De-extinction and cultivated meat: the technology, the disciplines behind it, and where both actually stood in 2026
Written 2026-08-21 Source material 13 MrBeast video transcripts (held locally, not in this repository), centred on $1 vs $1,000,000,000 Futuristic Tech!, published 2026-01-24. Search conditions The web-search budget for this review ran out partway through. The later half relies on targeted retrieval from authoritative domains — GFI, FSANZ, GOV.UK/FSA, Wikipedia, AgFunderNews, Green Queen, FoodNavigator. Everything that could not be confirmed is marked as such in the text. Nothing has been filled in speculatively.
One: the video is from January 2026, not 2025. The results it shows are from 2025 — the dire wolves in April, the woolly mouse in March — but the video itself went out on 2026-01-24, about seven months before this was written. A good deal happened in those seven months, and it happened to be the part that mattered most.
Two: a transcript is not a fact. These are auto-generated captions. They mishear terms and numbers — the cheetah episode renders "olfactory receptors" as "olfactors". More importantly, a MrBeast video is sponsored content inside an entertainment format. Colossal and Upside Foods are both being showcased, so what the video says is close to each company's own external messaging. This report separates three things throughout: what the video demonstrably shows, what the creator or company asserts, and what independent checking found.
Three, and most important: the useful part of this document is not how impressive the technology is. It is the size of the gap between the narrative and the data. Both fields contain real and difficult science. Both are also substantially over-narrated. The interesting part is that in 2026 they diverged completely — one reached a $10.2 B valuation while the other's funding fell 96% from its peak. That fork says more than any individual technical detail.
- Part 1 — De-extinction
- Part 2 — Cultivated meat
- Part 3 — Animal technology in the other 12 videos
- Part 4 — The two stories are the same story
- Part 5 — From review to original results
- Appendix — Fact-check summary
- Going deeper
Jimmy visits Colossal Biosciences, where founder Ben Lamm walks him through three things:
- Sampling a 43,000-year-old Alaskan mammoth tusk
- Using a laser to perforate a cell and a micromanipulator needle to transfer a nucleus — he calls himself "a cloner", then fumbles it
- Petting the woolly mouse — a mouse called Dale with edited coat genes
And three lines that matter:
"How many decades till you're able to make a woolly mammoth?" — "We will have mammoths this decade."
"We have over a thousand species already in this bio-vault alone."
"You're watching the only two dire wolves in the universe, literally resurrected from the past."
Of those three: the third is wrong — there are three, not two. The second could not be verified. The first is a company commitment rather than a scientific prediction. Each is taken apart below.
There are only three technical routes. Understanding how they differ is the key to the entire controversy.
How it works. The genes of an extinct species often have not disappeared entirely; they survive scattered through the populations of living descendants or close relatives. Selective breeding gathers those traits back into the same animals.
Examples. The Quagga Project, rebuilding the quagga (extinct 1883) from plains zebra. The Tauros / Taurus Project, rebuilding the aurochs from extant cattle breeds — with a release planned for the Scottish Highlands in 2026.
The limit. What you get is a lookalike. Behaviour and physiology differ. You have bred for stripes; you have not bred a quagga.
How it works. Transplant a preserved nucleus from the extinct species into an enucleated oocyte of a close relative, and have that relative carry it.
Examples.
- Celia, the Pyrenean ibex (bucardo), 2003 — the only time an extinct species has ever been born. The clone died shortly after birth from a lung defect. It remains the sole record of an extinct animal being born alive.
- Elizabeth Ann the black-footed ferret and Kurt the Przewalski's horse are clones of endangered species, not extinct ones. That is genetic rescue, discussed below.
The limit. You need living or well-preserved frozen cells. For anything that died out thousands of years ago this route is simply closed — fossils do not contain living cells.
How it works. Since a complete extinct genome is unavailable, sequence its ancient DNA, identify the key differences from its nearest living relative, and use CRISPR to write those differences into the living species' genome.
The nature of this route settles one thing by definition: what you produce cannot be the original species.
You take a grey wolf genome and change 20 sites; or an Asian elephant genome and plan to change a few dozen. The difference between grey wolf and dire wolf genomes is millions of years of accumulated evolution across thousands of loci. Twenty edits versus twenty thousand is not a difference of degree.
The field's accurate term for the product is a functional proxy — an engineered organism that imitates the extinct species' ecological role, not the species itself.
Colossal's own Chief Science Officer, Beth Shapiro, conceded this in May 2025, describing the animals as "grey wolves with 20 edits", calling "dire wolf" a common name rather than a scientific one, and stating plainly that it is not possible to recreate an organism identical to one that once lived.
This is the hardest constraint in de-extinction, and the one most easily talked past.
After death, DNA begins two chemical processes immediately:
- Hydrolytic depurination — single-strand breaks that cut DNA into progressively shorter fragments.
- Cytosine deamination — causing C→T and G→A miscalls, which account for the large majority of ancient DNA sequencing errors.
The second is actually good news: it is ancient DNA's fingerprint. Modern
contaminating DNA does not carry that damage signature, so tools like
mapDamage2.0 and PMDtools can verify whether a sequence is genuinely
old. Contamination is the field's worst technical problem, particularly for
ancient humans, where everyone in the lab is shedding DNA continuously.
| Quantity | Value | Source |
|---|---|---|
| Time for mitochondrial DNA to degrade to a mean length of 1 bp (−5 °C) | 6.83 million years | moa bone decay kinetics |
| Nuclear DNA decay rate | roughly 2× that of mtDNA | as above |
| Theoretical ceiling for current sequencing | 0.4 – 1.5 million years | field consensus |
| Oldest genome sequenced | over 1 million years — Siberian mammoth teeth (Krestovka, Adycha, Chukochya) | van der Valk et al., Nature, 2021 |
| Oldest DNA recovered (environmental) | 2 million years, Kap København sediments, Greenland | Kjær et al., Nature, 2022 |
⚠️ The widely quoted "DNA half-life of 521 years" (Allentoft et al. 2012, Proc R Soc B) could not be checked against the original here — the Royal Society site returned 403. Only the alternative formulation from the same study (first row above) was verifiable. Cite with care.
Dinosaurs are out, permanently. Non-avian dinosaurs died out 66 million years ago — more than forty times past the theoretical ceiling. Jurassic Park is chemically impossible. That is not an engineering difficulty; it is thermodynamics.
But the ceiling hurts Colossal less than you would expect, because the CRISPR route does not need a complete ancient genome. It only needs to know which sites to change and to what, and that can be assembled from fragmentary ancient DNA plus a complete reference genome from a close relative. However short the fragment, if it maps back to the reference, the difference can be read.
So the real bottleneck is not ancient DNA. It is the three steps after it: how many sites you can edit at once, whether you can turn an edited cell into an embryo, and who carries that embryo.
Announced 4 March 2025.
- 12 genes edited in total — 7 directly in zygotes, 5 via an embryonic stem cell route
- Named genes: FGF5 (hair growth cycle length), MC1R (melanin type, and therefore coat colour), and one lipid metabolism gene
⚠️ The detail that matters: only 2 of the 12 genes had direct mammoth counterparts. The other 7 were chosen from known mouse coat genetics.- Phenotype: 5 cm golden-brown long hair
- Preprint: Chen et al., Multiplex-edited mice recapitulate woolly mammoth hair phenotypes, bioRxiv, 2025-03-04
- No formal peer-reviewed publication was found at time of search
- Cold tolerance was never tested — the company said at the time it would be done "in the coming months"
This is real work and should not be dismissed. Editing a dozen loci simultaneously in one animal and getting the expected phenotype is technically solid. Multiplex editing efficiency, off-target control, and whether an edited genotype reliably yields a phenotype are all unavoidable steps on the road to a mammoth, and the woolly mouse ran all three.
- Not cold tolerance. Long hair is not the same as surviving cold. Mammoth cold adaptation was systemic: coat, subcutaneous fat, haemoglobin that releases oxygen efficiently at low temperature, and reduced ears and tail. The haemoglobin is the physiologically demanding part.
- Not that it extrapolates to elephants. Mouse genetics is among the best-resourced systems in biology — tools, cell lines, decades of knowledge. Elephants have none of that.
- 12 genes against the required scale. Estimates for a real mammoth run to dozens or hundreds of loci, and the genetic basis of many relevant traits is simply unknown.
Robin Lovell-Badge of the Francis Crick Institute put it precisely: acknowledging the technical achievement while noting the study does not establish mechanism, and questioning whether the resources would do more good preventing extinctions currently under way. Victoria Herridge of Sheffield added that mammoth genetic complexity is far higher and the mechanisms are far from understood.
The biggest event in the field in 2025, and the thing the video gets wrong.
Announced April 2025. Three animals, not two:
- Romulus and Remus (male, 6 months at announcement)
- Khaleesi (female, 2 months at announcement)
The method:
- Ancient DNA from a 13,000-year-old tooth and a 72,000-year-old inner ear bone
- Chassis species: grey wolf
- 20 edits across 14 genes
- Surrogates: two hound mixes
- 45 engineered ova prepared in total
The number that matters is divergence time.
Perri et al., Nature 2021, established that the dire wolf and grey wolf lineages split around 5.7 million years ago, that dire wolves belong in the genus Aenocyon rather than Canis, and that there is no sign of admixture with modern grey wolves or coyotes — once separated, the lineages never interbred again.
⚠️ A 2026 phylogenetic study is reported to revise that divergence to about 4.5 million years. The original could not be obtained; flagged as unverified.
Either figure means the difference between the two genomes is genome-wide. Colossal edited 14 genes.
There is a subtler point. Some of the edits do not use the dire wolf variant at all. Certain dire wolf alleles were expected to cause blindness or deafness in a grey wolf background, so the team used known-safe grey wolf variants that produce a similar appearance. Some "dire wolf traits" are therefore built from grey wolf parts.
The IUCN SSC Canid Specialist Group issued a statement in April 2025: these animals are "neither dire wolves nor proxies", and the project does not constitute a conservation contribution, because the animals have no corresponding ecological niche and will not restore any ecosystem function.
