competitive cultured meat is impossible at current technology, but has an explicit “this is where a miracle happens” final step where you get very cheap prices if you assume away all technical and engineering difficulties. Note that I do not have prior experience in either this field or adjacent fields, and am biasing towards overexplaining and communicating my current opinions instead of only saying things I’m certain is true.
Neil: To me the major distinction between the TEAs is what reference class they are using for technological breakthroughs. Humbird (2020) and Risner, et al. (2020) seem to imagine the correct reference class is one with real technological and cost limits, like biofuels, while CE Delft’s (2021) view is closer to reference classes where the technical challenges can be overcome with enough effort/support from researchers, investors, and government, like photovoltaic technology (solar panels). None of the TEAs provide a useful way to adjudicate if it’s possible to beat the amino acid prices the yeast industry has achieved or to design bioreactor systems that can reduce the costs of contamination at scale, but I think CE Delft (2021) undersells just how much effort is needed. I also found it concerning for the CE Delft (2021) TEA that Linch was able to spot a fundamental mathematical/conceptual error without any prior domain expertise. These are the only TEAs I’ve read so I don’t have a good sense where they fall on the quality scale of TEAs in general. The TEAs should just be one (important) part of your model on how likely cultured meat is to scale.
Investor timelines
In brief, Humbird thinks it would take at least 60 years for the required ten orders of magnitude manufacturing cost reductions to be achieved via normal economies of scale, and investors work on much shorter timelines with higher return expectations for the first successful plant. CE Delft argues cultured meat investors are a special socially motivated breed that will wait 30 years for their returns.
Humbird (2020) simply doesn’t believe a scale up can be achieved in any reasonable investor timeline. He argues that a scale up to conventional meat volume, on the order of 100M metric tonnes/year, requires a ten orders of magnitude increase from the production volume of edible cultured animal cells at the time- on the order of 1–10 kg/y. He argues (2020, p9-10) it took 60 years for “certain aspects of computing to have progressed by about ten orders of magnitude”, and “generally accepted economies of scale in the process industries do not hold within them multiple orders of magnitude in manufacturing cost reduction” . Therefore, “to fit this many powers of ten into a time scale that meets contemporary investor expectations, some would suggest that even more accelerationist “laws” are in play for cultured meat”. Humbird (2020, p11) argues cultivated meat optimists are following a reasoning of “CAPEX only matters for the first plant; all subsequent plants will be financed at a low cost of capital” but that “such treatment from lenders is, more accurately, only enjoyed after the first successful plant. This need to reconcile a sense of urgency with a reluctance to be “first” was an exceedingly common source of friction in the days of venture-backed biofuels startups”.
Humbird (2020, p10) also seems concerned that limits to multiple orders of magnitude progress is especially true “in biotechnology, where the equipment in question is much smaller and more specialized and where the precursor chemicals are only produced in limited quantities”. This is based on Humbird’s conclusion that pharma-grade safety standards will always be required and amino acid costs will be high even at scale, both of which seem more malleable to us than Humbird claims.
CE Delft (2021, p21) writes “typically, the hurdle for commercial investment projects to get funded is payback time < 4 year” but argue socially motivated investors are willing to wait longer for returns. The authors think the payback time should be 30 years. This drops the price from $17/kg to $7.74/kg.
However, it’s unclear why 30 years is actually justified or why socially motivated investors would prefer much longer payback time relative to normal investment that will get you more resources to use for donation or social investment later.
Linch: I also think that CE Delft’s (2021) assumption that maintenance costs are a flat 5% of costs is suspicious/insufficiently defended.
Linch: while it’s possible I’m misunderstanding, a huge related issue with having a capital expenditure investment that takes >30 years to pay off is that you are explicitly baking in not changing capital use (much) which means you’re planning to keep old equipment, which seems especially wild when “old equipment” is decomposed into specific items like “bioreactors in a new field like cultured meat.” So I think that the CE Delft assumption about socially motivated investor attitudes is both by itself dubious but also their conclusion is invalid given just that assumption.
Food versus pharma grade bioreactors
Humbird (2020) thinks that cell cultures cannot be manufactured at large scale using food-grade standards because cell contamination will destroy the product. Therefore, cultured meat is limited to smaller scale pharma-grade standards with virtually no contamination, which are very costly. CE Delft (2021) imagines a world where it will be possible for cell cultures to be produced in food-grade bioreactors. Risner, et al. (2020) allow for the “optimistic assumption” that advances in science will enable production using food-grade bioreactor systems, and that “operational issues related to bioreactor sanitation and fill rates are negligible”.
For Risner, et al. (2020) “cost estimates of food-grade bioreactors were calculated using a method which accounts for equipment scaling, installation, and inflation. This method applies a set unit cost of $50,000/m-3 for a food grade bioreactor and a common scaling factor of 0.6”, and a capital cost of $778K for a 20K L bioreactor. CE Delft (2012) rely on more perfusion bioreactors in their model at $600K a piece, but their 10K L stirred-tank bioreactors are suggested to cost $325K a piece. Humbird’s pharma-grade bioreactor costs are as much as an order of magnitude more expensive.
Linch: Note that in the CE Delft (2021) and Risner, et al. (2020) reports, using food-grade bioreactors is an assumption of the models, not a conclusion. In contrast, Humbird (2020) spends ~3 pages of his report explaining the modeling choice of pharma-grade reactors and why he believes those are more appropriate.
Neil: Humbird’s (2020) model assumes that pharma-grade is needed to avoid contamination. It’s unclear to us what percentage of a production batch could be spoiled and still be profitable for cultured meat companies. It may be higher than what pharmaceutical companies would accept. Fassler’s (2021) article suggests that a single particle on a glove or a poor weld on a piece of equipment is enough to contaminate the batch and would take hours to resolve.
Media/growth factors
Fassler (2021) points out that fetal bovine serum-free media is expensive and challenging to create. Here Humbird (2020) is more optimistic, thinking falling prices are possible with economies of scale and growth factors won’t be needed in high concentrations anyway. CE Delft (2021) also projects growth factor costs falling but still regards them as a big hurdle. Risner, et al. (2020) think Essential 8, an animal free growth medium, does not seem like a viable option and growth factor costs need to be completely eliminated.
