Biological Computing

Established “Biological computing” names three research programmes that share a slogan and almost nothing else. Established The first stores digital data in synthesised DNA and is a chemistry, coding-theory and capital-expenditure problem. The second computes with molecules — strand displacement, algorithmic self-assembly, molecular logic inside cells — and is a physical-chemistry and theoretical-computer-science problem. Frontier The third grows cortical neurons on electrode arrays and asks them to do something, and is a neuroscience problem with a commercial wrapper. Established They have no shared benchmarks, no shared conferences and no shared failure modes, and their binding constraints differ in kind: a synthesis fab that nobody will build, a diffusion limit that no engineering removes, and a replication that nobody has funded. Speculative A brief that treats them as one rising curve will be wrong about all three, so this one treats them as three curves.

The measured record is remarkable in density and dismal in throughput. Established DNA holds data at something like 687 terabytes per cubic millimetre; the only fully automated end-to-end storage device on record wrote and read the word HELLO — five bytes — in about twenty-one hours. Adleman’s founding computation solved a seven-node graph problem, solvable by inspection, in roughly a week. Frontier The best-known neural-culture result showed a statistically significant increase in average rally length between the first five minutes and the last fifteen of a twenty-minute session, and drew a formal published rebuttal in the same journal.

Established Two numbers in wide circulation are wrong, and both are corrected here with the arithmetic shown: the “700 terabytes per gram” density is the reported per-cubic-millimetre figure with the unit swapped, and the “million-fold energy advantage” of organoid computing compares a modelled brain to a real supercomputer while the only shipped wetware machine draws forty to a hundred times the power of an entire human brain. Speculative The exotic question — whether a general-purpose biological computer is a coherent engineering goal or a category error — is taken seriously in section 3, with both answers given their strongest form.

Established The storage demonstrations are real, they are small, and they are almost all one-off. Established Goldman and colleagues encoded 739 kilobytes of files — an estimated 5.2 × 10⁶ bits of Shannon information — into synthesised DNA and reconstructed them with 100% accuracy, published in Nature in February 2013; Church, Gao and Kosuri had encoded 5.27 megabits the previous year, a 53,400-word book, eleven images and a JavaScript program. Established Organick and colleagues pushed the scale two orders of magnitude in 2018: 35 files totalling over 200 megabytes into more than 13 million oligonucleotides, each recovered individually and without error by a random-access primer scheme. Established Erlich and Zielinski’s DNA Fountain encoded 2,146,816 bytes at 1.55 bits per nucleotide, which their own preprint calls “only 15% from the Shannon capacity of DNA storage,” against a stated capacity of at least 1.83 bits per nucleotide; a 2023 coding paper reports 1.78, closer still. On information-theoretic efficiency the field is essentially finished, and that is precisely why nothing left to solve is an information-theory problem.

Established The density figure everyone quotes is right in its numeral and wrong in its unit, and the arithmetic is worth doing on the page. Established The Church demonstration is reported as 5.5 petabits per cubic millimetre. 5.5 petabits is 5.5 × 10¹⁵ bits, which is 6.875 × 10¹⁴ bytes, which is 687.5 terabytes — per cubic millimetre. Established The folk figure “about 700 terabytes per gram” is therefore the paper’s own numeral with the unit changed from a volume to a mass. Established Convert it properly: dry DNA has a density of roughly 1.1 to 1.7 grams per cubic centimetre, so one cubic millimetre weighs about 1.1 to 1.7 milligrams, and 687.5 terabytes in 1.1 to 1.7 milligrams is roughly 400 to 625 petabytes per gram — the same order as the 215 petabytes per gram Erlich and Zielinski independently realised, and what the DNA Fountain preprint means when it says in its own words that DNA “offers tantalizing information density of petabytes of data per gram.” Established A second cross-check settles the reading. An independent 2023 paper puts the medium’s theoretical ceiling at approximately 10¹⁹ bits per cubic centimetre, and 5.5 petabits per cubic millimetre is 5.5 × 10¹⁸ bits per cubic centimetre — a little over half the ceiling, exactly where a good demonstration should sit. Read per gram, the same numeral would put the demonstration three orders of magnitude below the ceiling, which no author reporting a record would do. Established One honest caveat: the Church paper itself could not be opened for this brief, and the 5.5-petabit figure is quoted verbatim from a reference article rather than from the paper. The correction does not depend on it — it rests on the numeral identity between 687.5 and “about 700,” on Erlich’s independent per-gram figure, and on the ceiling cross-check, all three read directly.