The BBC headline was simply Experts question claim dire wolf brought back from extinction (2025-04-08).
In July 2025 New Scientist reported that critics Vincent Lynch, Flint Dibble, Victoria Herridge and Nic Rawlence were targeted by AI-generated smear posts and false YouTube copyright complaints. Colossal responded that neither the company nor its investors commissioned negative coverage of critics.
However that is ultimately attributed, it indicates how thoroughly scientific argument and public relations have become entangled in this field.
Ben Lamm says "We will have mammoths this decade"; the company's stated target is first calves in 2028.
| Date | Milestone |
|---|---|
| 2015 | Mammoth genes introduced into Asian elephant skin cells (60 genes targeted) |
| 2017 | 45 genes successfully added to the Asian elephant genome |
| 2022-07 | Complete Asian elephant genome sequenced with the Vertebrate Genomes Project |
| 2024 | First elephant iPSCs (induced pluripotent stem cells) |
The 2024 iPSC result is the genuine breakthrough. Elephant iPSCs are notoriously hard to establish — the cells respond poorly to standard reprogramming protocols and the field was stuck on this for years. Without iPSCs there is no source of cells that can be edited, expanded, and turned into germ cells. That step is what makes the route exist at all.
- Multiplex editing at scale. Dozens to hundreds of loci, all correct and stable in the same cell line.
- Surrogacy. Asian elephant gestation is about 22 months, the longest of any land animal. Asian elephants are themselves endangered, and using an endangered female to carry an engineered animal is a serious ethical and practical obstacle in its own right.
- Artificial womb. The only way around surrogacy. Colossal has an artificial womb prototype from the thylacine programme (2025-01, having carried a marsupial embryo past the halfway point of gestation) — but marsupial gestation is a few weeks with a barely developed placenta, making marsupials the easiest group in the animal kingdom for ex utero development. Going from that to a 22-month placental mammal is not a step up; it is a cliff.
An Asian elephant carrying some edited mammoth traits, born in 2028, is not impossible. A mammoth, in any meaningful sense, born in 2028, is not going to happen.
The distance between those two sentences is the entire controversy — and the company's external communication systematically blurs it.
| Species | Started | Progress | The hard part |
|---|---|---|---|
| Thylacine | 2022-08 | 2024-10: 99.9% of the genome reconstructed from a 110-year-old ethanol-preserved skull. 2025-01: complete genome, artificial womb prototype, marsupial embryo past halfway. 2025-05: key stem cells obtained. Estimated 8 years to release | The most technically feasible of the lot — short gestation, undeveloped placenta, young develop in the pouch, which sidesteps the biggest placental-mammal obstacle. The dunnart, its closest relative, is small and breeds quickly |
| Dodo | 2023-01 | Closest relative is the Nicobar pigeon; 2025-09, primordial germ cells (PGCs) cultured from rock dove as proof of concept | SCNT does not work in birds at all — the enormous yolk makes nuclear transfer physically impossible. The only route is editing PGCs and implanting them into a surrogate embryo, so the surrogate produces gametes carrying the target genes. An entirely different technology stack |
| Moa | 2025-07 | With Peter Jackson, Canterbury Museum and Ngāi Tahu | Same avian PGC route. The most contested: Philip Seddon (Otago) argues it has little to do with the global extinction crisis and is largely fundraising visibility; Nic Rawlence raises iwi opposition and data sovereignty |
| Bluebuck | 2026-04 | Colossal's first African programme; the species was hunted to extinction about 200 years ago | Announced after the video |
Great auk, Irish elk, ground sloth, sabre-toothed cat and woolly rhino are also listed as "under research".
The most surprising thing this review turned up: Colossal is a $10.2 B company with no product.
| Date | Round | Amount | Valuation |
|---|---|---|---|
| 2021-09-13 | Founding + seed | $15 M | — |
| 2022-03 | Series A | $60 M | — |
| 2023-01 | Series B | $150 M | >$1 B (unicorn) |
| 2024-10 | Colossal Foundation | $50 M | — |
| 2025-01 | Series C | $200 M | $10.2 B (Texas's first decacorn) |
$435 M raised in total as of 2025-01, roughly 170 staff and 95+ scientific advisers.
The video says "Colossal has spent over $400 million". That figure corresponds to total funds raised, not funds spent — a company founded in 2021 is unlikely to have burned $400 M. Treat it as promotional.
Largely not by selling anything. By turning technology into assets.
(1) Spinouts
- Form Bio (spun out 2022-09, $30 M) — AI platform for biological data
- Breaking (spun out 2024-04) — plastic degradation; claims its X-32 microbe breaks down plastic in about 22 months
This is a clever model. Building a mammoth requires developing a pile of general-purpose biotech tooling; that tooling is worth money in other markets, so it gets spun out and financed separately.
(2) Acquisitions
- 2025-08: the TIGRR lab at the University of Melbourne — first expansion outside the US, led by Andrew Pask
- 2025-11: ViaGen Pets & Equine — the important one. ViaGen is the commercial animal cloning company that clones pet dogs and racehorses for the wealthy. Colossal bought a real cloning business with actual cash flow, and with it hands-on large-mammal cloning capacity.
(3) Conservation work and government relationships
- The biovault, launched 2024-10, described as the world's largest distributed biobanking programme
- 2026-06: the U.S. Department of the Interior announced a biobanking partnership with Colossal through the U.S. Fish and Wildlife Service, framed officially around "President Trump's vision"
- 2026-01-13: Colossal Foundation announced genetic rescue programmes across six continents
This line deserves close attention. A private company has secured a partnership slot inside the US federal endangered species apparatus. That is both an enormous commercial moat and precisely what critics fear — it converts de-extinction from a contested scientific question into an established policy fact.
To be fair, the conservation side has produced real things:
- EEHV vaccine. Partnership with Baylor College of Medicine from 2022-10; first mRNA vaccination of an elephant in 2024-07. Elephant endotheliotropic herpesvirus is the leading killer of captive Asian elephant calves. This is work that saves lives.
- Northern quoll cane-toad toxin resistance (2024-05), amphibian chytrid immunity research (2024-12, $1 M donated), northern white rhino (with BioRescue), pink pigeon, Sumatran rhino
- Red wolf cloning (2025-04, three founder lineages) —
⚠️ contested. Joseph Hinton of the Wolf Conservation Center argues these are actually coyote-derived animals captured in southwestern Louisiana, since red wolves and coyotes historically hybridised and carry red wolf "ghost alleles"
⚠️ The video claims the biovault holds "over a thousand species". This could not be verified from any public source. Wikipedia lists roughly a dozen named species, and the "Colossal 100" endangered species list the company has referred to was planned but never published.
Covered in 1.2 and 1.5. The core is the IUCN SSC's determination and Beth Shapiro's own concession.
The most frequently made argument: if the public believes extinction is reversible, the urgency and funding for protecting living species get diluted.
- Nitin Sekar, Ars Technica 2025-04: Mammoth de-extinction is bad conservation
- The standard form of the argument: de-extinction money probably comes out of existing conservation budgets, which are zero-sum
- Notably, this worry is not new — it appeared in 1951, when the German association of zoo directors criticised revival attempts as "not only scientifically impossible but a diversion of resources from the pressing task of protecting threatened animals"
Surrogacy, miscarriage rates, and the health of cloned animals. Celia the bucardo's clone died shortly after birth from a lung defect, and that is not an isolated case — perinatal mortality in cloned animals is generally elevated. Using endangered Asian elephants as surrogates sharpens the problem considerably.
For cognitively sophisticated species there is a deeper issue: animals created this way may have the capacity for suffering and self-awareness, and in the absence of clear legal protection risk being treated as research tools or experimental assets rather than as sentient individuals.
- The niche may be gone. Thousands of years after extinction the environment has changed completely, and the original niche may be occupied by something else (competitive exclusion)
- It could become invasive
- Whatever killed it may still be there — put mammoths back in the Arctic and what happens about ivory poaching?
- Immunological naivety. A revived animal may be killed outright by diseases that living species adapted to long ago
- Legal ambiguity. How should the Endangered Species Act classify a gene-edited organism? This became contested in 2025 following statements from the Department of the Interior
Press release first, paper later is the most concentrated criticism. Both the dire wolves and the woolly mouse got large mainstream coverage first — the dire wolves made the cover of Time — with the preprint following and peer review still absent.
Ancient DNA professor Jeremy Austin was the bluntest: "more about getting scientists media attention than doing serious science."
This section should not be one-sided, because the supporting arguments are genuinely strong.
Technology spillover is the strongest one. Everything developed in order to build a mammoth — elephant iPSCs, multiplex CRISPR, artificial wombs, industrial-scale cloning capacity — applies directly to conserving living endangered species. The EEHV mRNA vaccine is the proof: it saves living Asian elephants, and it exists because Colossal needed to study elephants.
The money is additional, not diverted. Ben Lamm's repeated point: Colossal's $435 M came from technology venture capital, which would never have gone to conservation biology otherwise. Global conservation funding runs at roughly that order of magnitude per year, so calling this a diversion is not obviously fair.
Perfect is not the enemy of good. Beth Shapiro's position, roughly: since true resurrection is impossible, creating a proxy that performs the same ecological function — a cold-tolerant elephant that tramples tundra and helps restore mammoth-steppe ecology — is a worthwhile goal regardless of what it is called.
De-extinction is not a discipline. It is the intersection of five mature ones, and understanding what each cares about is more useful than memorising any individual technique.
Concerns. Extracting, sequencing and authenticating DNA from ancient remains, and using it to reconstruct evolutionary history, migration and population relationships.
Standing. This field has a Nobel Prize. Svante Pääbo received the 2022 Nobel in Physiology or Medicine for sequencing the Neanderthal genome. He showed that non-African modern humans carry 1–4% Neanderthal DNA, and — from a single finger bone — identified the previously unknown Denisovans. That is the discovery of a human species from DNA alone.