There is a large discrepancy between the costs for transforming growth factor beta (TGF-b). While Risner, et al. (2020) cites a study from GFI’s Liz Specht (2019) and a cost of $80.9M/gram from it, CE Delft cites an online market Qkine and that same paper from Specht but arrive at a range of $3.7M-$5M/gram for their baseline scenario. Specht’s paper also includes a cost for a hypothetical 20,000 liter batch of $3.2M, which is perhaps what CE Delft used. Humbird assumes two orders of magnitude cheaper costs are possible.
While this is not explicitly stated in any of the analyses, we think (following Humbird) there’s a reasonable argument that growth factors only need to be cheap for cultured meat, so not much prior research or optimization power has been spent on making growth factors cheap historically, so we should be much more optimistic about revolutionary change being possible here over eg.cheaper amino acids or bioreactors or changing the rate that animal cells double.
We’ve also talked to a few people working in adjacent fields who thought that massively cheaper growth factors were plausible, but none of them were very knowledgeable or confident about this area.
The viability of inexpensive amino acids is a major source of contention between the TEAs. Humbird (2020) thinks if cheaper amino acids were possible they would have been achieved already for the yeast industry. Risner, et al. (2020) also think amino acids are a huge prohibitive expense, but don’t discuss them in detail because they view growth factors as an obstacle that needs to be tackled before one can even think about amino acid costs. CE Delft (2021) uses a baseline price estimate for an amino acid product that would be poorly suited to cultured meat production as a feasible lower bound for prices.
Humbird’s (2020) best case model using enhanced hydrolysates says that macronutrients, i.e. amino acids, could reach $3.39/kg. Risner, et al. (2020) models the base media, which includes the amino acids, as costing $0.24-$3.12/L. CE Delft (2021) project amino acids from hydrolysate can be scaled to cost as little as $0.40/kg.
Fassler (2021) writes “Why the discrepancy? A footnote in the CE Delft report makes it clear: the price figures for macronutrients are largely based on a specific amino acid protein powder that sells for $400 a ton on the sprawling e-commerce marketplace Alibaba.com. That source, though, is not likely not suitable for cell culture. Via a chat tool, I asked the Alibaba vendor if the product would be acceptable for use in pharmaceutical-grade applications. “Dear,” she wrote back, “it’s organic fertilizer.” (In other words, it would not be.) As described on the webpage, the product is intended to be used in crop irrigation systems to help with plant nutrient uptake. The vendor did confirm it would be appropriate to use as an additive in livestock feed.” Fassler (2021) notes that “nutrition sources like the one sold on Alibaba will probably never work”: they’re not intended for human consumption, they may include heavy metals, arsenic, organic toxins, and so on which will get into the cells.
According to Humbird (2020), today’s production of certain amino acids would need to increase 6x to meet the demands of cultured meat, which he thinks is not going to happen. Humbird (2020) does explore the possibility that companies could derive a full amino acid profile from cheap commodity soy, reducing costs dramatically, however this would take years of research and a large scale up in the industry providing it
CE Delft (2021) is more optimistic about both alternatives like (soy) hydrolysate and lower grade conventional production of amino acids than Humbird (2020) is, meaning you don’t need to purchase expensive pure amino acids.
Individual amino acid costs don’t factor into Risner, et al’s (2020) written analysis mostly because they think the amino acids being a big variable in the cost calculation assumes that there has been a solution for the other expensive components like growth factors. The authors acknowledge that amino acid metabolism should be a consideration for commercial scale up. Their model does use a range of $0.24-$3.12/L for the base media which includes amino acids. These costs come from Specht (2019, p14). The actual range goes up to $376/L but Risner, et al (2020) only use the costs under conditions when growth factor requirements have been reduced and are produced at larger scales and higher efficiency. It’s also worth noting that p16 of this Agronomics slidedeck citing Specht (2019) includes a “current” media price of $10/L.
While Humbird’s (2020) take on amino acids is the most careful one so far, it is far from obviously correct. He curve-fits a regression model of existing amino acid volume: price pairs to the new “cultured meat” grade formulations, but some of the resulting prices are well above existing market rates for specific amino acids (human-supplement grade glycine can be bought at small scales for $27/kg from Amazon, considerably less than Humbird’s (2020, pg41) $87/kg estimates.).The synthetic biologists we talked to thought that revolutionary changes in amino acid production were not obviously implausible.
However, even if just one core amino acid at sufficiently high purity is too expensive, this may be enough to doom the entire project. At scale, amino acid prices can be lower than existing prices (as at-volume production is cheaper), but they can also be higher because the quality grade/purity needed is higher. For example, L-threonine (Thr), is currently produced at scale but using E.coli, meaning it needs to be filtered to be safe for consumption. Whether cheaper livestock-grade amino acids could be used is a testable hypothesis for cultured meat engineers.
According to Fassler (2021), “the cell density issue is one of the most intractable problems this emerging industry will face. Considering that the pharmaceutical industry has already likely spent billionson this very challenge—sums that make the total investment seen in cell meat look like a drop in the bucket—solving it would be a stunning accomplishment.”
As Humbird (2020) notes, talking about biological “max cell density” and “average cell volume” hides the actual issue that only so much of a bioreactor’s volume could be filled by cells. Humbird argues (citing some Chinese Hamster Ovary cell studies) that the practical upper bound for volume density is 25%.
In contrast, CE Delft (2021) starts with a volume density of 17.5% (below Humbird’s limit), but quickly goes on to
In a middle scenario, increasing max cell density by 4x, getting to a volume density of 70%, not only hitting way above Humbird’s practical upper bound but also coming dangerously close to the theoretical sphere packing limit.
In the final scenario, CE Delft increases the “Average cell volume” by almost 50%, from 3.5 x 103 µm3 to 5.0 x 103 µm3.
Unfortunately, they chained the two scenarios, implicitly claiming a 100% volume density (in other words, the entire bioreactor has only cells).
This is biologically/physically implausible as
There is no way for resources to get to cells if there’s no nutrient solution in the tanks
Animal cells are not perfect cubes
We’ve informed CE Delft of this error.