Established The economics, not the chemistry, decide this field, and the gap is eight to nine orders of magnitude. Established As reported for the Goldman scheme, writing cost about $12,400 per megabyte and reading about $220 per megabyte in 2013 prices; Erlich and Zielinski’s two-megabyte payload cost about $7,000 to write and $2,000 to read in 2017. Those come to roughly $1.24 × 10¹⁰ and $3.5 × 10⁹ per terabyte written, against LTO tape at the order of $10 per terabyte. Frontier No primary or industry figure for oligonucleotide synthesis cost per base later than 2017 could be found for this brief, which is itself a finding: every cost comparison in circulation rests on a decade-old price. Established The reason writing is expensive is structural. Reading DNA is a massively parallel measurement problem and its cost fell faster than Moore’s law for a decade; writing is serial chemistry in which per-base coupling efficiency compounds over length, so at 99.5% efficiency the full-length yield of a 200-mer is 0.995²⁰⁰, about 37%. Established That asymmetry is why the “sequencing costs are collapsing, therefore storage costs are collapsing” argument fails: the curves run on different physics and only one has moved.

Established Throughput is worse than cost, and the number that calibrates the field is five bytes. Established Takahashi, Nguyen, Strauss and Ceze built the first fully automated end-to-end DNA storage device in 2019; the payload was the word HELLO and the write-store-read cycle took approximately 21 hours. Density and throughput are independent axes, and this field sits at opposite ends of both. Established What has genuinely progressed is coding: fountain codes, Reed–Solomon outer codes and Grass and colleagues’ sol–gel silica encapsulation together turn an error-prone medium into a recoverable and archivally stable one. Frontier Coding is also the cheapest lever on cost, because a code tolerating an order of magnitude more synthesis error lets you buy worse chemistry — and worse chemistry is the only chemistry that will ever be cheap.

Established DNA computing began with a demonstration of principle that was immediately shown not to scale. Established Adleman’s 1994 experiment solved a seven-node Hamiltonian path instance in roughly a week in about a hundred microlitres of solution; by 2002 his group had solved a twenty-variable 3-SAT instance. A seven-node instance is solvable by inspection, and the point was that molecules could compute at all, not that they computed well. Established Two years later Boneh, Dunworth, Lipton and Sgall published the objection that has governed the field since: brute-force molecular search buys an enormous constant-factor parallelism, but the mass of DNA required still grows exponentially in problem size, so somewhere around sixty or seventy variables the reaction needs more DNA than the Earth contains. Established Molecular parallelism does not defeat exponential scaling; it postpones it by a constant. The serious response was to change the question rather than argue, and the field moved from “solve SAT” to “implement algorithms” — algorithmic self-assembly, in which a tile set grows a computed pattern such as the Sierpinski triangle, is the canonical pivot.

Established What molecular computing is actually good at is small, slow, embedded logic. Seelig, Soloveichik, Zhang and Winfree showed in 2006 that AND, OR and NOT gates can be built from nucleic-acid hybridisation alone with no protein enzymes; Qian and Winfree scaled strand-displacement seesaw circuits to a four-bit square-root circuit in 2011. Established Cherry and Qian’s 2018 winner-take-all network sorted 100-bit patterns into nine classes, each pattern presented as twenty distinct DNA molecules drawn from a set of a hundred, and classified correctly with up to thirty bits flipped — a real pattern-recognition result in a test tube, at roughly 2.5 hours of kinetics per answer. Established Soloveichik, Seelig and Winfree supplied the theoretical backbone — DNA can implement arbitrary chemical reaction networks, which makes it a general analogue substrate on paper. Speculative The strongest live argument for the field is not speed at all: Benenson and Shapiro’s 2004 molecular automaton diagnosed a transcriptional signature and released a therapeutic oligonucleotide, inside a droplet, with no instrument. A DNA circuit can be where the molecules are — in a cell, in blood, in the field — and silicon cannot. Handwave On that reading the right benchmark is decisions per molecule of analyte in situ, not operations per second.

Frontier The cultured-neuron result is narrower than its reputation and better controlled than its critics allow. Frontier Kagan and colleagues cultured roughly 800,000 cortical neurons — a mixture of human iPSC-derived and mouse cells — on a high-density multielectrode array, embedded them in a simulated Pong environment, and reported that average rally length increased significantly between the first five minutes and the last fifteen of a twenty-minute session. Established The controls were not trivial: media-only cultures, rest sessions, in-silico random-noise paddles, and three feedback conditions comparing structured stimulus against silent and no-feedback arms, with only the mouse and human cortical-cell groups showing the effect. The title claims the cultures “learn and exhibit sentience,” and the paper drew a formal published response in the same journal, Neuron, signed by thirty neuroscientists arguing the evidence does not support the claim. Thirty authors on a two-page correspondence is not a stray objection; it is close to a disciplinary position statement, and any account omitting it is one-sided. Frontier The framing the field has partly adopted since is reservoir computing: a neural culture is a high-dimensional nonlinear dynamical system and a trained linear readout does the classifying — brain-organoid reservoir computing has been demonstrated on speech recognition on exactly that model.