Key figures. Svante Pääbo; Beth Shapiro (now Colossal's CSO, previously a leading ancient DNA researcher at UC Santa Cruz and author of How to Clone a Mammoth); Love Dalén (mammoth genomes); Eske Willerslev (environmental DNA).
Its own core problems. Contamination control and authentication, sequencing chemistry for ultra-short fragments, and population-genetic inference from degraded sequence. Note that the mainstream concern of this field is not resurrection at all — it is understanding the past.
Concerns. The genetics of small populations — inbreeding depression, drift, accumulation of deleterious alleles — and how to intervene.
This is the field most relevant to de-extinction, and far more serious. Two cases everyone should know:
- Florida panther (success). By the mid-1990s roughly 22 animals remained, with inbreeding symptoms including kinked tails, cowlicks, sperm defects and heart abnormalities. After introducing 8 Texas pumas, the population grew 4% annually, rising from about 25 to over 100 within a decade, with hybrid kitten survival three times that of the pure line. A textbook win.
- Isle Royale wolves (cautionary). In 1997 a single wolf immigrated naturally into an isolated population of about 25. Fitness spiked initially, but by 2016 only two animals remained. Possibly the immigrant's deleterious alleles spread with its genes; possibly one migrant could not offset the accumulated genetic load. It demonstrates that genetic rescue outcomes are not predictable.
Risk concepts. Outbreeding depression, genetic swamping, and the most fundamental criticism — that it may treat the symptom while ignoring habitat loss, the actual cause.
Big-science projects. Vertebrate Genomes Project; Earth BioGenome Project, which aims to sequence every eukaryote on Earth.
Concerns. Treating biological systems as designable, manufacturable, standardisable engineering objects.
Key figure. George Church (Colossal co-founder, Harvard, and an early CRISPR developer). His lab's signature work includes minimal genomes and large-scale multiplex editing — knocking out 62 PERV viral loci in pig cells at once, the landmark multiplex record.
Core problems. The ceiling on edit scale, off-target effects, gene drives and their ecological risk, and governance frameworks for "should we do this at all".
Concerns. The engineering role large animals play in ecosystems, and what removing them did.
The entire ecological argument for de-extinction rests on this field. Colossal's case for putting mammoths back on Arctic tundra is that mammoths trample snow and knock over trees, holding the tundra in a mammoth steppe state — an ecosystem with higher carbon storage and higher albedo. In theory this slows permafrost thaw.
Precedent. Pleistocene Park in Russia, where the Zimovs have run the same experiment with living large herbivores for thirty years.
Criticism. The carbon argument has never been validated at scale, and the animal numbers required are in the hundreds of thousands — not a handful.
Concerns. Freezing genetic material before species disappear.
Examples. San Diego Zoo's Frozen Zoo (founded 1972; the world's oldest and largest wildlife cell bank, over 1,200 species) and Nature's SAFE in the UK. Elizabeth Ann the black-footed ferret exists because the Frozen Zoo banked cells from a female called Willa in 1988 — thirty years later, those cells restored genetic diversity to the species.
Colossal's biovault does the same thing at larger scale and more commercially. It is also its least contested and most widely accepted line of work.
Jimmy visits Upside Foods and is walked through the full four-step process:
- Cells taken from an egg in 2018, frozen at −80 °C
- Thawed and grown in an environment "simulating the inside of a chicken"
- Fed protein, fat and sugar — Jimmy pours it in himself: "I'm feeding those chicken cells so they multiply into chicken meat"
- Piped into a large tank, "through this you can see the chicken cells swimming around"
Then a blind tasting: chef Nick DiGiovanni makes two chicken sandwiches, one from a bird slaughtered that day and one cultivated. Jimmy guesses wrong.
The headline claims:
"produced hundreds of thousands of pounds of chicken meat without killing a single chicken" "spent over $200 million in research and development" "Every two seconds this produces a chicken's worth of chicken meat?" — "Yeah, every few seconds." "over 220 million chickens are killed every single day"
The first and third of those need the most scrutiny. But something else comes first.
The most unexpected finding in this review, and the single thing most worth knowing:
The video was published 2026-01-24. According to industry reporting, Upside Foods has not sold anything since it stopped supplying restaurants in 2024. So at the time of broadcast, the company had been at zero sales for roughly two years.
And more pointedly:
On 2026-01-23 — the day before the video went out — Upside Foods founded Lucius Labs, pivoting to sell cell culture media and analytical services to the life sciences industry.
A company formally redirected its main business toward selling media to biopharma the day before being filmed as the future of food. That coincidence is the best single summary of where this industry stood in 2026.
To be clear: this does not mean the video is fake. The facility is real, the process is real, the blind tasting is real. EPIC exists and can produce cultivated chicken. The issue is scale and commercial status — the video presents an industry changing the world, and the reality is an industry contracting, pivoting and looking for a way through.
- Upside's EPIC facility claimed 50,000 lb/year capacity at its 2021-11 opening, with a stated ceiling at full production of 400,000 lb/year
- But a Wired investigation on 2023-09-15 found significant problems with the company's bioreactor technology, and that it was relying mainly on far less efficient roller bottles — lab-scale equipment, labour-heavy and low-output
- GFI's own characterisation — and GFI is the industry's main advocacy organisation — is that the sector is "almost entirely pre-revenue", with only a few thousand consumers ever having bought a cultivated meat product
If "hundreds of thousands of pounds" means cumulative biomass produced, it is not impossible. If a viewer takes it to mean hundreds of thousands of pounds sold, it is seriously misleading.
Rough arithmetic: an edible broiler yields roughly 2–3 lb. At one chicken every 3 seconds, a year gives about 10.5 million chickens, or over 20 million lb — fifty times EPIC's own claimed full-production ceiling of 400,000 lb/year.
This fails on order of magnitude. The charitable reading is that it describes how many chickens' worth of cells are in the tank, not a production rate, and the caption or edit blurred it.
Upside has raised roughly $608 M in total. $200 M of R&D spend is credible and possibly low.
Global broiler slaughter runs about 70–75 billion per year; divided by 365 that is 190–210 million per day. 220 million is within range. This number is true, and it is the moral foundation the whole industry rests on.
Four pillars, each stuck in its own way.
Starting material. Skeletal muscle satellite cells, fibroblasts, mesenchymal stem cells, iPSCs, embryonic stem cells, adipose-derived cells. Each is a trade-off: satellite cells naturally become muscle but proliferate poorly; iPSCs proliferate indefinitely but are harder to induce and differentiate.
Why immortalisation is mandatory. Normal somatic cells hit the Hayflick limit and senesce after a few dozen divisions. Industrial production needs cells that passage indefinitely. Two routes: introduce hTERT or SV40 — genetic modification, which raises public "is this GMO / is this a cancer cell" concerns — or wait for spontaneous immortalisation.
A November 2025 study is genuinely good news here: bovine cells can spontaneously immortalise after 500 days in culture, with no genetic modification (chicken cells immortalise spontaneously more readily anyway). That sidesteps a major regulatory and perception obstacle.
The underlying difficulty. Cell biology for farm species is far less developed than for human, mouse or hamster cells. Decades of CHO cell tooling from biopharma start from zero in cattle and chickens. GFI's figures: establishing and characterising one cell line takes 6–18 months, and as of 2025 only about 90 were tracked across the entire industry.
On the −80 °C in the video. That is the master/working cell bank system. A −80 °C freezer suits medium-term storage — months to a few years — but genuine long-term preservation uses liquid nitrogen at −196 °C, because at −80 °C there is still slow molecular motion and ice recrystallisation, which accumulates damage. The video is not wrong, but that is almost certainly a working bank rather than a master bank.
This is the single determining factor for the industry.
Conventional cell culture depends on foetal bovine serum (FBS) — literally serum drawn from calf foetuses of pregnant cows. Two fatal problems: it is ethically absurd for a technology premised on not killing animals, and it is economically impossible at roughly $1,000/L. It also varies by donor animal and cannot be chemically defined, which is a disaster for food regulation.
So serum-free media is mandatory: a basal medium such as DMEM/F12 plus recombinant growth factors — FGF2, TGF-β, insulin, transferrin, albumin.
Cost structure. Growth factors and recombinant proteins dominate. Deco Labs, a Tufts spinout, gave a specific breakdown in 2026-06: albumin alone contributes about $100/kg of production cost.
This is the one direction with repeatedly verified real progress in 2026:
| Source | Media cost | Status |
|---|---|---|
| Conventional pharma grade | hundreds of $/L | — |
| GFI target | <$0.25/L (99.9% below pharma grade) | target |
| Published research | $0.63/L | achieved |
| Clever Carnivore | $0.07/L | achieved at pilot scale |
| Meatly | £1/L (2024) → 22p/L (achieved 2025) → 1.5p/L (industrial target) | partly achieved |
| Deco Labs cAlbumin | target $0.02/L | in development |
That curve is real and falling fast. Looking at this line alone, you would be optimistic about the industry.
Animal cells are nothing like microbes. Brewing yeast has a cell wall, tolerates vigorous agitation, has simple metabolism, and doubles in 90 minutes. Animal cells:
- Extremely shear-sensitive — impellers and bursting bubbles tear them apart (Humbird singles out bubble-induced cell damage)
- Mostly anchorage-dependent — they must attach to a surface, hence microcarriers
- Doubling time 20–24 hours, an order of magnitude slower
- Metabolite accumulation — ammonia and lactate self-poison the culture; CO₂ inhibition is another bottleneck Humbird identifies explicitly
- Extreme sterility requirements — animal cell media is also perfect bacterial media, and one contamination writes off the whole vessel. This drives capital expenditure directly
Actual scale as of 2026-08:
| Organisation | Scale |
|---|---|
| Vow (Sydney) | 22,000 L single vessel — claimed the world's largest food-grade cell culture bioreactor |
| Meatly (London) | 20,000 L (under construction) |
| Aleph Farms | 5,000 L, emphasising off-the-shelf equipment |
| Zhou Zi Future (Nanjing) | 2,000 L, China's largest pilot plant, designed for 10–50 t/year |
| Biopharma's practical ceiling | around 25,000 L |
| The 2021 ABEC / GOOD Meat contract | multiple 250,000 L vessels |
Read the last two rows together and you have the whole story of 2021 to 2026: that 250,000 L contract was never built, and ended in a $4.4 M settlement (ABEC had sued for over $60 M).