Thus, the qualitative take from Fassler (2021), that CE Deflt’s (2021) numbers are unrealistic, rather than essentially physically impossible, is arguably still too optimistic about CE Delft’s (2021) perspective here.
Risner, et al. (2020) repeats this 100% cell density implausibility intentionally in their all-problems-resolved scenario 4. “This was done to basically account for any innovation in vascularization and fusing of the cells” (supplementary material, p2). They recognise it is definitely biologically implausible and nearly physically impossible. Their use of 9.50E+07 max Cell density (cells/ml) was simply a midpoint between scenarios 1 and 4 and didn’t try to take into account Humbird’s suggested practical upper limit.
Max Cell density (cells/ml)
Average Cell volume (µm^3)
Total cell volume
Volume density
CE Delft lower limit
5.00E+07
3.50E+03
1.75E+11
17.5%
CE Delft upper limit
2.00E+08
5.00E+03
1.00E+12
100%
Humbird upper limit
5.00E+07
5.00E+03
2.50E+11
25%
Risner lower limit
1.00E+07
5.00E+03
5.00E+10
5%
Risner middle limit
9.50E+07
5.00E+03
4.75E+11
47.5%
Risner upper limit
2.00E+08
5.00E+03
1.00E+12
100%
Humbird (2020) takes the example of Chinese hamster ovary cells—immortal, non-cancerous, and morphologically stable cells that “can be grown in high-density culture” and “require only a limited number of growth factors, making it possible (though still challenging) to design cost-optimized, serum-free media” (Humbird 2020, p13). This particular cell line “which has benefitted from more than 60 years of constant research and development—is “probably not efficient enough for low-cost production of bulk cell mass,” according to Humbird in Fassler’s (2021) article. “The chemical reactions occurring inside a living cell proceed with the usual thermodynamic efficiency of 30–80%. If an organism lives at all, it does not do so orders of magnitude away from its physical limits” (Humbird 2020, p9).
Risner, et al. (2020) starts with human embryonic stem cells likely exhibiting a Warburg metabolism (aerobic glycolysis) glucose uptake rate of 4.13*10-13 moh/h but argue cell-engineering is needed to change to a more efficient oxidative phosphorylation metabolism to allow a rate as low as 4.13*10-14 moh/h, even after eliminating growth factor and amino acid costs completely.
Neil: Glucose specific metrics are used as a proxy for overall media consumption requirements as it is the most consumed nutrient in cell culture. We toyed around with their interactive online model. Moving glucose concentration and consumption rates to the limits allowed in their model (beyond those used in their scenarios and beyond known demonstrated limits as far as I can tell), the price is still prohibitively expensive when media costs are not reduced. Taking scenario 3 and removing other media costs but keeping the preset glucose variables holds the price above $60/kg. It’s unclear whether glucose consumption metrics are a good proxy for overall media consumption requirements.
Humbird (2020) and Risner, et al. (2020) also mentions other critical (but presumably solvable) scientific challenges to getting cultured meat at scale, including producing immortal cell lines breaking the Hayflick limit and cells that don’t automatically adhere to the sides of tanks (which animal cells are wont to do).
Concluding thoughts
This post is part of a project that aims to come up with concrete forecasts and opinions about the likelihood and timelines of cultured meat technology. The publication of existing cultured meat industry predictions and this summary of TEAs was expedited by articles in the media doing much of the same work. The next step in the project moves into forecasting. We intend to produce our own forecasts on cultured meat timelines, taking into account relevant reference classes, the TEAs, expert outreach, and our own analysis.
As this is both an important topic and a highly technical one where we do not have prior experience, we also plan to hire external experts to red-team our analysis and check for relevant flaws post-publication.
Acknowledgements
This research is a project ofRethink Priorities. It was jointly written by Neil Dullaghan and Linch Zhang. Thanks to Peter Wildeford and Marcus A. Davis for helpful feedback. Thanks also to everybody who responded to our emails and social media inquiries, including the TEA authors and relevant experts. If you like our work, please considersubscribing to our newsletter. You can see more of our work here.
Cultured meat: A comparison of techno-economic analyses
That is the takeaway from a new article by Joe Fassler (2021) in The Counter that draws heavily on two techno-economic analyses (TEA) of cultured meat (CE Delft 2021 & Humbird 2020). For full disclosure, we at Rethink Priorities were independently reviewing these TEAs (plus a third by Risner, et al. 2020) and in the process of writing a summary and comparison of them with our main takeaways.
The article addresses many of the issues we also noticed, and supplements them with interviews from industry experts. Here we want to acknowledge that they beat us to the punch somewhat, add a few relevant things we think the article left out from the comparison, and what the next steps are in our project.
The main cruxes of disagreement across the TEAs are:
Approach to the research question
Investor payback timelines
Food grade versus pharmaceutical grade bioreactors
The costs of media (growth factors and amino acids) at scale
The limits of cell-engineering needed to reduce media consumption needs
First though, we provide our quick summaries of the three TEAs so readers have a background before diving into the comparisons.
As we are investigating a scientific question that sometimes hinges on deep technical expertise (which neither Neil nor I have), we will likely have some errors in the summaries and (especially) personal takeaways. In addition, this report is less thoroughly checked than usual for Rethink Priorities reports. It should best be viewed as our current tentative understanding of the existing literature, rather than a final definitive summary of the existing literature.
tl;dr: We reviewed 3 TEAs on cultured meat. Our summary is that Humbird is very high quality and suggests cultured meat cost-competitiveness is hard and needs everything to go right. CE Delft outlines some of what will need to go right, but doesn’t provide much evidence that any of it is possible, has internal validity errors, and arguably has too much motivated reasoning. Risner, et al. is decent, within the narrow limits it sets itself, but too many details are underespecified for it to reflect the full costs and challenges of scaling up cultured meat. Reading the TEAs and doing surrounding research has turned Linch from a cultured meat optimist to being broadly pessimistic. Neil wants to be more agnostic until further research from Rethink Priorities and others.
Our TEA Summaries
As part of a project forecasting the potential for cost-competitive cultured meat to displace conventional meat, my colleague Neil and I investigated three techno-economic analyses (TEAs) estimating the economic feasibility of cost-competitive cultured meat ($2.50-$8/kg, akin to different estimates for existing wholesale meat prices).