Frontier The infrastructure is real; the scale is not. Frontier FinalSpark’s Neuroplatform is a remotely accessible wetware rig with genuine operating numbers: more than 1,000 brain organoids over three years, lifetime extended from hours to up to 100 days at best, more than 18 terabytes of data and over 20 billion recorded action potentials — on 32 electrodes in total. That is an access platform, not a computational result, and no published task performance from its external users was found. Established The reference organoid in the field’s own 2023 manifesto is under 500 micrometres across with fewer than 100,000 cells, and the manifesto’s target of 10 million neural cells is still four orders of magnitude short of a mouse brain. Frontier Cortical Labs’ CL1, announced at Mobile World Congress on 2 March 2025 at about US$35,000, is the first commercially sold biological computer; it draws 850 to 1,000 watts and keeps the neurons alive for up to six months. Established Two passes of source-gathering found no independent benchmark of it; the nearest thing is a vendor comparative claim in a press release qualified as holding “when learning opportunities were constrained to real-world timescales” — a handicap chosen by the party being compared. There does not appear to be a published independent benchmark of any commercial wetware computer, and that absence is the single most important fact about the product category.

Frontier The live question in storage is whether anyone can escape base-by-base writing. Two routes are open. Frontier Template-independent enzymatic synthesis using terminal deoxynucleotidyl transferase — demonstrated terminator-free for digital storage by Lee, Kalhor, Goela, Bolot and Church in 2019 — replaces phosphoramidite chemistry with an enzyme, removing the organic-solvent and coupling-efficiency structure that makes long oligos expensive. Frontier The other is to stop writing bases at all: CATALOG’s combinatorial assembly encodes information in which pre-made DNA parts are joined rather than in which bases are added, with an associated claim of a roughly one-megabit-per-second writer. That throughput number is the most-quoted figure in the field and this brief could not confirm it against a peer-reviewed source; the underlying work is a company-originated preprint. Speculative Neither route has shipped at archival cost, and both are engineering-limited rather than science-limited.

Speculative The most interesting unbuilt idea is a nanopore that writes. Nanopore sequencing already reads single molecules electrically as they thread a pore. Handwave If a pore could also modify a strand passing through it — addressable, base-specific chemistry under electrical control, at error rates inside the correction capacity of a fountain code — then write and read collapse into a single device and the entire cost structure of DNA storage changes at once. Handwave No demonstration exists and this brief found no paper claiming one; it is offered as a workback target, not a result, and it is the highest-leverage speculative object in the subject.

Frontier In molecular computing the frontier is the clock, not the gate. Frontier Diffusion sets the rate limit in solution-phase strand displacement, so the two live attacks are spatial localisation — tethering gates to a surface or an origami so reactants do not have to find each other — and enzyme-assisted displacement; the localised architecture of Chatterjee, Dalchau, Muscat, Phillips and Seelig is the reference result. Speculative Renewable or reversible gate designs attack the other structural problem: DNA gates are consumed on first use, so a circuit is single-shot unless the chemistry resets. Enzyme-free nucleic-acid dynamical systems — oscillators built from DNA alone — show the substrate can sustain autonomous dynamics rather than one-way cascades. Speculative None of this yet produces restorative logic, and without signal restoration circuit depth stays in single digits.

Frontier In wetware the frontier is whether the tissue does anything the decoder does not. Frontier The unresolved question about every neural-culture demonstration is whether the culture implements task-relevant credit assignment, or whether it is a fixed reservoir whose useful behaviour is produced entirely by the trained linear readout. Speculative The test is cheap in principle: compare the culture’s closed-loop performance against a linear readout trained on the same recorded activity with the feedback loop opened, and show the closed loop wins. Frontier Kagan and colleagues published a nomenclature paper in 2024 proposing agreed terms for diverse intelligent systems, defensibly read as a partial concession on the “sentience” framing and the most constructive thing to come out of the controversy. Speculative A live and under-discussed possibility is that the answer is no for cortical cultures and yes for engineered ones — that a culture with a designed reward pathway rather than borrowed spontaneous plasticity would behave differently — but nobody has built one.

Speculative The category question, stated as sharply as it can be. Speculative The strongest case that a general-purpose biological computer is a category error: a dish of neurons has no addressable memory, no instruction set, no separation of program from data, no clock and no mechanism for composing sub-results; what it has is a high-dimensional nonlinear response, which is the definition of a reservoir, and reservoirs are useful only because something else does the learning. On that reading “organoid intelligence” is reservoir computing with a wetter, costlier, shorter-lived reservoir. Speculative The strongest case for the other side: free-energy accounts predict that any system minimising surprise under a generative model will show goal-directed adaptation without an engineer specifying an objective, the Pong protocol is a direct test of that prediction, and it produced a differential effect against four control conditions. Handwave What would settle it is a comparison nobody has run — a neural culture, a DNA winner-take-all network and a two-dollar microcontroller on the same task at matched energy, latency and capital cost, with all three results published.