The 2021 blueprint and the 2026 reality differ by an order of magnitude.
Why mince is easy and steak is extremely hard. One reason: the diffusion limit.
In tissue without vasculature, oxygen and nutrients penetrate only by diffusion, effective over roughly hundreds of micrometres. Beyond that thickness the cells at the centre become hypoxic and die, forming a necrotic core. A 2 cm steak is a hundred times that limit.
Real meat solves this with a capillary network — no cell is more than tens of micrometres from a capillary. To make genuine whole-cut meat you have to build a vascular system, which is the hardest problem in tissue engineering and the reason artificial organs remain out of reach.
Current approaches:
- Edible scaffolds — polysaccharides (chitosan, alginate, cellulose), proteins (zein), mycelium, decellularised plants (strip the cells from a spinach leaf and the remaining vasculature is already a capillary network — an elegant idea), 3D bioprinting
- Hybrid products — cultivated cells plus plant protein. Currently the only commercially realistic route
- Cultivated fat as a flavour ingredient — fat needs no complex structure and carries most of meat's flavour, which is why selling cultivated fat to plant-based companies as an ingredient became a popular niche
How hybrid, exactly? Look at the real products:
- Aleph Farms' cultivated steak, approved in Singapore 2026-08, contains only about 10–20% cultivated beef cells; the rest is a soy/wheat protein matrix
- GOOD Meat's 2024 product at Huber's Butchery in Singapore contained 3% cultivated chicken
- Meatly's dog treats at Pets at Home in the UK contain 4% cultivated chicken
- Clever Carnivore states its strategy openly: 10% cultivated cells, 90% plant-based
This is almost never emphasised in public communication, and it is the single most important thing to understand about the industry.
Mark Post (Maastricht University) made the first cultivated beef burger in 2013: over twenty thousand muscle strands, costing more than $325,000, two years of work, publicly tasted in London on 2013-08-05.
Costs have fallen several orders of magnitude since. The question is whether they can fall several more.
The most important critical paper in the field: "Scale-up economics for cultured meat" (Biotechnology & Bioengineering, 2021, funded by Open Philanthropy).
⚠️ The Wiley version returned 403 here. What follows is taken from the engrxiv preprint abstract and is verified original wording. The widely quoted figure of "about $37/kg" comes from the journal version and could not be verified first-hand.
Humbird is a chemical process engineer, not an industry insider, and what he did was a capital and operating cost analysis of a conceptual facility. The core conclusion, in his words:
"capital- and operating-cost analyses of conceptual cell-mass production facilities indicate production economics that would likely preclude the affordability of their products as food."
And the crucial judgement: improved metabolic efficiency and low-cost plant-hydrolysate media are "necessary but insufficient".
The barriers he identifies:
- Low growth rate, low metabolic efficiency
- Catabolite inhibition and CO₂ inhibition
- Bubble-induced cell damage — these three together cap feasible reactor volume and achievable cell density
- Excessive capital expenditure on contamination-proof facilities
- No food-grade production capacity for amino acids and protein growth factors
His argument has force because it is not "this is expensive now" but "there is a physical and engineering floor here." Learning curves reduce cost; they do not go through physics.
- CE Delft 2021 (commissioned by GFI) produced an optimistic techno-economic pathway
- GFI's own position is careful: the models "adopt different optimistic or pessimistic assumptions and therefore reach different conclusions", and should be treated as "a roadmap for understanding cost drivers and prioritising research" rather than as prediction
- Aleph Farms' independent TEA: at parity with conventional beef, unit production cost $6.45/lb, 47% gross margin, 2.5-year payback, using 5,000 L off-the-shelf reactors. CTO Neta Lavon says profitability is achievable "with non-GMO cells and equipment you can buy today" — but this is a model, not a measurement
The comparison that matters most:
| Item | Figure | Nature |
|---|---|---|
| Arthur D. Little estimate (via Parima, 2026-07) | finished product around €40+/kg | the only finished-product cost estimate with any third-party character |
| Same estimate's end-of-decade projection | around €10/kg (−75%) | projection |
| US wholesale chicken | about $2–3/lb ≈ $4.4–6.6/kg | the real benchmark |
€40/kg against $5/kg — still an order of magnitude apart.
This argument had a public collision in April 2025 worth recording in full (source: AgFunderNews, 2025-04-10).
Food-tech investor Lever VC published A Second Generation of Cultivated Meat Companies Breaks Through Projected Cost Barriers, calling Humbird "spectacularly wrong":
| Metric | Humbird's 2021 estimate | Lever VC's claimed 2025 reality |
|---|---|---|
| Media cost | about $6.50/L | under $1/L, several below $0.50/L; Believer Meats has a peer-reviewed $0.63/L |
| Cell density ceiling | around 60 g/L | leaders at 60–90 g/L, 50% above that ceiling |
| Cell mass COGS | — | $10–15/kg, leaders below $10/kg |
Humbird's rebuttal was sharp, and technically sound:
- He never said media would stop at $6.50/L; after publication his own figure was $3/L
- "Comparing $/L without cell concentration and litres-per-kilogram data is odious" — the decisive point. Media at half the price used in twice the volume is not a saving
- Lever VC gave no citations, resting on press releases that were themselves uncited
- In his words, the report read like "a poor ad hominem attack on a five-year-old paper nobody is really discussing any more"
Lever VC's Nick Cooney conceded Humbird's technical point on feed conversion but said they held NDA-protected diligence data including concentrations and L/kg that could not be published; they did not cite published research because "there is so little out there", with formulations treated as trade secrets.
GFI chief scientist Elliot Swartz was the most neutral: he has become more optimistic about techno-economics since Humbird's paper, but could not comment on Lever VC's specific figures.
But two counter-data points from Vow deserve equal recording, because they do undercut Humbird's CAPEX argument:
- A 20,000 L bioreactor built for "well under $1 million" — if accurate, this substantially weakens the "contamination-proof capex is prohibitive" claim
- Spontaneously immortalised quail fibroblasts needing only two growth factors and no albumin — sidestepping the most expensive inputs directly, and echoing the 2025-11 bovine spontaneous immortalisation result
The optimists are arguing about learning curves: every generation gets cheaper, media has gone from $1,000/L to $0.07/L, so why should reactors and scale-up be different?
The pessimists are arguing about physical ceilings: media got cheap because it is a chemicals procurement problem, but oxygen transfer, shear, the diffusion limit and sterile capex are hard process-engineering constraints that do not yield to spending.
Humbird's "comparing $/L alone is odious" identifies the thing most easily missed in this argument: every cost-reduction announcement reports the unit price of an input, and almost nobody reports how many litres are needed per kilogram of product. The first number looks good; the second is the denominator.
Assessment. The media curve shows the optimists are right about inputs. The gap between 22,000 L and 250,000 L shows the pessimists are right about scale-up. Both can be true at once, and together they suggest the likely outcome is neither replacing meat nor total failure, but finding a viable position in premium niches and as a hybrid ingredient.
The hardest data in this document, and the most sobering.
| Year | Global cultivated meat funding (GFI figures) |
|---|---|
| 2021 | $1.3 B (peak) |
| 2022 | $917 M |
| 2023 | $230 M |
| 2024 | $144 M |
| 2025 | $73.9 M, −49% year on year |
| 2026 Q2 | $25 M |
⚠️ Three mutually inconsistent figures exist for the 2021 peak: $1.3 B (GFI), $1.8 B (Green Queen), $989 M (AgFunderNews' own basis). Cite with the source attached.
Alternative protein as a whole raised $359 M in H1 2026, of which cultivated meat's Q2 was $25 M. Extrapolating, 2026 will likely land between $50–100 M — still flat on the historical floor, with no sign of recovery.
Other indicators:
- Company count: 155 in 2024 → 140 in 2025 (net −15)
- Alternative protein raised $881 M in 2025, below $1 B for the first time in seven years, and one-eighth of the 2021 peak of $6.9 B
- Cultivated meat accounts for only 6.3% of all alternative protein jobs
| Company | Date | Raised | Note |
|---|---|---|---|
| SCiFi Foods | mid-2024 | — | "no funding, ran out of time" |
| Upstream Foods (cultivated fish fat) | mid-2025 | — | Netherlands |
| CellRev | 2025-08 | — | UK |
| Believer Meats | ceased 2025-12, bankrupt 12-31 | $387 M | see below |
| Meatable | dissolved 2025-12-19 | ~100 staff | Agronomics wrote its £11.9 M stake down to zero |
| Cultimate Foods | bankrupt 2026-04 → liquidated 2026-08 | €2.3 M | Germany, cultivated fat |
Believer Meats' collapse is the industry's harshest lesson and deserves its own account:
- Founded 2018, raised $387 M in total
- Wilson, North Carolina plant: 200,000 sq ft, $154 M to build, designed for 12,000 t/year of cultivated chicken, described as the world's first and only large-scale cultivated meat production facility
- Obtained both FDA and USDA clearance during 2025
- And closed a few months later
- 2025-12: contractor Gray Construction sued over $34 M owed; the company announced closure a week later
- At bankruptcy filing on 2025-12-31 it had roughly $86,000 in cash
- Upside Foods bid $50 M for the plant — 32 cents on the dollar for a $154 M asset
- Auction 2026-08-17, sale hearing 2026-08-20.
⚠️ No public report of the outcome as of 2026-08-21
Full regulatory clearance, the world's largest facility built, and then insolvency — that says more about where the problem is than any technical argument.