A quick note on terms. The studies we looked at (and us) are only investigating “cultured meat”, that is, animal cell based meat of target conventional meats (usually beef). They do not investigate other technological alternatives to conventional meats such as plant*-based meats (like Impossible Foods or Beyond Meat), fungi-based meats (like Quorn), or brain-dead/genetically happier chickens (no known products).
In addition, the studies only look at the production side of one intermediate product: that of undifferentiated animal cell slurry (70-77% water). They do not investigate the question of scaffolding (converting cell slurry into structured meat products, eg. steak or chicken wings), nor the marketing and consumer acceptance questions of whether consumers will wish to buy cultured meat if it’s widely available and cheap.
To give a quick sense of the difference between the TEAs, here are the range of prices estimated in the TEAs given various assumptions, though note the lowest estimates from Humbird (2020) and Risner, et al (2020) are not regarded as realistic by the authors.
Price estimates explored in the TEAs, USD$/kg
Humbird (2020)
Risner, et al (2020)
CE Delft (2021)
$225.03
$437,000.00
$22,422.00
$236.58
$57,200.00
$1,707.00
$36.79
$44,500.00
$149.00
$21.00
$1.95
$116.00
$17.00
$8.00
$6.52
$5.74
Humbird (2020)
Executive summary
Humbird (2020) presents a study (funded by Open Philanthropy) which argues that a number of very difficult scientific and engineering problems need to be solved to get cultured meat to less than $25/kg (the reference point of the more expensive end of plant-based meats), at least with production processes anywhere similar to known processes.
Humbird models an ecosystem producing 100 kilotons/year of warm-blooded mammal cell mass. He believes that after a number of difficult engineering problems are solved to get to viable cultured meat at scale, the cost of cultured meat is set by expensive capital expenditures and media costs, particularly amino acids. Ultimately, Humbird believes that significant scientific and engineering progress is necessary but insufficient for cultured meat to displace conventional meat.
The core argument is that while microbial processes (for yeast, bacteria) have scaled over decades to extremely large production volumes (and create a lower bound for attainable prices given known limits), producing animal cells at industry scale present very large challenges in comparison:
Animal cells have a slower doubling time (24h-48h) than microbes (20min-3h)
Practical bioreactor volume is smaller, as animal cells are less tolerant of both spatial heterogeneity (eg temperature, pH) and fixes to heterogeneity (sparging, agitation)
Oxygen uptake/transfer rate and catabolite inhibition (lack of cell walls means harder to control osmolality) creates a limit to the practical final cell density
Animal cells need to be fed (expensive and globally low supply of) high-quality media components, particularly near-pure amino acids.
There are significant capital costs to maintain very high hygiene (“aseptic”) conditions, as any viral/bacterial contamination in an animal cell bioreactor is fatal.
Significant genetic engineering is necessary to produce animal cells with the desirable properties for bioreactor use (like breaking the Hayflick limit and achieving anchorage independence)
Humbird first models a Techno-Economic Analysis (TEA) where all parameters are similar to baker’s yeast, except for constraints imposed by the first three challenges (doubling time=24 h, maximum fermentor size=20 m3, maximum oxygen uptake rate=25 mol O2/m3-h).
The analysis comes to $3.87/kg of wet (70% water) cell mass for the constrained yeast process. As yeast production at scale is already a highly optimized process over many decades, and the additional constraints mentioned so far are pretty close to the fundamental biological nature of animal cells, it seems unlikely that we can do better than a lower bound of $3.87/kg. Unfortunately, there are other constraints.
For the rest of the study, Humbird then considers what steps are needed for a more realistic model that incorporates all the real constraints of cellular meat, but also at prices and quantities with a large ecosystem of producers and suppliers (commensurate with a production of 100 kilotons/year), and then does sensitivity analysis on the relevant prices.
The general upshot:
Significant metabolic engineering is necessary so enhanced cells have metabolisms at close to what Humbird believes are theoretical limits
Cheap at scale (<$200/kg) cultured meat with “wild-type” cells is simply not feasible
The cost of cultured meat is then bounded by the costs of high-quality media components, particularly near-pure amino acids, which Humbird models to be very expensive even at scale.
Humbird speculatively considers using soy hydrolysate instead of individual amino acids in media, which will be significantly cheaper if viable.
Secondarily, capital costs (especially but not limited to the cost of aseptic animal cell bioreactors) are also a large cost.
In the most optimistic scenario Humbird considered (cells with enhanced metabolisms designed, soy hydrolysate absorption, all other technical problems solved), Humbird estimates a cost of $21/kg wet cell mass.
Note that this does not include flavoring or structuring cell slurry into meat, nor does it include the amortized costs of R&D.
Humbird gives five key recommendations for future study. Note that in his model, solving these problems are necessary but insufficient for getting cultured meat to $21/kg.
Cell-line engineering
Plant protein hydrolysis
Heat-stable media formulation
Aseptic bioreactor design
Nitrogen integration
CE Delft report (2021)
Executive summary
Vergeer, et. al (2021) presents a study, using both public and private data, to argue that while substantial cost reductions are necessary, they are feasible to bring cultured meat to sufficiently low cost such that we can have a large burgeoning industry of cultured (“cultivated” in their words) meat by 2030. This study is conducted from the consultancy CE Delft and funded by the Good Food Institute.
The core argument goes: Existing models of cultured meat via current production technology and costs for inputs will result in very expensive cultured meat. However, based on their analysis of sometimes private data from cultured meat companies, new changes just around the horizon across a wide range of avenues should be enough to cut cultured meat to fairly affordable prices, perhaps close to conventional meat baseline by 2030.
The authors model an ecosystem producing 10 kilotons/year of wet (70-77%) animal cell mass, with some intentional vagueness about which animal types they ultimately refer to. They gathered data from 16 different (cultured meat and other) companies to do this analysis. They assume food bioreactors rather than traditional animal cell bioreactors as the baseline capital cost.
CE Delft first lightly models three “baseline scenarios” for the expected cost of a kilogram of cultured meat, using different ranges of ingredient costs and amounts required (high is upper end, mid is geometric means of all estimates, low is lower end for everything).