Established The three programmes have three different binding constraints, and only one is scientific. Established For archival DNA storage the binding link is industrial capacity: the chemistry works, the codes work, random access works, and what is missing is a synthesis fab. Established Nobody has built one because the addressable market for cheap oligonucleotides today — research reagents and gene synthesis — is satisfied at current prices, so there is no forcing demand to justify the capital expenditure, and there is no demand because there is no cheap synthesis. Frontier That is a chicken-and-egg problem of exactly the kind public procurement exists to break, and it is not a research problem at all.

Speculative For molecular computing the binding link is a scientific unknown: the diffusion floor on gate delay. Established In solution phase two molecules must find each other, and the time that takes is set by diffusion and concentration, not by any parameter an engineer controls. It is why benchmark strand-displacement circuits run in hours. Speculative Escaping it means abandoning solution-phase operation altogether — surface tethering, spatial organisation, or a physical mechanism that is not hybridisation — and that is a genuinely open physical-chemistry problem, not an engineering roadmap. Established This is a different binding link from storage, which is why treating biological computing as one field gives wrong answers about all of it.

Frontier For neural-culture computing the binding link is institutional: an independent, blinded, pre-registered replication that has not happened. There is no scientific obstacle to it. The obstacles are that high-density multielectrode rigs and iPSC-derived cultures are expensive, that some protocols are partly proprietary, and that no funder is paying for a replication whose most likely outcome is a null result embarrassing a commercially active laboratory. Frontier Until that experiment exists, every downstream claim in the field — scaling, energy efficiency, product benchmarks — rests on a single laboratory’s single protocol.

Established Two secondary constraints bind across all three programmes. The first is the absence of an interchange standard: an archive buyer cannot commit to a format with one supplier and no published specification, and this brief could not read the specifications of the industry body that exists to produce one. Frontier The second is the total absence of published watts-per-inference or dollars-per-operation accounting for any wetware system, which makes every efficiency claim in that literature unfalsifiable as stated. Speculative Neither is glamorous, and between them they are probably worth more than any result on the list.

Established What this subject waits on, in order of how much it would change. Established First, a cost curve for enzymatic DNA synthesis published head-to-head against phosphoramidite chemistry by a party that is not selling either. Everything in the storage workback plan is priced against figures from 2013 and 2017, and no later primary figure could be found for this brief, so the field is currently arguing about an economy nobody has measured this decade. Frontier Second, single-molecule electrical chemistry at a pore, which is where the write-read collapse would come from if it comes at all, and which currently exists only on the read side. Speculative Both are inputs this subject buys rather than produces, and neither has a customer other than this subject.

Frontier Third, from neuroscience rather than engineering: a mechanistic account of plasticity in dissociated cortical culture strong enough to predict, rather than describe after the fact, which stimulation protocols produce task-relevant weight change. Established Without it, wetware computing has no design rules and every result is empirical, which is also why replication is so brittle. Frontier Fourth, from stem-cell biology, vascularised organoids — the 500-micrometre ceiling is an oxygen-diffusion ceiling, and the manifesto’s ten-million-cell target is unreachable without perfusion.

Speculative Fifth, and least discussed: a coding-theory result establishing how much synthesis error a fountain-style architecture can absorb before recovery fails. Established That number sets the price of the chemistry the field is allowed to buy, and it is the one dependency here that could be settled by simulation plus a modest wet-lab sweep rather than by new capital. Speculative Notably, none of the five is a result any other brief on this map is currently producing, and three of them — the cost curve, the plasticity account and the error-tolerance bound — are the kind of unglamorous measurement that no laboratory advances its own career by making. Established This subject has to commission them itself or wait, and it has largely waited.

Established The storage chain, in order, with the measurement each link needs. Established One: enzymatic synthesis at greater than 99.9% per-base fidelity and 100-mer-plus length, measured as a full-length-yield-versus-length curve published head-to-head against phosphoramidite. Established Two: massive writer parallelism — more than 10⁶ independent synthesis features at under 10^-9 dollars per base, measured as dollars per base at the array and features per wafer, reported by an independent party rather than the vendor. Established Three: error-tolerant codes that permit sloppy chemistry, measured as file-recovery rate against synthesis error rate, swept. Four: random access holding at 10³ to 10⁴ times the pool size already demonstrated. Established Five: a read path that does not need a six-figure sequencer per site, measured in dollars per gigabyte read including consumables at 10^-3 residual error. Six: unattended closed-loop operation, measured as bytes per hour per instrument end to end, against the current benchmark of five bytes in 21 hours. Link two is binding, and it is industrial rather than scientific.