In a gold rush the people who profit are selling shovels. So when the shovel-sellers leave, they have concluded there is no gold:
| Company | Was | Pivoted to |
|---|---|---|
| Ark Biotech | cultivated meat bioreactors | pharmaceutical software |
| Future Fields | recombinant proteins | life sciences |
| ProFuse Technology | media supplements | drug discovery (GLP-1-related sarcopenia) |
| Uncommon Bio | cultivated meat platform | therapeutics |
| UPSIDE Foods | cultivated meat | Lucius Labs — the reverse: a food company becoming a supplier |
ProFuse CEO Guy Nevo Michrowski put it plainly: "meaningful revenue is not coming within five years."
- B2C → B2B — selling ingredients or contract manufacturing rather than meat
- Own megafactory → asset-light / CDMO — Aleph, BlueNalu and Vow all pivoted
- Whole cuts → mince and hybrid formulations
- Pure food → adjacent verticals — pharma, cosmetics, leather, pet food
Point 4 has an excellent example: Japan's IntegriCulture posted its first full-year profit in FY2025, roughly a decade after founding — about $1.5 M in annual sales, with cell-based cosmetics as the main source of profit, not meat.
The only part of the industry with real retail revenue, and the recipient of 2026's largest single funding round.
Three structural advantages: a far simpler regulatory path (in the EU it registers as a feed material rather than novel food), higher consumer acceptance, and higher margin tolerance — pet treats are expensive anyway.
- Meatly (UK): £10.4 M Series A closed 2026-05-07 — the largest single equity round in cultivated meat so far in 2026; London factory broke ground 2026-06-04 (Europe's largest); Chick Bites on shelves at Pets at Home since 2025-02
- Bene Meat + Forza10: Coolty Meat dog food launched in the EU 2026-05-13 — the EU's first cell-cultured pet food on the market
- Friends & Family Pet Food: approved by Singapore's AVS 2025-06 (first in Asia); 8 SKUs across 8 retail locations in 2026, roughly 400 packs sold in the first six weeks
Note that last figure: 400 packs. That is the industry's real commercial scale today.
| Jurisdiction | Human food | Pet food |
|---|---|---|
| Singapore | ✅ 6 approvals — the only multi-product, multi-company market on Earth | ✅ |
| Israel | ✅ 1 (2024-01) | — |
| United States | ✅ 4 companies (Upside, GOOD Meat, Mission Barns pork fat, Wildtype salmon) | — |
| Australia / NZ | ✅ 1 (Vow quail, effective 2025-06) | — |
| United Kingdom | ❌ | ✅ Meatly (2024-07, Europe's first) |
| European Union | ❌ | ✅ via the feed pathway |
| China / Korea / Japan / Thailand / India / Brazil / Canada / Switzerland | ❌ | ❌ |
The key conclusion: outside Singapore, no other jurisdiction approved a new cell-cultured human food anywhere in the world during 2025–2026.
- Israel: Aleph Farms was approved in 2024-01, but as of 2026-08 there is no evidence it ever went on commercial sale. The website says only "Now approved in Israel", and 2026 coverage describes Israeli launch as a plan "by 2027"
- Singapore: of six human-food approvals, the only one confirmably on sale in 2026 is pet food. GOOD Meat's status is unclear (Eat Just suspended Singapore operations in 2024-03)
- United States: Upside has zero sales
"Number of approvals" is becoming a distorted industry metric.
- State bans. Florida (manufacture and sale is a criminal offence), Alabama (illegal since 2024-10), South Dakota (effective 2026-07-01, expiring 2030-06-30). Labelling restrictions: Missouri, South Carolina, Texas, Washington. Under consideration: Arizona, Kentucky, Tennessee, West Virginia
- Federal. The bipartisan FAIR Labels Act was introduced 2026-05-05 (backed by the beef industry), which would bar cultivated and plant-based products from using meat-related terms on labels
- GFI states plainly that the US substantially cut all federally funded alternative protein R&D in 2025
- Upside sued Florida in 2024-08.
⚠️ Outcome could not be verified
⚠️ Details of the 2025 Montana, Indiana, Nebraska and Texas measures could not be verified.
- The UK is the most likely second Western human-food market in 2027 (the FSA regulatory sandbox runs to 2027-02, with Gourmey duck and Vital Meat chicken furthest along). But there is a fatal variable: on 2026-07-10 the FSA warned around 600 businesses that the new UK–EU SPS trade agreement — expected mid-2027, including dynamic alignment — may invalidate independent UK approvals, forcing companies onto the EU pathway, which takes up to 6 years. This is the only variable in the global regulatory picture that could move backwards.
- Italy legislated a full ban in 2023-12. But when the European Commission closed its TRIS review in 2024-01 it found Italy had breached EU procedure by passing the law before the standstill period ended, which under CJEU case law may render it unenforceable in domestic courts. Hungary also passed a ban on 2025-11-19
- The Netherlands leads Europe: €4.1 M to three cellular agriculture projects on 2026-03-31; the world's first cultivated meat farm opened at Schipluiden, South Holland, on 2026-06-04 — built on an existing dairy farm, purely demonstrative, selling nothing
- China: the 15th Five-Year Plan (finalised 2026-03-12) brings "novel
protein sources" into national food security strategy, explicitly citing
synthetic biology.
⚠️ But cultivated meat is not named specifically, and there is still no approval framework at all. GFI APAC relays an industry expectation of "possibly late 2026" — that is company expectation, not official commitment- The industrial side is far stronger. Nanjing's Zhou Zi Future (from Professor Zhou Guanghong's group at Nanjing Agricultural University, which made China's first cultivated meat in 2019) completed China's largest pilot plant in 2025-12 — 2,000 L, 10–50 t/year — running the world's first 2,000 L scale cultivated pork production trial, and holds the second-largest patent family after UPSIDE Foods
- This is the mirror image of the Western situation: China has industrial capability ahead of regulation, the West has regulation ahead of capacity. Chinese companies' current practical route is through Singapore
The term was coined by Isha Datar in 2015, in a New Harvest Facebook group. New Harvest is the field's founding organisation, established in 2004 by Jason Matheny — who later became Director of IARPA and is now CEO of RAND, an interesting career trajectory in its own right.
It splits in two, and the division matters:
(A) Cell-based = tissue engineering = cultivated meat, milk, egg. Hard, expensive, slow.
(B) Acellular = precision fermentation — not culturing animal cells at all, but using engineered microbes to produce a single target molecule.
Branch B's record dwarfs branch A's:
- Chymosin — the first GRAS transgenic food, in 1990, and now over 90% of the global rennet market. Most cheese you eat is made with it, and nobody objects
- Perfect Day's whey protein β-lactoglobulin (FDA "no questions" in 2020, produced in Trichoderma)
- Impossible Foods' soy leghemoglobin — the secret of the plant burger that bleeds
- Precision fermentation market: $2.1 B in 2023, projected above $100 B by 2034
The comparison is itself an important lesson: producing a molecule is orders of magnitude easier than producing a tissue. Cultivated meat raised $73.9 M in 2026; the fermentation segment raised $357 M — nearly five times as much.
Cultivated meat inherited its technology directly from here. Mark Post's 2013 burger came out of Maastricht's vascular tissue engineering lab — his actual research was making vascular grafts for cardiac patients.
That migration, from medical technology down into food, is the key to the cost problem.
Medical tissue engineering tolerates costs of hundreds to thousands of dollars per gram — one patient needs a small piece, and insurance pays. Food tolerates a few dollars per kilogram.
That is six orders of magnitude. A technology entirely viable in its native domain becomes entirely unviable when moved into a cost environment six orders of magnitude away. This is not a matter of insufficient effort; it is a mismatch of setting.
Humbird's paper is, ultimately, saying exactly this.
GFI's standard three-way split: plant-based, fermentation, cultivated, in descending order of maturity.
What this field cares about is "why bother", and those arguments are solid:
- Livestock accounts for roughly 11–14.5% of global greenhouse gas emissions (FAO basis; contested, but the order of magnitude holds)
- Land use: the large majority of agricultural land goes to livestock and feed
- Antimicrobial resistance — a large share of global antibiotic use is in livestock
- Zoonotic risk — H5N1 continued to spread through 2024–2026 and has crossed into dairy cattle. This is the most urgent one
- Animal welfare: 200 million chickens a day
Note that these arguments holding does not make cultivated meat the solution. The evidence currently favours the plant-based and precision-fermentation legs over the cell-based one.
The least glamorous field and the one that decides the outcome. It concerns exactly what Humbird's paper concerns: reactor configuration, oxygen transfer coefficient (kLa), shear stress, scaling laws, sterile design, CAPEX/OPEX modelling.
A structural problem for this industry is that it is led by biologists and food-tech founders, while what actually blocks it is chemical engineering.
Cultivated meat's environmental benefit is not settled. Work including a 2023 UC Davis preprint argues that if pharmaceutical-grade inputs continue to be used, cultivated meat's carbon footprint could be several to several dozen times higher than beef, because upstream purification is extremely energy-intensive. The industry's response is that this reflects pharma-grade inputs and food-grade would be far lower — which is correct, but concedes that the environmental case depends on a premise not yet achieved.
⚠️ The specific multiple and the peer-review status of that LCA could not be verified first-hand.
The two subjects of this report met on 28 March 2023. And nobody ate the result.
The facts (Smithsonian Magazine 2023-03-30; Wikipedia, Vow (company)):
| Date | Tuesday 2023-03-28 |
| Unveiled at | NEMO Science Museum, Amsterdam (later displayed at Rijksmuseum Boerhaave, Leiden) |
| Made by | Vow, the Australian cultivated meat company that later built the 22,000 L "Andromeda" reactor |
| Target protein | Mammoth myoglobin — the protein that determines meat's colour and much of its flavour |
| Sequence gaps | filled using African elephant genetic data, the mammoth's nearest living relative |
| Expression host | Sheep cells, grown to 20 billion cells |
| The object | Between a softball and a volleyball in diameter; slow-roasted, then blowtorched |
| Who ate it | Nobody |
Why nobody ate it — Vow's Chief Scientific Officer James Ryall, verbatim:
"I've got no idea what the potential allergenicity might be of this particular protein."