They then consider six scenarios for future cost reduction:
Note that scenario 4 onwards builds from scenario 3 (low end of all estimates) rather than scenario 2 (mid-range of estimates). This modeling choice is not explained.
Each scenario builds on top of the previous scenario.
Scenario 4: Lower prices for growth factors
“As in other literature, we adopt the scenario that a price reduction of a factor > 1,000 may be possible. “
Scenario 5: Cheaper or lower usage of recombinant proteins
“We have credible industry sources that point towards a feasible cost reduction of a factor 100.”
Scenario 6: Capital expenditures can be repaid at a slower rate than normal
While traditional investors require a payback time of <4 years for commercial investment projects, CE Delft argues that socially motivated investors can instead opt for a payback time of 30 years.
Linch (For judgment calls that one of us looked into in some detail, and the other broadly agrees with but is not vetted, we specify author names in Bold so it’s clearer that we’re referring to just one author’s opinion): It is unclear to me whether a 30 year payback time, even if true, is correctly pricing in what may presumably be higher maintenance costs, for 30+ year old equipment using novel technology.
Scenario 7: Increase in maximum cell density during proliferation
“50*106 cells/ml is the baseline cell density modeled. According to experts in the field, it may be feasible to increase cell densities by a factor 4 (to 20*107 cells/ml)”
“Higher cell densities would mean that more meat cells can be grown in a reactor of the same volume”
Though note that a factor of 4 increase in cell density means that the percentage of bioreactor volume that is now cells would increase from 17.5% to 70%, very close to sphere packing density limits.
Scenario 8: Shortening of production run time
“It is estimated that a reduction in production run time of 25% is feasible.”
Scenario 9: Increase in cell volume
“As the companies involved in this study produce a range of species and cell types, we used an average cell volume for the baseline scenario and determined potential larger cell volume (5,000 µm3 ) on primary data collected and literature.”
However, note that combining scenario 7 with scenario 9 translates to a pure cell density (% of bioreactor volume that is cells) of 100%.
Risner, et al. (2020)
Executive Summary
Risner, et al. (2020) present a study (funded by the Innovation Institute for Food and Health at UC Davis) on the costs of a 20m3 (20,000 L) food-grade core bioreactor based-system producing 121 kilotons of bovine cell wet mass (1% of current US beef production). The authors argue that “technological performance will need to approach [or surpass] known technical limits for animal cell based meat to achieve profitably as a commodity”. They think cultured meat “may not be economically viable as a commodity for some time” (undefined how long), but “may enter the market place sooner as a minor ingredient which lends desirable organoleptic qualities to an otherwise plant-based products” or become more like luxury food items; almas beluga caviar (US$10,000/kg), Atlantic bluefin tuna (US$6,500/kg), and foie gras (US$1,232/kg).
The study employs six global sensitivity analysis algorithms of 67 parameters, chosen using cellular biology and chemical/process engineering conventions. The list is narrowed down to the nine factors that most influenced capital and annual operating expenses by consolidating the top five parameters across all six algorithms:
Average single cell volume
Average single cell density
Fibroblast growth factor-2 (FGF-2) concentration
FGF-2 cost
Glucose concentration
Glucose consumption rate per cell
Maturation time
Transforming growth factor beta TGF-b (a growth factor) concentration
Oxygen consumption per cell
In the paper itself the authors don’t spend much time discussing TGF-b or oxygen consumption per cell, despite TGF-b being a major cost variable and oxygen uptake rates being a major obstacle in Humbird’s TEA. Their interactive cost calculator illustrates that media/recombinant protein costs also substantially affect cost (especially transferrin, L-Ascorbic acid 2-phosphate, and insulin). Furthermore, they also cite bioreactor capital costs as a significant factor in costs, even after assuming food-grade bioreactors.
These nine factors were then clustered into technological components. Four technology development scenarios are then developed, based on altering the values of some of these nine variables (see table below). The four technology development scenarios are:
Scenario 1 US$437,000 per kg : baseline scenario based on existing animal cell based meat production, in a food-grade bioreactor, including 2019 pricing (obtained from a GFI’s Specht (2019) for animal serum-free media and growth factors, and reported human embryonic stem cells’ glucose uptake rates.
Scenario 2 US$57,200 per kg: halves the cost of growth factors, reduces cell doubling time from 24hrs to 16hrs, halves glucose consumption rate per cell, increases total cell volume density 5x—based on 2x108 cells/ml average max cell density, which the authors note is a physical limit and the number in reality would be much lower.
Scenario 3 US$44,500 per kg: Same as scenario 2 but eliminates the cost of growth factors.
Scenario 4 US$1.95 per kg: An intentionally pie-in-the-sky scenario was designed to see what it would take to reach ~$2/kg (for conventional beef price parity). Nearly all technical challenges are resolved, including reducing growth factor costs to zero, increasing myoblasts/myosatellite cells (MSC) tolerance to glucose concentrations, glucose/media consumption is reduced by an order of magnitude, halving again myoblast/MSC doubling time and lower maturation time 6x, reducing media/recombinant proteins costs (base media containing over 50 ingredients, TGF-b, transferrin, L-Ascorbic acid 2-phosphate, and Insulin) to zero, and increasing total volume density to 100% (which is biologically implausible but unlike CE Delft (2021), Risner, et al (2020) acknowledge this).
Note that this TEA modelled only the core bioreactor system costs, as if every other part of the process was free, so is intended as a minimum cost.
The key three recommendations of the paper in order to lower costs are:
cell selection/engineering to lower the media consumption rate via a more glucose-efficient metabolism
“Engineering and/or screening for cell lines which shift rapidly from a Warburg metabolism (aerobic glycolysis) to a more glucose-efficient metabolism (oxidative phosphorylation i.e. production of 2 ATP vs. a theoretical 38 ATP per glucose molecule)”: from 4.13*10-13 moh/h cell glucose consumption rate per cell to 4.13*10-14 moh/h cell. Note that their argument is less about glucose specifically, which after all is just a simple sugar, and more about using it as a proxy for total media consumption requirements and thus a measure of cell-engineering.
reducing or eliminating the cost of growth factors (especially Fibroblast growth factor-2)
They suggest one way to eliminate FGF-2 costs is by “leveraging the ability of cancer cells to increase glucose uptake rates and exhibit cell proliferation without the presence of growth factors, or cell lines could be engineered/identified to express oncogenes related to these traits.” However, they note this “could be challenging from both a regulatory perspective as well as for consumer acceptance.”
scaling up of perfusion bioreactors.