Speculative The general-purpose-computer chain. Speculative One: a gate library with restorative logic, so output levels are pulled back to standard values and arbitrary depth becomes possible; measured as achievable circuit depth before signal-to-noise collapse, which today is single digits. Speculative Two: reusable gates, measured as cycles before exhaustion; three: sub-second gate delay, measured in milliseconds. Handwave Four: random-access read/write memory in the same chemistry as the logic, measured as addressable bits, retention and read disturb — nothing resembling this exists. Speculative Five: a compiler and a verified abstraction boundary, so a designer never reasons about hybridisation thermodynamics. Speculative Six: one application where all five beat a two-dollar microcontroller, which realistically means in vivo, at the site of a molecular signature. Link three is binding, and it is a scientific unknown.

Frontier The wetware chain. Frontier One: demonstrate credit assignment inside the culture — synaptic change causally attributable to a task-relevant error signal rather than to generic stimulation-induced plasticity — with a control delivering the same stimulation statistics decorrelated from performance. Frontier Two: an independent, blinded, pre-registered replication of the original result. Established This has not happened and it is the single most important missing result in the field. Frontier Three: scaling, as performance against neuron count over at least two orders of magnitude — if the curve is flat the substrate is not doing the work. Frontier Four: beat the linear readout trained on the same recorded activity. Five: no performance decay across the full device lifetime, which the shipped product caps at six months. Frontier Six: full-system energy accounting, watts per inference including incubation, published. Link two is binding, and it is institutional.

Established A DNA archive is a factory problem wearing a laboratory’s clothes. Established The write side needs an array with more than a million independently addressable synthesis features, reagent handling at industrial volume, and quality control on a product that is a population of molecules rather than an object. The read side needs sequencing amortised across many customers, because a per-site sequencer at six figures destroys the cost case for any archive smaller than a hyperscaler’s. Established The middle needs the boring parts — liquid handling, pooling, primer libraries, indexing, and physical cataloguing of a medium that is invisible and arrives in tubes.

Established The storage half of the engineering is the easy half: silica encapsulation with error-correcting codes gives DNA archival stability at ambient conditions and, unlike tape, needs no migration cycle, no powered environment and no reader that must still exist in a century. Frontier That single property carries the entire long-horizon case, and it is the one thing tape cannot match at any price. Speculative It is also the property least often engineered for, because the demonstrations optimise density and the archive would optimise recoverability after five hundred years of nothing happening.

Frontier A wetware computer is mostly a life-support machine. Established The shipped product’s 850 to 1,000 watts go overwhelmingly to pumps, gas mixing, temperature control and filtration rather than to computation, and the culture dies within six months regardless. Amortised, a US$35,000 unit over a 4,380-hour life is about eight dollars an hour before power, and the power alone comes to some 4,380 kilowatt-hours — which is to say the appliance costs roughly what a small rack costs and delivers no measured throughput at all. Frontier The engineering agenda that would change this is perfusion and vascularisation, a tissue-engineering problem shared with Lab-Grown Organs rather than a computing one. Speculative Until an organoid can be fed through a vasculature rather than by diffusion, both the size ceiling and the power budget are fixed by biology and not by design.

Established The nearest neighbours are not the ones the name suggests. Established DNA storage sits closest to Synthetic Biology and Genetic Engineering, because its binding input is oligonucleotide synthesis capacity and that is the same industrial base gene synthesis runs on — which means a storage-driven synthesis fab would be paid for by one field and consumed by another. Frontier It shares a structural predicament with Energy Storage Revolutions: both are waiting on a first buyer large enough to justify a factory, and in both cases the technology is not what is missing.

Established Molecular computing is adjacent to Nanomedicine in the way that matters most, because the strongest application for an embedded molecular circuit is a diagnostic or smart therapeutic acting at the site of a signature, which is the same design brief written twice. Frontier It is adjacent to Xenobiology through the chemistry of non-canonical nucleic acids — mirror-image DNA in particular is a candidate for nuclease-resistant, orthogonal molecular circuitry that could run in a body without being digested. Established It is adjacent to Designer Organisms wherever the computation is meant to run inside a living chassis rather than a tube — which is where the twenty-two-year-old Benenson automaton has been waiting for a delivery vehicle ever since it was built.

Frontier Neural-culture computing is adjacent to Brain-Computer Interfaces through shared multielectrode hardware and shared decoding mathematics; to Synthetic Consciousness and Consciousness Research through the sentience claim and the moral-status question behind it; and to Lab-Grown Organs through organoid culture technique, which is where its size ceiling will be lifted if it is lifted at all. Established All three programmes are adjacent to Ultra-Efficient Computing Energy Systems, which owns the comparison biological computing keeps making and losing, and which supplies the only honest denominator — joules per useful operation — that any of these substrates could be judged on if anyone measured it.

