Molecular Computing — Part II | Tetheron
Tetheron · Molecular Computing · Part II of a series

What biology actually demonstrates

Part I argued that the leverage in engineering has moved to the architectural layer. This part goes one scale down, into the only system anyone has ever pointed to as proof that matter can think — and is careful about what that proves.
Scroll ↓
A scattered field of identical rod-shaped molecular units at the base, from which three visibly different structures rise: a dense hexagonal lattice, an open branching tree, and a layered stacked column. Faint lines trace from the shared field up into each structure.
One parts bin, three architectures — assembled independently.

Every instance of intelligence anyone has ever pointed to runs on molecular machinery. Not on a logic layer above chemistry, not on some organizing principle still to be found — on molecules, and on how they are arranged.

That is the claim this piece defends. There is no hardware–software split in a cell and no separate substrate where the thinking happens. Channel gating, receptor binding, scaffold reorganization, vesicle fusion: the primitives are molecular events, and cognition is what those events do once they are organized. A ciliate with no nervous system learns to ignore a repeated stimulus, and the investigators who measured it suspect complex molecular computation upstream of the change. Biology is our inspiration here, not our evidence — the evidence comes from our own campaign, which began in August.

What biology hands us is a design brief, and a remarkably specific one.

The molecules came before the machine. The architecture came after — and it came more than once.

Four things the record shows

Arrangement beats content, inside a single synapse. Using localization microscopy, Tang and colleagues in Thomas Blanpied’s group showed that the proteins driving vesicle release cluster into nanometer-scale domains that align across the synaptic cleft with concentrated receptors and scaffolds — a trans-synaptic “nanocolumn.” The consequence is the part that matters here: perturbing that alignment disrupts transmission even when the synapse’s receptor content is unchanged. Same molecules, same amount of them, different arrangement, different function. This is Part I’s finding stated at the scale of a single junction. Tang, Chen, Li, Metzbower, MacGillavry & Blanpied · Nature 536, 210–214 · 2016
Two panels showing the same two membrane surfaces with the same number of protein clusters. Left: clusters aligned in vertical columns and light passes cleanly through. Right: the same clusters offset and staggered, and the light is scattered and diffuse.
Schematic. Same components, same quantity — aligned on the left, offset on the right.
The parts predate the nervous system. Voltage-gated sodium and calcium channels are present in choanoflagellates — single-celled organisms with no neurons and no nervous system. The scaffold backbone that organizes the postsynaptic density (Shank, Homer, DLG) traces to a unicellular ancestor of choanoflagellates and animals. In sponges, which have no nerve cells at all, the residues forming the critical contacts in the dlg scaffold are identical to ours. The molecular toolkit was not invented for thinking. It was already there. Liebeskind, Hillis & Zakon · PNAS · 2011 • Sakarya et al. · PLoS ONE · 2007 • Alié & Manuel · BMC Evolutionary Biology · 2010
And the architecture was assembled from them more than once. Ancestral gene-content reconstruction across the major ion-channel families finds that animals with nervous systems carry broadly similar channel complements — but that those complements evolved independently. Ctenophores, cnidarians and bilaterians each underwent separate expansions in the families that shape synaptic transmission and action potentials. Evolution reached into the same parts bin repeatedly and built different machines out of it. Liebeskind, Hillis & Zakon · PNAS · 2015
Between species, architecture is the variable that moves. Songbird and parrot brains contain roughly twice the neurons of primate brains of the same mass, at packing densities exceeding anything in mammals, and concentrated in the pallium. Large corvids and parrots reach forebrain neuron counts equal to or above primates with far larger brains. A honeybee runs a genuinely rich behavioral repertoire on about a million neurons packed into roughly a cubic millimetre — some two orders of magnitude denser than mammalian cortex. And density is not a function of size: clades scale by different rules, so a large brain can be sparsely populated and a small one densely. The molecular machinery of excitability is broadly shared across all of them. What differs — dramatically — is packing, allocation and wiring. Hold the molecular parts constant and vary how they are arranged, and the four look nothing alike: Olkowicz et al. · PNAS · 2016 • Menzel · Nature Reviews Neuroscience · 2012 • Herculano-Houzel, Collins, Wong & Kaas · PNAS · 2007 • Kazu et al. · Frontiers in Neuroanatomy · 2014
Four neural architectures built from one shared molecular substrate Four panels. Primate neocortex: six layers with a dense granular input layer IV and pyramidal neurons. Artiodactyl cortex: the same six layers, heavily folded, with fewer neurons in them. Avian pallium: rounded nuclear clusters rather than layers, with an iterated column-like circuit running through them. Insect mushroom body: no cortex at all, cell bodies lining the rim of a calyx, and a microglomerulus with one central bouton encircled by claw-like dendritic endings. A band across the bottom shows the shared molecular substrate of voltage-gated channels and postsynaptic scaffold proteins common to all four. PRIMATE NEOCORTEX Six layers. Dense granular input layer IV. I II III IV V VI granular layer IV ARTIODACTYL CORTEX (SHEEP) The same six layers, heavily folded, with fewer neurons in them. I II III IV V VI more surface, fewer cells AVIAN PALLIUM Nuclear, not laminar — and the same iterated column runs through it. iterated column circuit INSECT MUSHROOM BODY No cortex at all. Somata line the rim; synapses are microglomeruli. pedunculus one bouton, many claws SHARED MOLECULAR SUBSTRATE Voltage-gated channels and postsynaptic scaffold proteins: present in all four, and older than every one of these architectures. channel · scaffold · channel · scaffold
Four architectures, one parts bin. The panels differ in layering, folding, grouping and synaptic form; the band beneath them is common to all four and predates every one of them. Schematic.