Doing so allows for multiple changes of media per batch and higher cell concentrations. However, the TEA’s costs are based on estimates for standard food-grade bioreactors and Risner, et al. (2020) acknowledge more sophisticated bioreactors (i.e. single-use or novel perfusion systems) may substantially increase capital costs.
An implicit fourth recommendation is also to substantially lower the costs of the media/amino acids used.
“The amino acids being a big variable in the cost calculation assumes that there has been a solution for the other expensive components like growth factors” (private correspondence) . . .“Our model did not account for amino acid uptake rates due to glucose being the most consumed nutrient in cell culture, however amino acid (AA) metabolism should be a consideration for commercial scale up” (supplementary material, p2)
Why do the TEAs disagree?
Overall approach of the TEAs
Humbird asks, if cultured meat scaled to a global volume of 100 kilotons per year, what would it cost? CE Delft imagines a world where 10,000 metric tons per year of cost-competitive cultured meat can be produced, and then reverse engineers what would need to change for that to happen (unrelated to how likely those changes are). Risner, et al. (2020) follow a similar model to Humbird and then explore what changes to the major cost factors would be needed to get to price parity (acknowledging that they are unlikely).
Our takeaways on the TEAs overall:
Linch: My general impression is that Humbird’s analysis was very carefully and rigorously done (though I wish his uncertainty was sometimes made more numeric, as from an EA perspective many of these investments are worth doing even if there’s only a 5-20% chance of cultured meat working out in a thirty year time horizon). In contrast, my impression is that both the CE Delft report and its subsequent heavy publicity can best be viewed as a textbook exercise in motivated reasoning (akin to this, but less subtle). Various decision points and errors in the report shade in favor of cultured meat being cheap/quickly viable, likely without corresponding errors in the opposite direction. I have not read Risner et.al as carefully as Neil and feel less qualified to opine on it. My understanding is that Risner concludes cost-competitive cultured meat is impossible at current technology, but has an explicit “this is where a miracle happens” final step where you get very cheap prices if you assume away all technical and engineering difficulties. Note that I do not have prior experience in either this field or adjacent fields, and am biasing towards overexplaining and communicating my current opinions instead of only saying things I’m certain is true.
Neil: To me the major distinction between the TEAs is what reference class they are using for technological breakthroughs. Humbird (2020) and Risner, et al. (2020) seem to imagine the correct reference class is one with real technological and cost limits, like biofuels, while CE Delft’s (2021) view is closer to reference classes where the technical challenges can be overcome with enough effort/support from researchers, investors, and government, like photovoltaic technology (solar panels). None of the TEAs provide a useful way to adjudicate if it’s possible to beat the amino acid prices the yeast industry has achieved or to design bioreactor systems that can reduce the costs of contamination at scale, but I think CE Delft (2021) undersells just how much effort is needed. I also found it concerning for the CE Delft (2021) TEA that Linch was able to spot a fundamental mathematical/conceptual error without any prior domain expertise. These are the only TEAs I’ve read so I don’t have a good sense where they fall on the quality scale of TEAs in general. The TEAs should just be one (important) part of your model on how likely cultured meat is to scale.
Investor timelines
In brief, Humbird thinks it would take at least 60 years for the required ten orders of magnitude manufacturing cost reductions to be achieved via normal economies of scale, and investors work on much shorter timelines with higher return expectations for the first successful plant. CE Delft argues cultured meat investors are a special socially motivated breed that will wait 30 years for their returns.
Humbird (2020) simply doesn’t believe a scale up can be achieved in any reasonable investor timeline. He argues that a scale up to conventional meat volume, on the order of 100M metric tonnes/year, requires a ten orders of magnitude increase from the production volume of edible cultured animal cells at the time- on the order of 1–10 kg/y. He argues (2020, p9-10) it took 60 years for “certain aspects of computing to have progressed by about ten orders of magnitude”, and “generally accepted economies of scale in the process industries do not hold within them multiple orders of magnitude in manufacturing cost reduction” . Therefore, “to fit this many powers of ten into a time scale that meets contemporary investor expectations, some would suggest that even more accelerationist “laws” are in play for cultured meat”. Humbird (2020, p11) argues cultivated meat optimists are following a reasoning of “CAPEX only matters for the first plant; all subsequent plants will be financed at a low cost of capital” but that “such treatment from lenders is, more accurately, only enjoyed after the first successful plant. This need to reconcile a sense of urgency with a reluctance to be “first” was an exceedingly common source of friction in the days of venture-backed biofuels startups”.
Humbird (2020, p10) also seems concerned that limits to multiple orders of magnitude progress is especially true “in biotechnology, where the equipment in question is much smaller and more specialized and where the precursor chemicals are only produced in limited quantities”. This is based on Humbird’s conclusion that pharma-grade safety standards will always be required and amino acid costs will be high even at scale, both of which seem more malleable to us than Humbird claims.
CE Delft (2021, p21) writes “typically, the hurdle for commercial investment projects to get funded is payback time < 4 year” but argue socially motivated investors are willing to wait longer for returns. The authors think the payback time should be 30 years. This drops the price from $17/kg to $7.74/kg.
However, it’s unclear why 30 years is actually justified or why socially motivated investors would prefer much longer payback time relative to normal investment that will get you more resources to use for donation or social investment later.
Linch: I also think that CE Delft’s (2021) assumption that maintenance costs are a flat 5% of costs is suspicious/insufficiently defended.
Linch: while it’s possible I’m misunderstanding, a huge related issue with having a capital expenditure investment that takes >30 years to pay off is that you are explicitly baking in not changing capital use (much) which means you’re planning to keep old equipment, which seems especially wild when “old equipment” is decomposed into specific items like “bioreactors in a new field like cultured meat.” So I think that the CE Delft assumption about socially motivated investor attitudes is both by itself dubious but also their conclusion is invalid given just that assumption.