He noted the protein "has not existed for 5,000 years". That is an honest and deeply awkward reason: you built a molecule no human immune system has ever encountered, and then found you were not willing to eat it.
Co-founder Tim Noakesmith gave a purely narrative purpose: to "create a symbol of a more exciting future that is better for us and better for the planet". The site mammothmeatball.com is still live, with the tagline "Let's eat ourselves out of extinction". Wikipedia's characterisation of the whole thing is three words: a publicity stunt.
Why this case deserves its own section:
- It shows that de-extinction's most realistic product may not be an animal but a protein. Reviving a mammoth requires surrogacy, 22 months of gestation, and hundreds of edits. Expressing mammoth myoglobin requires a sequence and an expression system. The difficulty differs by an unquantifiable number of orders of magnitude, and the narrative effect is nearly identical.
- It exposes the shared structure of these projects precisely: a real technical core — ancient protein reconstruction plus heterologous expression are both genuine — attached to an end promise that never lands. Nobody ate it, and the methodology was never published.
- It was not even the first time. In 1951 the Explorers Club in New York held a famous "mammoth dinner". In 2016 Glass et al. genetically tested a surviving sample and concluded it was green sea turtle (Was Frozen Mammoth or Giant Ground Sloth Served for Dinner at The Explorers Club?, PLOS ONE 11(2): e0146825). The authors' characterisation was the same: a publicity stunt. — Two "eat a mammoth" publicity events, 72 years apart, and neither involved eating any mammoth.
Between those two events, in 2011, the Chinese palaeontologist Xing Lida — now an associate professor at China University of Geosciences (Beijing) — ate mammoth meat on a live stream, to some controversy.
On the meat's origin, per Shanghai's SHINE (Tan Weiyun, 2018-12-12):
"It was from a leg preserved in Siberian permafrost for about 4,000 years. Xing cooked it thoroughly and flavored with salt, but told his online audience that it still 'tasted bad, weird and coarse, like soil.'"
The same piece describes him as the first person in China to eat 4,000-year-old mammoth and drink the resulting broth.
The "4,000 years" is worth pausing on. Mainland mammoths died out around 10,000 years ago at the end of the Pleistocene, but two relict populations survived far longer: St. Paul Island, Alaska, to about 5,600 years ago, and Wrangel Island in the Russian Arctic to about 4,000 years ago — the last known mammoths anywhere. So "Siberia + about 4,000 years" most likely points to the Wrangel Island lineage. For scale: when the last mammoths were still alive, the Egyptian pyramids had already been standing for centuries.
Incidentally, Xing's own comment on reviving dinosaurs in that same piece matches the argument in section 1.3:
"There is no possible channel with current technology... amber is at more than 100 million years old, so we cannot revive dinosaurs today."
📌 Citation correction: the English Wikipedia article citing this SHINE piece gives the publication year as 2011. The actual publication date is 2018-12-12; 2011 is when the meal happened, not when the article appeared.
There is more, and it is heavily concentrated in one episode.
This episode carries nearly all the solid biology in the set, and is the only one besides the futuristic-tech episode worth discussing in its own right.
1. Cheetah population data (conservation biology)
"today there are less than 8,000 cheetahs living on earth" / "when the first car was made more than a century ago, over 100,000 cheetahs were alive in the wild"
From 100,000 to under 8,000 in a century. And cheetahs connect directly to Part 1, because they are the textbook case of a genetic bottleneck — the species passed through an extreme bottleneck in the last glacial period, and diversity is now so low that skin grafts between unrelated individuals are not rejected. That is close to unique among wild mammals, and it makes the cheetah a paradigm target for genetic rescue and biobanking.
2. Thermoregulation (the video gets an important mechanism right by accident)
"before this cheetah races a Formula E car, we need to prevent it from overheating"
Not showmanship. A cheetah at full sprint runs on anaerobic metabolism, generating heat far faster than it can shed it, and can hold top speed for only 20–30 seconds. That is the real reason cheetahs are sprinters rather than endurance hunters — not muscle, heat.
3. Purring (felid anatomy)
"The cheetah is purring insanely loud."
Cats divide by hyoid ossification into those that roar (Panthera: lion, tiger, leopard, jaguar) and those that purr continuously. Cheetahs are the latter — a big cat that is physiologically closer to a house cat.
4. Search dog olfaction (with a caption error)
"How can dogs smell this good? It's called their olfactors. The longer the snout, the more olfactors that these dogs have."
5. Elephant painting — an excellent anthropomorphism trap
"This elephant is the best painting elephant on the planet."
The mainstream view in animal cognition is that elephant painting is a repeated motor pattern guided by subtle handler cues, not spontaneous aesthetic creation. It is the best available material for discussing how we over-read animal behaviour — and it rhymes with Part 1. We project what we want to see onto animals very readily.
6. Checkable data points: a 500 lb lion, 1,500 lb bulls.
"This is Black Panther. Modeled after the cheetah, it can reach speeds over 22 miles per hour."
This pairs perfectly with the cheetah episode:
| Speed | |
|---|---|
| Real cheetah | 60–70 mph |
| Cheetah-inspired robot | 22 mph |
Nearly a threefold gap — and that is exactly what makes biomimetics worth discussing. A cheetah's speed comes from spinal flexion storing energy like a spring, elastic recoil in the tendons, semi-retractable claws for grip, and a long tail as an inertial balance rod. Those mechanisms rely on compliant, variable-stiffness biological materials; the robot uses rigid links and motors.
Between copying and copying well lies materials science.
1. Cataract surgery (2023-01-28) — the highest science content of the three
"the surgeon uses a tiny vacuum to suck up the clouded lens... replace it with an artificial one" "half of all the blindness in the world is people who need a 10-minute surgery"
That is phacoemulsification plus intraocular lens implantation. Lens clouding happens because crystallin proteins denature and aggregate with age — those proteins barely turn over and are among the oldest in the body.
But the striking part is Jeremiah's case:
"Because Jeremiah was born with cataracts, his right eye has never received light. Which means this surgery only has a 50% chance of working."
That goes straight to the critical period in visual neuroscience. Hubel and Wiesel won the 1981 Nobel for monocular deprivation experiments: if one eye receives no visual input in early development, the corresponding connections in visual cortex never form properly, and restoring the optical path later does not restore sight (amblyopia). Behind "only a 50% chance" sits an entire theory of neural plasticity.
2. Prosthetics (2025-01-11)
"I was in a car wreck when I was 17, had osteomyelitis, and that's how I lost my feet."
Osteomyelitis — bacterial bone infection — is a common cause of amputation. The detail about a global inventory of recycled components is also interesting, touching on the trade-off between socket fit and gait biomechanics.
3. Hearing devices (2023-05-06)
The biosphere in the later part of the video is Biosphere 2 in Arizona. It is worth covering because, like de-extinction, it is a case where the real story is better than the packaging.
| Claim | Fact | Verdict |
|---|---|---|
| "half a billion dollar biosphere" | Built for $150 M (1987–1991, funded by Ed Bass) | ❌ overstated — even inflation-adjusted to 2026 it is about $350 M |
| "five different biomes" | 7 areas in total, of which exactly 5 are wilderness biomes: rainforest, ocean with coral reef, mangrove wetland, savanna, fog desert | ✅ correct — the other two are agriculture and habitat |
| "these plants have been living in here, some of them, for over 30 years" | Sealed in 1991, so 35 years by 2026 | ✅ correct |
First closure mission (1991–1993), 8 people sealed in for two years:
- The agricultural system supplied only 83% of required food; the crew was chronically hungry
- Oxygen fell from 20.9% to 14.5% within 16 months — equivalent to above 4,000 m altitude, and the crew developed serious hypoxia symptoms
- The strangest part: CO₂ did not rise correspondingly. Where did the oxygen go?
- The answer: CO₂ reacted with exposed concrete in the structure to form calcium carbonate, fixing carbon and oxygen together. A pure, entirely unanticipated building-materials chemistry problem nearly destroyed the mission
- The crew split into two factions
The second mission (1994) was stranger still: begun in March; in April investor Ed Bass brought in Steve Bannon — yes, that one — to take over management; on 5 April former crew members opened the airlocks from outside, exchanging around 10% of the air; terminated early in September.
Why the failure is more valuable than a success would have been. Biosphere 2 remains the largest closed ecosystem humans have ever built, now operated by the University of Arizona. Its central lesson is that the failure mode of a closed life-support system tends to come from the variable you never modelled. That is invaluable for Mars habitat design — and it is what the video is gesturing at without saying.
Incidentally, the hydroponics in the bunker episode (2025-08-30) belongs to the same lineage: soilless cultivation with closed-loop nutrient solution, the food module of every closed life-support system.
The $1 vs $10,000,000 Job! episode (2023-11-25) includes a dinosaur fossil dig:
"because this bone is so ancient, we have to cover it in a cast before we pick it up. Or else it could literally break in half."
That is a field jacket, standard palaeontological practice.
But the thing worth pointing out: non-avian dinosaurs died out 66 million years ago, and the theoretical ceiling for recoverable DNA is 0.4–1.5 million years. More than forty times over. There is no DNA in those fossils, only bone structure replaced by minerals.
So: Colossal can work on mammoths (43,000 years) and dire wolves (13,000 years), and will never work on dinosaurs. Putting that comparison side by side is clearer than either half alone.
- Clean water (2025-08-01): mentions E. coli and Giardia — one a bacterium, one a protozoan parasite, with entirely different pathogenic mechanisms. A good entry point for microbial classification
- Pyramids (2025-02-08): palaeopathology (stress markers on workers' skeletons), confined-space hypoxia
- Nothing relevant: the underground city, the plane tickets, and the $100 million car episodes
Superficially one is about reviving dinosaurs and the other about growing meat in a lab. Underneath they are the same thing.