What biology spends its energy on

Biological computation is not free, and we should stop anyone who says it is. The brain is about two per cent of body mass and consumes roughly twenty per cent of the body’s energy budget. What is interesting is not that the number is low. It is where the number goes.

Attwell and Laughlin’s budget for signaling in grey matter apportions it: about 47% to action potentials, 34% to postsynaptic currents, 13% to maintaining the resting potential, and 3% to recycling the transmitter itself. Roughly ninety-four per cent of the signaling cost is spent moving ions across membranes to hold and change electrical state. The molecular events that carry the computation — binding, gating, conformational change — are the cheap part. The expensive part is the electrical layer wrapped around them.

Where the brain’s signaling energy goes Stacked bar of the Attwell and Laughlin energy budget for signaling in grey matter: action potentials 47 per cent, postsynaptic currents 34 per cent, resting potential 13 per cent, glutamate recycling 3 per cent, presynaptic vesicle release about 3 per cent. The first three, totaling 94 per cent, are the cost of moving ions across membranes. WHERE THE BRAIN’S SIGNALING ENERGY GOES 94% is spent moving ions across membranes — the electrical layer 47% 34% 13% Action potentials Postsynaptic currents Resting potential 6% — transmitter release and recycling SOURCE: ATTWELL & LAUGHLIN, J. CEREB. BLOOD FLOW METAB. 21, 1133–1145 (2001) The four named categories total 97%; the balance is presynaptic vesicle release, giving 6% for presynaptic work overall.
The cost is in the wiring, not the switching.

That is a finding, and it is the single most useful thing biology tells an engineer. Nature’s twenty watts is not the price of molecular computing. It is mostly the price of the charge-shuttling infrastructure that biology needed in order to move molecular events around a body.

Which is what we are building against. If the expensive layer is the electrical one, then a substrate whose operand is the incoming field itself — transformed in the material, before digitization — does not have to pay that bill. That is ODMA™ in one sentence.

The whole argument turns on one measurable number: how the write energy scales with the area you address. Scale it one-for-one and the advantage is gone; scale it sub-linearly and the advantage is real, and grows the further below one-for-one it lands. That is an exponent, not a verdict, and we put the measurement in the first campaign rather than the fifth. We will have the number this year.

And the density that architecture buys

The other half of the brief is packing. A cubic millimeter of human cortex holds on the order of a billion synapses — close to one per cubic micrometer, threaded through several kilometers of wiring in that same cubic millimeter. That is the densest computing substrate anyone has measured, and it is built by arranging molecules, not by lithography.

A synapse is nonetheless a large object. It is assembled from thousands of proteins and occupies roughly a cubic micrometer. Molecular spacing is three orders of magnitude finer in each dimension. The geometric headroom between “one computing element per cubic micrometer” and “one per cubic nanometer” is not subtle, and it is the reason to work at this scale rather than a smaller version of the last one.

That headroom is geometric rather than speculative — it follows from spacing, and nobody has to take our word for it. How much of it turns out to be addressable is the question our first campaign opens. We would rather earn that number than estimate it.

What we are claiming

Every known instance of intelligence runs on molecular machinery, and what varies between instances is how that machinery is arranged. Arrangement is the operative variable, one scale below where Part I found it. Biology is the proof that this class of thing works at room temperature, at low power, at densities no fab has approached. We intend to build the second member of that class.