Established The storage case fails on procurement, not on physics, and the missing institution is a buyer with a long horizon. The comparison everyone runs is DNA against tape on dollars per terabyte per year. Speculative The comparison nobody runs is DNA against paper, microfiche, etched nickel and fused silica for the hundred-to-thousand-year national-archive tier, where tape is not a competitor at all because it must be migrated every seven to ten years and migration is an organisational commitment no institution can guarantee across centuries. Speculative On that tier the required buyer is a national library, a records agency or a nuclear-waste-marker programme willing to pay a one-time write cost of order 10⁴ to 10⁵ dollars per terabyte for a five-hundred-year no-migration guarantee. Established That buyer does not exist, and creating it is a policy act rather than a research programme. Frontier It is the most fundable unfunded item in the subject.

Established An archive format needs a standards body with published specifications, and this brief could not read the one that exists. An industry alliance for DNA data storage exists; its domain was unreachable for this research, so nothing is said here about its membership, outputs or roadmap. Established What can be said is structural: no archive can commit to a format with a single supplier and no independently readable specification, and a fifty-year interchange guarantee is a legal artefact before it is a technical one.

Frontier On the wetware side the missing institution is a provenance and consent regime for cultured human neurons. Frontier Cortical cultures derived from human induced pluripotent stem cells are, legally, tissue; the donor consented to research use, not usually to the sale of a compute product built on their cells. Speculative A growing legal literature on brain organoids exists, and no jurisdiction has yet drawn the line between a research culture, a commercial substrate, and an entity with any protected interest. Frontier Until a consent and chain-of-custody standard exists, a commercial wetware industry is building on a legal foundation that could be revised retroactively.

Speculative And one experiment only an institution can commission: the matched-budget bake-off. Speculative Run one task at matched energy, latency and capital cost on a neural culture with a linear readout, a DNA winner-take-all network and a microcontroller, and publish all three results. Handwave Nobody has done it because it would embarrass at least two of the three communities and the funder would be paying for a negative result. That is precisely the sort of purchase a public research institute exists to make.

Frontier The sentience question is not settled by asserting that it is silly. Established A paper claiming cultured neurons “exhibit sentience” drew a signed rebuttal from thirty neuroscientists, and that rebuttal is about evidential standards rather than about moral status. Speculative The moral-status question is separate and genuinely open: nobody has an agreed criterion that would tell us whether a 100,000-cell cortical organoid with no sensory organs, no body and no vasculature has any interests at all, and the ethics literature on it is small and mostly precautionary. Speculative The honest position is that current systems are almost certainly too small and too disorganised for the question to bite, and that the field’s own stated target of ten million neural cells is a deliberate move toward the regime where it does.

Frontier Consent is the sharper near-term problem. Frontier A donor whose skin cells became an iPSC line consented to research; a compute appliance sold at US$35,000 running that donor’s derived cortical tissue is a use nobody explicitly agreed to. Speculative Reasonable regimes exist on both sides — broad consent with no downstream claim, or a provenance chain with a revocable interest — and the field has adopted neither. A commercial wetware industry that reaches scale before this is settled is inviting the retroactive litigation that has followed every previous case of commercialised human tissue.

Established DNA storage has a quieter ethical structure and one real hazard. Established Data written into DNA is data written into a molecule readable by any sequencer, forever, with no cryptographic guarantee unless one is added at the coding layer and no revocation mechanism at all. Frontier An archival medium with a five-hundred-year retention promise and no delete operation is a privacy commitment of a kind no institution has previously made. Speculative The dual-use question — that a general-purpose cheap DNA writer is also a general-purpose cheap pathogen-sequence writer — is why the synthesis-screening regimes discussed in Synthetic Biology apply here directly, and it is the strongest reason to expect cheap synthesis to arrive regulated or not at all.

Speculative The civilizational case for DNA storage is not capacity. It is the migration cycle. Established Every digital medium a society currently uses requires periodic copying by an institution that must continue to exist, to care, and to be funded. Tape needs migration every seven to ten years, so a thousand-year archive on tape is roughly a hundred and thirty consecutive successful acts of institutional continuity. Speculative Silica-encapsulated DNA with error correction is the only high-density medium proposed that needs none of them, and the failure mode of an unattended archive is decay of a physical object rather than the quiet lapse of a budget line. What DNA storage actually offers a civilization is not a shoebox holding the internet but a way of writing something down that does not require the future to keep paying attention.

Speculative The second implication is about who gets to be remembered. Speculative At current write costs a medium that survives five hundred years without maintenance is affordable only for what an institution deliberately selects. Speculative A world with cheap DNA archiving and a world with expensive DNA archiving differ not in how much is preserved but in who chooses, and the second concentrates that choice in a very small number of national bodies.