Lay the two toolboxes side by side and the overlap is striking:
| Technology | In de-extinction | In cultivated meat |
|---|---|---|
| Cell culture | elephant / wolf / thylacine cell lines | cattle / chicken / fish cell lines |
| iPSC / stem cells | turning somatic cells into gamete-competent cells | turning somatic cells into indefinitely proliferating seed cells |
| Multiplex CRISPR | writing in extinct-species traits | immortalisation, improved metabolic efficiency |
| Bioreactors | expanding edited cells | producing meat |
| Cryopreservation / biobanking | the biovault | master and working cell banks |
| Artificial womb / ex utero development | avoiding surrogacy | (not needed yet) |
They are two outlets of the same technological shift: treating life as a readable, editable, manufacturable engineering object. That shift is called synthetic biology, and its two branches — conservation biotechnology and cellular agriculture — differ only in the application they chose.
Both sides show the same pattern:
- Press release first, paper later — or never
- One visually arresting result standing in for an entire technical programme — the woolly mouse, the blind-tasted sandwich
- Skipping the inference between "we did X" and "X implies Y" — editing 12 genes ≠ being able to make a mammoth; making one piece of meat ≠ making cheap meat
- Avoiding the question of scale — three wolves is not a population; 400 packs of dog treats is not an industry
This is not a moral accusation, it is structural. Both sides are venture-backed private companies that must keep telling a story to keep raising, while doing the kind of long-horizon basic research public science normally carries. Funding cycles run 18–24 months; these technologies run on 10–30 year timescales. The narrative has to run ahead of the data, or the money stops.
Understand that and you can debunk any "future technology" story yourself.
| Colossal (de-extinction) | Cultivated meat | |
|---|---|---|
| 2021 | just founded, $15 M seed | funding peak of $1.3 B |
| 2026 | $10.2 B valuation, acquiring companies, federal partnership | $73.9 M, down 96%, the leading firm bankrupt and at auction |
Why?
One reason, in my view: Colossal does not have to compete in a commodity market, and cultivated meat does.
- Colossal sells a story, technology assets, government relationships and conservation services. Its customers are a small number of large investors, government agencies and philanthropic capital. It never has to prove a mammoth can be produced below some price. Its product does not even need to exist — an edited wolf that makes the cover of Time has discharged its commercial function.
- Cultivated meat sells chicken, and chicken wholesales at $2–3/lb — one of the most thoroughly optimised commodities in human history, behind which sit a century of breeding, feed conversion optimisation and slaughter automation. You must compete at that price or you have no market. €40/kg against $5/kg does not close because the story is good.
That generalises into a rule. When assessing any "future technology", ask first: does it ultimately have to compete on price in a mature commodity market?
- If not — pharmaceuticals, defence, space, luxury goods, government procurement — then technical difficulty and capital patience are the main variables, and a good story sustains it for a long time.
- If so — food, energy, bulk materials — then anything an order of magnitude more expensive than the incumbent dies, however advanced and however morally correct.
The tragedy of cultivated meat is that its moral argument (200 million chickens a day) is unimpeachable, its technical progress (media at 7 cents a litre) is real, and it picked the wrong arena.
Everything it is doing now — pivoting to pet food (higher margin tolerance), cosmetics (IntegriCulture's actual profit source), pharma supply (Upside's Lucius Labs), hybrid formulation (10% cultivated cells) — amounts to escaping that commodity market.
Not as frauds. The opposite:
De-extinction. As species revival it has failed and cannot succeed by definition. As an accelerator for conservation biotechnology it has real value. Elephant iPSCs, the EEHV mRNA vaccine, artificial womb prototypes, large-scale biobanking — these will outlast the programme and will be used on genuinely endangered species. Its biggest risk is not technical failure but that it is institutionalising the false idea that extinction is reversible.
Cultivated meat. As a replacement for animal agriculture it is unrealistic, at least for the foreseeable future. As a technology for high-value niches — pet food, cultivated fat as an ingredient, cosmetics inputs, space and polar food, meat substitutes for endangered species — it has real room. Its greatest value may not be in food at all, but in having pushed down the cost curve for cell culture generally: media from $1,000/L to $0.07/L is a result that benefits cell therapy, regenerative medicine and biopharma.
The shared lesson: to judge a technology, do not look at the problem it claims to solve. Look at the conditions under which it can actually survive.
Parts 1 to 4 answer "which of these claims are true". This part answers a different question: is there a number everyone should be reporting that nobody is?
There is, and more than one. Everything below comes from real sequences in public databases. No figure is invented, and every one can be recomputed by a script in this repository.
| Package | Scope | Core mechanism |
|---|---|---|
chimaera |
composition itself | five typed operators (blend / express / edit / graft / cross), each with independent feasibility constraints |
walghvogel |
flavour channel | a corpus of historical testimony with negative controls, across seven trait channels |
tzeentch |
morphology channel | developmental competence gating — the talpid² chicken-tooth route has to be expressible |
As Part 1 argued, "fusing A and B" is not one operation but five. Separating them produces a result immediately:
You can combine a pig and an apple — 1.53 billion years apart — far more easily than a baboon and a sheep, 94 million years apart. As long as you do not do it genetically.
That is not a witticism. It is why every cultivated meat product on the market
is a blend at 3–20% cells: blend is the only operator whose feasibility barely
decays with divergence.
Vow's 2023 meatball used mammoth myoglobin with "gaps filled from African elephant". Vow never said how large the gaps were. Against African elephant — the filler they named — the decomposition is:
| Share | |
|---|---|
| Distinctively mammoth | 2.60% — 4 residues: E27L, F28E, L30F, E84Q |
| Ancient-supported but identical to elephant anyway | 79.87% |
| No mammoth data at all; pure elephant | 17.53% |
Against Asian elephant it is 1.95% with 3 substitutions. Both are correct; they
answer different questions. Quote the African elephant column when discussing
Vow, because that is what Vow actually did. The corpus table in
provenance_results.json stores the Asian pair, so the two will not match, and
are not meant to.
One thing confirmed along the way: exactly one mammoth myoglobin exists in the entire public protein record — a 127-residue fragment (UniProt R9RZK4 = NCBI AGM75745.1), missing 27 residues that constitute the whole C-terminal third. There is no better version available.
Median distinctive fraction 1.84%. This is the largest publicly available provenance catalogue of extinct-species proteins found in this work.
Regressing distinctive fraction against TimeTree divergence gives r = +0.794 for nuclear genes and r = +0.617 for mitochondrial (five species pairs each).
NADH5 and NADH6 are in fact mitochondrial ND5 and ND6; counted as nuclear
they placed a 10.80% point at 25.4 Ma, which is nothing like a nuclear
substitution rate, and it dragged the whole trend apart. The nuclear
compartment is the one carrying the main conclusions, and it went from barely
correlated to strongly correlated. The improvement was not sought — it fell
out of listing the corpus gene symbols to set up the human–chimpanzee control.
The harder result is the negative control: the aurochs, 0.05 Ma (54,000 years) from cattle, returns exactly 0.00% distinctive on both nuclear proteins, ZFX and ZFY. Insufficient divergence yields nothing detectable. The method does not manufacture signal.
De-extinction coverage always gives the numerator — "45 genes edited" — and never the denominator. Fitting a substitution rate across 43 nuclear observations gives λ = 5.7 × 10⁻⁴ per residue per Ma, which sits inside the empirical range for protein evolution. Nothing in the fit was told to land there, making it a free consistency check.
From that: Colossal's announced 45 edits correspond to a completion of 0.32%–0.50% (precisely, 0.321%–0.495%).
A sequence listing is part of a legal claim, so a listing naming an extinct species asserts something about that species. Twelve extinct taxa were surveyed; one genuine hit.
US 12478079 B2, granted 2025-11-25, is a cheese patent containing mammoth casein. The same document also contains elephant casein. The two differ by not a single amino acid — and casein is an intrinsically disordered, weakly constrained protein that differs noticeably even between close relatives. Zero substitutions across 6.4 million years is not what divergence looks like.
The two aurochs hits are artefacts, recorded rather than deleted: the
peptides begin GRKKRRQRRR, the HIV-1 Tat cell-penetrating tag, making them
engineered constructs — and Bos primigenius also denotes domestic cattle.
The dire wolf has zero protein sequences in UniProt, NCBI and the patent
literature alike.
This generalises the FGF5 finding, and is the strongest result here.
Colossal edited 12 genes in the woolly mouse; FGF5 is one of them. Across four proboscidean species, FGF5 is 271 residues with exactly one variable site in the whole family — and the mammoth carries the majority allele. In other words: no mammoth-specific FGF5 protein variant exists to copy.
How common is that? 12 of 43 nuclear proteins in the corpus are identical. But identity is not itself a finding — 27 residues across 6.4 Ma will match by arithmetic alone. So each null is scored against the odds it faced:
| Species | Gene | Length | Ma | P(identical) | |
|---|---|---|---|---|---|
| Steller's sea cow | CNR1 |
396 | 13.3 | 1.3% | finding |
| Steller's sea cow | EDG1 |
328 | 13.3 | 2.8% | finding |
| Mammoth | FGF5 |
271 | 6.4 | 38.5% | finding |
| Mammoth | IRBP |
27 | 6.4 | 90.9% | underpowered |
| Aurochs | ZFY |
801 | 0.05 | 97.6% | underpowered |
Extrapolated to whole proteomes, the two flagship programmes fail in opposite directions:
| Target | Divergence | Share with nothing to copy | Genes |
|---|---|---|---|
| Mammoth → Asian elephant | 6.4 Ma | 41% (30–55%) | ~8,265 |
| Thylacine → dunnart | 44 Ma | 6% (3–11%) | ~1,183 |
The mammoth is close enough to an elephant that a large part of the genome offers no distinctive residue at all — the edit list is short because there is little to change. The thylacine is far enough from a dunnart that nearly every gene differs, so its list is not short; it is unmanageable. Neither is the tractable middle a headline implies.
That 41% cannot be checked inside the corpus. The corpus produced it, and there is no second mammoth.