Food versus pharma grade bioreactors
Humbird (2020) thinks that cell cultures cannot be manufactured at large scale using food-grade standards because cell contamination will destroy the product. Therefore, cultured meat is limited to smaller scale pharma-grade standards with virtually no contamination, which are very costly. CE Delft (2021) imagines a world where it will be possible for cell cultures to be produced in food-grade bioreactors. Risner, et al. (2020) allow for the “optimistic assumption” that advances in science will enable production using food-grade bioreactor systems, and that “operational issues related to bioreactor sanitation and fill rates are negligible”.
For Risner, et al. (2020) “cost estimates of food-grade bioreactors were calculated using a method which accounts for equipment scaling, installation, and inflation. This method applies a set unit cost of $50,000/m-3 for a food grade bioreactor and a common scaling factor of 0.6”, and a capital cost of $778K for a 20K L bioreactor. CE Delft (2012) rely on more perfusion bioreactors in their model at $600K a piece, but their 10K L stirred-tank bioreactors are suggested to cost $325K a piece. Humbird’s pharma-grade bioreactor costs are as much as an order of magnitude more expensive.
Linch: Note that in the CE Delft (2021) and Risner, et al. (2020) reports, using food-grade bioreactors is an assumption of the models, not a conclusion. In contrast, Humbird (2020) spends ~3 pages of his report explaining the modeling choice of pharma-grade reactors and why he believes those are more appropriate.
Neil: Humbird’s (2020) model assumes that pharma-grade is needed to avoid contamination. It’s unclear to us what percentage of a production batch could be spoiled and still be profitable for cultured meat companies. It may be higher than what pharmaceutical companies would accept. Fassler’s (2021) article suggests that a single particle on a glove or a poor weld on a piece of equipment is enough to contaminate the batch and would take hours to resolve.
Media/growth factors
Fassler (2021) points out that fetal bovine serum-free media is expensive and challenging to create. Here Humbird (2020) is more optimistic, thinking falling prices are possible with economies of scale and growth factors won’t be needed in high concentrations anyway. CE Delft (2021) also projects growth factor costs falling but still regards them as a big hurdle. Risner, et al. (2020) think Essential 8, an animal free growth medium, does not seem like a viable option and growth factor costs need to be completely eliminated.
There is a large discrepancy between the costs for transforming growth factor beta (TGF-b). While Risner, et al. (2020) cites a study from GFI’s Liz Specht (2019) and a cost of $80.9M/gram from it, CE Delft cites an online market Qkine and that same paper from Specht but arrive at a range of $3.7M-$5M/gram for their baseline scenario. Specht’s paper also includes a cost for a hypothetical 20,000 liter batch of $3.2M, which is perhaps what CE Delft used. Humbird assumes two orders of magnitude cheaper costs are possible.
While this is not explicitly stated in any of the analyses, we think (following Humbird) there’s a reasonable argument that growth factors only need to be cheap for cultured meat, so not much prior research or optimization power has been spent on making growth factors cheap historically, so we should be much more optimistic about revolutionary change being possible here over eg.cheaper amino acids or bioreactors or changing the rate that animal cells double.
We’ve also talked to a few people working in adjacent fields who thought that massively cheaper growth factors were plausible, but none of them were very knowledgeable or confident about this area.
7-9
12-15
Amino Acids
The viability of inexpensive amino acids is a major source of contention between the TEAs. Humbird (2020) thinks if cheaper amino acids were possible they would have been achieved already for the yeast industry. Risner, et al. (2020) also think amino acids are a huge prohibitive expense, but don’t discuss them in detail because they view growth factors as an obstacle that needs to be tackled before one can even think about amino acid costs. CE Delft (2021) uses a baseline price estimate for an amino acid product that would be poorly suited to cultured meat production as a feasible lower bound for prices.
Humbird’s (2020) best case model using enhanced hydrolysates says that macronutrients, i.e. amino acids, could reach $3.39/kg. Risner, et al. (2020) models the base media, which includes the amino acids, as costing $0.24-$3.12/L. CE Delft (2021) project amino acids from hydrolysate can be scaled to cost as little as $0.40/kg.
Fassler (2021) writes “Why the discrepancy? A footnote in the CE Delft report makes it clear: the price figures for macronutrients are largely based on a specific amino acid protein powder that sells for $400 a ton on the sprawling e-commerce marketplace Alibaba.com. That source, though, is not likely not suitable for cell culture. Via a chat tool, I asked the Alibaba vendor if the product would be acceptable for use in pharmaceutical-grade applications. “Dear,” she wrote back, “it’s organic fertilizer.” (In other words, it would not be.) As described on the webpage, the product is intended to be used in crop irrigation systems to help with plant nutrient uptake. The vendor did confirm it would be appropriate to use as an additive in livestock feed.” Fassler (2021) notes that “nutrition sources like the one sold on Alibaba will probably never work”: they’re not intended for human consumption, they may include heavy metals, arsenic, organic toxins, and so on which will get into the cells.
According to Humbird (2020), today’s production of certain amino acids would need to increase 6x to meet the demands of cultured meat, which he thinks is not going to happen. Humbird (2020) does explore the possibility that companies could derive a full amino acid profile from cheap commodity soy, reducing costs dramatically, however this would take years of research and a large scale up in the industry providing it
CE Delft (2021) is more optimistic about both alternatives like (soy) hydrolysate and lower grade conventional production of amino acids than Humbird (2020) is, meaning you don’t need to purchase expensive pure amino acids.
Individual amino acid costs don’t factor into Risner, et al’s (2020) written analysis mostly because they think the amino acids being a big variable in the cost calculation assumes that there has been a solution for the other expensive components like growth factors. The authors acknowledge that amino acid metabolism should be a consideration for commercial scale up. Their model does use a range of $0.24-$3.12/L for the base media which includes amino acids. These costs come from Specht (2019, p14). The actual range goes up to $376/L but Risner, et al (2020) only use the costs under conditions when growth factor requirements have been reduced and are produced at larger scales and higher efficiency. It’s also worth noting that p16 of this Agronomics slidedeck citing Specht (2019) includes a “current” media price of $10/L.