Handwave The far implication, if the general-purpose version ever works, is different in kind. Handwave A computational substrate that self-assembles, self-repairs and manufactures itself from ambient feedstock would change the economics of computing the way biology changed the economics of chemistry — not by being faster but by being growable. Speculative Nothing in the present evidence supports expecting it, and the mainstream position is precise about why: biology’s advantages are in fabrication, not in operation. Speculative The interesting counter is that embeddedness is an operational advantage silicon structurally cannot have, and that is the one claim in this subject nobody has properly tested.

These horizons track three clocks running at different speeds: a storage cost curve that has barely moved, a molecular-logic programme rate-limited by diffusion, and a wetware field whose next milestone is a replication rather than a result.

• 10 yr: Frontier Enzymatic synthesis reaches commodity scale for short oligonucleotides, or the archival storage case is effectively closed for the decade; this is the one link that decides the rest. Frontier A national archive runs a funded DNA pilot at the hundred-terabyte scale, priced as a physical good rather than a service. Frontier An independent, blinded replication of the cultured-neuron result is published, or the field’s commercial wing loses its remaining credibility. Speculative An independent wetware benchmark appears with watts per inference stated; the honest prior is that it will not flatter the product. Frontier Molecular logic ships as a diagnostic rather than as a computer, the outcome the evidence has favoured for twenty years.
• 25 yr: Speculative DNA archiving is either an ordinary line item in national records budgets or a permanently interesting laboratory technique, and the deciding variable is synthesis capital expenditure. Speculative A restorative, reusable DNA gate set exists and circuit depth passes double digits, or solution-phase molecular computing is a solved and bounded field. Speculative Vascularised million-cell organoids exist for reasons having nothing to do with computing, and the computing field inherits them. The matched-budget bake-off is finally run, most likely by someone with no stake in the answer.
• 50 yr: Speculative If enzymatic writing became cheap, DNA occupies the cold tier of the archive stack and tape migration becomes a historical practice; if it did not, DNA storage occupies the niche microfiche does — real, used, small. Embedded molecular computation is routine in medicine and invisible as computing, the way a pregnancy test is not filed under information processing. Handwave Cultured neural tissue is either a standard laboratory instrument for studying learning or a discontinued product category; the strong version, beating silicon at matched energy, has no supporting evidence today.
• 100 / 250+ yr: Handwave The single-molecule read-write pore, if possible at all, is what makes all of this ordinary: storage, logic and sensing in one device at the scale of a single strand. A growable computational substrate — self-assembling, self-repairing, made from feedstock rather than fabricated — is coherent as a goal and has no demonstrated component. Handwave The most defensible long claim in this subject is the least exciting one: a silica-encased DNA archive written now, with no maintenance, will still be readable when every other medium in this sentence has been migrated a hundred times or lost.

• Depends on This brief waits on Synthetic Biology and Genetic Engineering for the one input that decides the storage programme: oligonucleotide synthesis at a price and a scale nobody currently needs for any other purpose. It waits on Lab-Grown Organs for vascularised, perfused neural tissue, without which the organoid size ceiling and the kilowatt life-support budget are both fixed by oxygen diffusion. It waits on Xenobiology for orthogonal nucleic-acid chemistries that would let a molecular circuit run in a biological environment without being degraded by it. And it waits on Brain-Computer Interfaces for the high-density recording hardware and the decoding mathematics every wetware demonstration borrows wholesale. What it does not wait on is any result in information theory: on that axis the storage field is already within a few per cent of the ceiling of its medium.
• Requires (not on this map) Two constraints sit outside the map of research briefs because neither is a scientific result. The first is industrial: template-independent enzymatic DNA synthesis manufactured at archival cost, meaning a fab with more than a million addressable synthesis features running below a nanodollar per base. The chemistry is demonstrated; the plant does not exist, because the only market that would justify it is the market the plant would create. The second is institutional: a provenance and consent regime for cultured human neurons that distinguishes a research culture from a sold compute product, and says what the donor of the originating cell line is owed. Neither constraint waits on a discovery, and both are load-bearing — the first decides whether DNA archiving is ever a product, the second decides whether commercial wetware computing survives its first serious legal challenge.
• Enables Cheap, reliable DNA synthesis is the input that Designer Organisms and Synthetic Biology are both rate-limited by, so a storage-driven synthesis fab enables them whether or not a single archive is ever sold — that spillover is the strongest industrial-policy argument for funding this subject at all. Embedded molecular logic supplies Nanomedicine with autonomous in-situ decision-making that needs no instrument and no power source. A maintenance-free archival medium supplies Civilization Resilience Planning with the only proposed high-density store that does not depend on institutional continuity.
• Adjacent to Ultra-Efficient Computing Energy Systems, which owns the energy comparison this field keeps making badly; to Synthetic Consciousness and Consciousness Research, which own the moral-status question cultured cortical tissue raises and this brief cannot settle; to Memory Engineering, which shares the problem of writing information into a biological substrate and reading it back intact; and to Artificial General Intelligence, which is the comparison class the wetware programme has chosen for itself and against which it has so far published no benchmark.