One distinction we would rather draw than have drawn for us

Biology’s molecular computation is networks of molecules signaling through conformational change and diffusion, on millisecond timescales. Ours is the internal dynamics of a selected vibrational mode in the far infrared, roughly eight orders of magnitude faster. There is no organized dipolar molecular array in a neuron, and we are not reverse-engineering one — we are taking the principle and building something evolution never had the chemistry to try.

So anyone who tells you biology has already validated a terahertz molecular device is not reading carefully. And anyone who tells you the principle stops at the water’s edge owes an account of why arrangement should stop mattering at the one scale nobody has built at yet.

Where Tetheron sits

We are not building a faster transistor, and we are not entering the qubit race with another modality. We are building an architectural substrate for computing, organized at the molecular scale, in solid-state configurations that today’s semiconductor infrastructure can integrate with.

What we are aiming at follows directly from the two findings above. Low power, because the expensive layer in biology is the one we are trying not to build. Speed set by the material, because the operating mode’s intrinsic dynamics are sub-picosecond. Density, because the geometric headroom below the synapse is where the room is. And therefore edge computing — inference where the signal arrives, rather than shipped to a building and back.

Where that leads is a different shape for the industry: computation that does not have to be centralized, because it no longer needs the power and cooling that force centralization. We think that is the more plausible route to general machine intelligence than continuing to scale the current one, and we are building toward it on purpose.

What that means concretely, this year. Our first fabrication and characterization campaign began with our university partners in August 2026. It makes films, not devices, and it is built to answer one question before it answers any others: whether the operating point this architecture requires is physically reachable at all. That question has a pre-registered answer of “no” available to it, and if we get it, we will say so.

Biology proves that matter can compute at room temperature, at molecular scale, on almost no power. It took evolution one parts bin and four billion years. We think there is a second way to do it.

Part II of a series. Part I sets out the six-field case that architecture, not composition, carries the functional behavior.

Start a conversation →
Tetheron is developing organized dipolar molecular array (ODMA) technologies for next-generation computing.
ODMA™ is a trademark of Tetheron Inc.
References, in order of appearance — A. H. Tang, H. Chen, T. P. Li, S. R. Metzbower, H. D. MacGillavry & T. A. Blanpied, “A trans-synaptic nanocolumn aligns neurotransmitter release to receptors,” Nature 536, 210–214 (2016), doi:10.1038/nature19058 · B. J. Liebeskind, D. M. Hillis & H. H. Zakon, “Evolution of sodium channels predates the origin of nervous systems in animals,” PNAS (2011) · O. Sakarya et al., “A post-synaptic scaffold at the origin of the animal kingdom,” PLoS ONE (2007) · A. Alié & M. Manuel, “The backbone of the postsynaptic density originated in a unicellular ancestor of choanoflagellates and metazoans,” BMC Evolutionary Biology (2010) · B. J. Liebeskind, D. M. Hillis & H. H. Zakon, “Convergence of ion channel genome content in early animal evolution,” PNAS (2015), doi:10.1073/pnas.1501195112 · S. Olkowicz et al., “Birds have primate-like numbers of neurons in the forebrain,” PNAS (2016), doi:10.1073/pnas.1517131113 · R. Menzel, “The honeybee as a model for understanding the basis of cognition,” Nature Reviews Neuroscience 13, 758–768 (2012), doi:10.1038/nrn3357 · S. Herculano-Houzel, C. E. Collins, P. Wong & J. H. Kaas, “Cellular scaling rules for primate brains,” PNAS (2007), doi:10.1073/pnas.0611396104 · R. S. Kazu et al., “Cellular scaling rules for the brain of Artiodactyla include a highly folded cortex with few neurons,” Frontiers in Neuroanatomy (2014), doi:10.3389/fnana.2014.00128 · D. Attwell & S. B. Laughlin, “An energy budget for signaling in the grey matter of the brain,” Journal of Cerebral Blood Flow & Metabolism 21, 1133–1145 (2001) · P. R. Huttenlocher, “Synaptic density in human frontal cortex — developmental changes and effects of aging,” Brain Research (1979) · B. Staffler et al., “SynEM, automated synapse detection for connectomics,” eLife (2017), doi:10.7554/eLife.26414 · Habituation in Stentor coeruleus: W. F. Marshall and colleagues, “Single-cell analysis of habituation in Stentor coeruleus,” Current Biology (2022), PMID 36435177; associative learning in the same organism is reported in a 2026 bioRxiv preprint and is not peer-reviewed.