While Humbird’s (2020) take on amino acids is the most careful one so far, it is far from obviously correct. He curve-fits a regression model of existing amino acid volume: price pairs to the new “cultured meat” grade formulations, but some of the resulting prices are well above existing market rates for specific amino acids (human-supplement grade glycine can be bought at small scales for $27/kg from Amazon, considerably less than Humbird’s (2020, pg41) $87/kg estimates.).The synthetic biologists we talked to thought that revolutionary changes in amino acid production were not obviously implausible.
However, even if just one core amino acid at sufficiently high purity is too expensive, this may be enough to doom the entire project. At scale, amino acid prices can be lower than existing prices (as at-volume production is cheaper), but they can also be higher because the quality grade/purity needed is higher. For example, L-threonine (Thr), is currently produced at scale but using E.coli, meaning it needs to be filtered to be safe for consumption. Whether cheaper livestock-grade amino acids could be used is a testable hypothesis for cultured meat engineers.
Some more thoughts here.
Cell engineering
According to Fassler (2021), “the cell density issue is one of the most intractable problems this emerging industry will face. Considering that the pharmaceutical industry has already likely spent billions on this very challenge—sums that make the total investment seen in cell meat look like a drop in the bucket—solving it would be a stunning accomplishment.”
As Humbird (2020) notes, talking about biological “max cell density” and “average cell volume” hides the actual issue that only so much of a bioreactor’s volume could be filled by cells. Humbird argues (citing some Chinese Hamster Ovary cell studies) that the practical upper bound for volume density is 25%.
In contrast, CE Delft (2021) starts with a volume density of 17.5% (below Humbird’s limit), but quickly goes on to
In a middle scenario, increasing max cell density by 4x, getting to a volume density of 70%, not only hitting way above Humbird’s practical upper bound but also coming dangerously close to the theoretical sphere packing limit.
In the final scenario, CE Delft increases the “Average cell volume” by almost 50%, from 3.5 x 103 µm3 to 5.0 x 103 µm3.
Unfortunately, they chained the two scenarios, implicitly claiming a 100% volume density (in other words, the entire bioreactor has only cells).
This is biologically/physically implausible as
There is no way for resources to get to cells if there’s no nutrient solution in the tanks
Animal cells are not perfect cubes
We’ve informed CE Delft of this error.
Thus, the qualitative take from Fassler (2021), that CE Deflt’s (2021) numbers are unrealistic, rather than essentially physically impossible, is arguably still too optimistic about CE Delft’s (2021) perspective here.
Risner, et al. (2020) repeats this 100% cell density implausibility intentionally in their all-problems-resolved scenario 4. “This was done to basically account for any innovation in vascularization and fusing of the cells” (supplementary material, p2). They recognise it is definitely biologically implausible and nearly physically impossible. Their use of 9.50E+07 max Cell density (cells/ml) was simply a midpoint between scenarios 1 and 4 and didn’t try to take into account Humbird’s suggested practical upper limit.
5.00E+07
3.50E+03
1.75E+11
17.5%
2.00E+08
5.00E+03
1.00E+12
100%
5.00E+07
5.00E+03
2.50E+11
25%
1.00E+07
5.00E+03
5.00E+10
5%
9.50E+07
5.00E+03
4.75E+11
47.5%
2.00E+08
5.00E+03
1.00E+12
100%
Humbird (2020) takes the example of Chinese hamster ovary cells—immortal, non-cancerous, and morphologically stable cells that “can be grown in high-density culture” and “require only a limited number of growth factors, making it possible (though still challenging) to design cost-optimized, serum-free media” (Humbird 2020, p13). This particular cell line “which has benefitted from more than 60 years of constant research and development—is “probably not efficient enough for low-cost production of bulk cell mass,” according to Humbird in Fassler’s (2021) article. “The chemical reactions occurring inside a living cell proceed with the usual thermodynamic efficiency of 30–80%. If an organism lives at all, it does not do so orders of magnitude away from its physical limits” (Humbird 2020, p9).
Risner, et al. (2020) starts with human embryonic stem cells likely exhibiting a Warburg metabolism (aerobic glycolysis) glucose uptake rate of 4.13*10-13 moh/h but argue cell-engineering is needed to change to a more efficient oxidative phosphorylation metabolism to allow a rate as low as 4.13*10-14 moh/h, even after eliminating growth factor and amino acid costs completely.
Neil: Glucose specific metrics are used as a proxy for overall media consumption requirements as it is the most consumed nutrient in cell culture. We toyed around with their interactive online model. Moving glucose concentration and consumption rates to the limits allowed in their model (beyond those used in their scenarios and beyond known demonstrated limits as far as I can tell), the price is still prohibitively expensive when media costs are not reduced. Taking scenario 3 and removing other media costs but keeping the preset glucose variables holds the price above $60/kg. It’s unclear whether glucose consumption metrics are a good proxy for overall media consumption requirements.
Humbird (2020) and Risner, et al. (2020) also mentions other critical (but presumably solvable) scientific challenges to getting cultured meat at scale, including producing immortal cell lines breaking the Hayflick limit and cells that don’t automatically adhere to the sides of tanks (which animal cells are wont to do).
Concluding thoughts
This post is part of a project that aims to come up with concrete forecasts and opinions about the likelihood and timelines of cultured meat technology. The publication of existing cultured meat industry predictions and this summary of TEAs was expedited by articles in the media doing much of the same work. The next step in the project moves into forecasting. We intend to produce our own forecasts on cultured meat timelines, taking into account relevant reference classes, the TEAs, expert outreach, and our own analysis.
As this is both an important topic and a highly technical one where we do not have prior experience, we also plan to hire external experts to red-team our analysis and check for relevant flaws post-publication.
Acknowledgements
This research is a project of Rethink Priorities. It was jointly written by Neil Dullaghan and Linch Zhang. Thanks to Peter Wildeford and Marcus A. Davis for helpful feedback. Thanks also to everybody who responded to our emails and social media inquiries, including the TEA authors and relevant experts. If you like our work, please consider subscribing to our newsletter. You can see more of our work here.