Established “One gram of DNA holds about 700 terabytes.” Established This is a volume-to-mass unit swap, wrong by roughly a thousandfold. The reported density is 5.5 petabits per cubic millimetre, which is 687.5 terabytes — the folk figure is that same numeral with the wrong unit attached. Established Realised per-gram densities are in petabytes: 215 petabytes per gram measured, 400 to 625 petabytes per gram if you convert the cubic-millimetre figure properly, and the DNA Fountain authors’ own phrase is “petabytes of data per gram.” It is not a conservative version of the claim; it is the wrong unit. Established The original paper could not be opened for this brief and the correction does not require it — the numeral identity, the independent per-gram measurement and the theoretical-ceiling cross-check all come from documents that were read.

Established “DNA storage will be cost-competitive soon.” Established The best rebuttal is the founding paper’s own abstract. Goldman and colleagues wrote in Nature in February 2013 that “current trends in technological advances are reducing DNA synthesis costs at a pace that should make our scheme cost-effective for sub-50-year archiving within a decade.” Established That decade expired in 2023, and by no public accounting is DNA storage cost-competitive for fifty-year archiving; the present gap against tape is eight to nine orders of magnitude. This is the most instructive over-promise in the subject precisely because it is not a journalist’s. It is in the primary literature, in the abstract, with a clock attached.

Established “DNA storage is fast because it is massively parallel.” Established Reading is parallel; writing is serial per base, and the one fully automated end-to-end device on record moved five bytes in about 21 hours. The related enthusiast error is that falling sequencing costs imply falling storage costs. They do not: reading and writing are different problems with different cost curves, and the write curve is the binding one.

Established “Adleman proved DNA computers can crack hard problems.” Established He solved a seven-node Hamiltonian path instance — trivial by hand — in about a week, as a demonstration of principle, and two years later Boneh and colleagues showed the resource requirement grows exponentially with problem size, so a few dozen more variables would need more DNA than the planet holds. Frontier The skeptic-side mirror of this error is “DNA computing is dead.” Established The theoretical-computer-science framing is dormant; molecular programming as embedded computation is active and produced Nature-level results as recently as 2023. Do not confuse the death of one framing with the death of a field.

Established “Organoid intelligence is a million times more energy-efficient than silicon.” Established The ratio comes from contrasting a human brain at a modelled one exaFLOP for 10 to 20 watts against the Frontier supercomputer at 1.102 exaFLOPS for 21 megawatts. Established Two things are wrong with it. First, the brain’s exaFLOP is a modelling assumption rather than a measurement, so this is a hypothetical divided by an actual. Second, and decisively, the argument collapses against its own product: the only shipped wetware computer draws 850 to 1,000 watts, which is forty to a hundred times the power of an entire human brain, to run a culture that dies within six months. Established Per unit of useful computation the delivered biological machine is enormously less efficient than a laptop, and no independent benchmark exists that would even permit the ratio to be computed.

Established “Lab-grown neurons were shown to be sentient,” and “DishBrain learned to play Pong.” Established The word sentience appears in the paper’s title in a specific technical sense and drew a published rebuttal in the same journal signed by thirty neuroscientists. Established The measured effect was a statistically significant increase in average rally length between the first five minutes and the last fifteen minutes of a twenty-minute session, against media-only, rest, random-noise and three feedback control conditions. That is a real, controlled effect, and it is not playing Pong. Frontier The steelman for the authors is that a differential against four control arms is the substantive result and survives the naming dispute entirely; the steelman for the critics is that a differential response to structured stimulation is what any plastic excitable tissue does, and calling it learning imports the conclusion.

Established “Brain organoids are miniature brains.” Established The reference organoid in the field’s own manifesto is under 500 micrometres across with fewer than 100,000 cells, no vasculature, no sensory input and no body, at about 40% myelination against roughly 50% in human brain. The manifesto’s own target of ten million neural cells would still be four orders of magnitude short of a mouse brain.

Frontier Two smaller corrections, offered because this subject rewards checking. Frontier The figure sometimes quoted for a benchmark strand-displacement square-root circuit — “over 100 hours” — appears in a survey rather than in the primary paper, and published seesaw-circuit runtimes are more commonly quoted around ten hours; this brief could not resolve which is right and states only that the runtime is in hours. Speculative And the founding idea of the field was not Adleman’s: M. S. Neiman published three papers in the Soviet journal Radiotekhnika in 1964 and 1965 on molecular-level information storage, three decades early, in Russian, in a radio-engineering journal, and the field did not notice for thirty years. Established That is a checkable instance of the Institute’s own thesis — an unfashionable venue is not the same thing as an unserious idea, and the cost of that mistake was measured in decades.