But the divergence can be matched. TimeTree 5 puts human–chimpanzee at 6.4 Ma — the same median it gives mammoth–Asian elephant. Two living species, complete proteomes, no ancient DNA, no imputation, no fragments, and not a single primate anywhere in the fit.
The gene set is the corpus's own nuclear genes, so it could not be chosen after seeing the answer.
| Predicted identical | 26.3% |
| Observed identical | 25.7% — 9 of 35, 95% CI 14.2%–42.1% |
| Published, all orthologues | 29% (Nature 437:69, 2005) |
A rate fitted entirely on degraded ancient sequence from elephants, sirenians and marsupials predicts a living primate pair to within 0.6 percentage points. The measurement also agrees with the published figure, which independently checks the alignment code.
The mechanism is visible with nothing tuned to this pair: short proteins come
back identical, long ones do not, and the exceptions are ones evolutionary
biology already names. Protamine PRM1 is 51 residues with 8 substitutions —
among the fastest-evolving proteins known. RUNX2 is 569 residues and identical
against 12% odds — a skeletal transcription factor under heavy constraint.
The control was also used to adjudicate a modelling choice, and could not. Length in the fit is the proxy's full length, but a substitution is only visible where the ancient fragment covers the protein, so covered length is strictly more correct. Median nuclear coverage is 93%, mean 81%, so the two differ.
| Convention | λ | Predicts | Error | Shift on dropping 2 low-coverage entries |
|---|---|---|---|---|
| Full length (default) | 5.7e-4 | 26.3% | +0.6 pts | −2% |
| Covered length | 6.8e-4 | 22.6% | −3.1 pts | −7% |
Both sit inside the observed interval, so 35 genes cannot decide it. The
control leans toward full length, and covered length leans harder on the two
least trustworthy observations — mammoth VWF compares 19 of 411 residues and
finds 2 substitutions. Full length also yields the weaker form of this
project's own claim (41% rather than 47%), so it stays the default. A larger
control could settle this; this one cannot.
Listed separately, because this says more about how far the numbers should be trusted than the numbers themselves do.
-
Ambiguity codes were counted as substitutions. Seven of the sea cow HBB's twenty-four apparent substitutions were the letter
X. Found while sanity-checking an outlier.Xnow counts as imputed, not distinctive. -
Rate limiting silently destroyed a dataset. A throttled run overwrote a 67-pair index with 21 pairs and raised no error. There is now retry with backoff, per-species failure status, and a writer that refuses to shrink the dataset without
--force. -
Compartment classification was wrong in both directions at once.
ATP7is "ATPase Cu++ transporting alpha polypeptide" — ATP7A, a nuclear copper transporter — which theATPprefix called mitochondrial and dropped from the nuclear analysis.NADH5andNADH6are mitochondrial ND5 and ND6, but neither begins withND, so both were counted as nuclear — and mitochondrial genes evolve several times faster, inflating the fitted rate.The two errors ran opposite ways and partly cancelled, which is how they survived: fixing them moved every headline number by only 2–3% (λ from 5.9 to 5.7 × 10⁻⁴; the mammoth's nothing-to-copy share from 42% to 41%). The conclusions were robust to the bug, but that could only be known afterwards. The classifier is now a closed set of the thirteen mitochondrial protein-coding genes plus aliases — prefix rules will always catch nuclear-encoded ATPases (ATP1A1, ATP7B) and complex I subunits (NDUFA1) by accident, and a closed set should be enumerated.
-
The
crossoperator missed its own calibration point by a factor of three. Its anchor text read "horse × donkey (~4 Ma) ... halving at 4 Ma puts the mule at the midpoint". Horse and donkey are 11.1 Ma apart, so the mule scored 0.146 rather than 0.5 — the model called the most-bred hybrid in history barely feasible, and a registered commercial breed implausible. Caught by scoring the operator against eleven crosses that have actually been carried out. Re-checking every other anchor then found the same fault ingraft(rat × mouse is 13.1 Ma, not "~20"). Every operator that decays with distance now declares its anchor pair as data, with a test asserting the half-life equals that pair's divergence. -
The FGF5 inference was overstated — the most important one. Comparing only the mammoth–Asian elephant pair, it appeared the mammoth carried a phylogenetically diagnostic state, which would have shown the sequence was genuinely read rather than imputed from an elephant. Adding the fourth sequence destroyed that argument: the mammoth carries the majority allele, which diagnoses nothing. Corrected, with the failed argument preserved in
fgf5_proboscidea.yamlrather than deleted — and a test asserting the record still contains it.
projects/chimaera/ composition engine, provenance, saturation model, patents
projects/walghvogel/ flavour channel, worked examples, illustrations
projects/tzeentch/ morphology channel, developmental competence gating
Each package README carries the full results, the method, and the places its conclusions do not hold.
| Claim | Finding | Verdict |
|---|---|---|
| "the only two dire wolves in the universe" | There are three — Romulus, Remus, Khaleesi | ❌ wrong |
| "We will have mammoths this decade" | Company target is 2028; an edited Asian elephant is perhaps possible, an actual mammoth is not | |
| Colossal "spent over $400 million" | $435 M raised as of 2025-01, not spent | |
| bio-vault "over a thousand species" | Could not be verified from any public source | ❓ unverified |
| Mammoth tusk "43,000 years old from Alaska" | Consistent with the field's sample age range, and far below the aDNA ceiling | ✅ plausible |
| Woolly mouse "mixed its DNA with that of a mouse" | Actually CRISPR editing of 12 genes, only 2 with direct mammoth counterparts | |
| Upside "hundreds of thousands of pounds" of chicken | EPIC claimed 50,000 lb/year (400,000 lb ceiling); Wired found production relied on roller bottles; GFI calls the sector "almost entirely pre-revenue" | ❌ highly doubtful |
| "every two seconds a chicken's worth of meat" | Implies 20 M lb/year, 50× the claimed full-production ceiling | ❌ fails on order of magnitude |
| Upside "over $200 million in R&D" | About $608 M raised in total | ✅ plausible, if conservative |
| "220 million chickens killed every single day" | 70–75 billion/year globally = 190–210 M/day | ✅ right order of magnitude |
| Not mentioned in the video | Upside has had zero sales since 2024; Lucius Labs was founded 2026-01-23, the day before broadcast, pivoting to sell media | |
| Biosphere "half a billion dollar" | Built for $150 M (1987–1991) | ❌ overstated ~3× |
| Biosphere "five different biomes" | Exactly 5 wilderness biomes among 7 areas | ✅ correct |
| Plants "living in here for over 30 years" | Sealed 1991, so 35 years by 2026 | ✅ correct |
- The "over a thousand species" figure for Colossal's biovault
- The 2026 source revising dire wolf divergence from 5.7 to 4.5 Ma
- Whether the woolly mouse preprint has been formally peer-reviewed
- Allentoft 2012, "DNA half-life of 521 years", original text (Royal Society 403)
- The specific "~$37/kg" figure in Humbird's paper (Wiley 403)
- The specific multiple and peer-review status of the UC Davis LCA preprint
- The outcome of Upside Foods v. Florida (ij.org 403)
- Details of the 2025 Montana, Indiana, Nebraska and Texas measures
- The final result of the Believer Meats asset auction (auction 2026-08-17, hearing 2026-08-20; no public report as of 2026-08-21)
- Whether GOOD Meat is still selling in Singapore
- Whether Aleph Cuts ever went on genuine commercial sale in Israel
- The accurate 2021 cultivated meat funding peak (three sources give $1.3 B / $1.8 B / $989 M)
Books
- Beth Shapiro, How to Clone a Mammoth: The Science of De-Extinction — start here. The author is now Colossal's CSO, but the book predates that, and is restrained and technically solid. Its central argument is that you cannot clone a mammoth.
- Svante Pääbo, Neanderthal Man: In Search of Lost Genomes — a first-person account of Nobel-winning work, and of how ancient DNA went from ridiculed to mainstream.
Essential papers
- Perri et al. 2021, Nature — dire wolf phylogeny, the scientific basis of the entire dispute
- van der Valk et al. 2021, Nature — million-year mammoth genomes
- Kjær et al. 2022, Nature — two-million-year environmental DNA, the current record
Who to follow
For: Beth Shapiro, George Church, Andrew Pask, Ben Lamm. Against: Victoria Herridge, Nic Rawlence, Vincent Lynch, Jeremy Austin, Philip Seddon. Follow both — the most valuable information in this field is generated where the two sides meet.
Institutions: IUCN SSC, particularly Canid Specialist Group statements; Revive & Restore, a de-extinction non-profit considerably more restrained than Colossal; San Diego Frozen Zoo.
Essential
- David Humbird 2021, Biotechnology & Bioengineering, "Scale-up economics for cultured meat" — if you read one thing, read this. The single most important sceptical document in the field.
- GFI's State of the Industry Report, published each April. GFI's position is supportive, but its data is the most systematic available and it reports bad news honestly.
- Mark Post's original papers, for the technical starting point.
Who and what to follow
- AgFunderNews and Green Queen — the best industry reporting, especially AgFunderNews' 2026-06-22 piece After the crash, who's still standing?
- New Harvest — non-profit, funds basic research, more objective than industry sources
- Sceptics: Paul Wood (former Pfizer animal health executive, among the most systematic critics), Derrick Risner (UC Davis, on LCA)
Regulatory tracking: Singapore SFA's Approved Novel Foods list; UK FSA sandbox progress (results due 2027-02); the FSANZ application register; the EFSA novel food database.
The most valuable angle on this material is not how impressive the technology is. That has been covered thoroughly, and covered more entertainingly than any report can manage.
What is scarce is an account of how the gap between what a technology is said to be and what it actually does gets created, how large it is, and how to measure it yourself.
The material here contains an unusually clean case. Two companies appear in the same video; one went on to a tenfold valuation increase while the other's entire industry fell 96%. One of them formally pivoted its business the day before broadcast. The video says there are two dire wolves; there are three, and they are not dire wolves.
None of that is error-spotting. It is a complete story about how we come to believe things about the future.