What a material has to do to compute
Biology is the one existence proof that matter can compute at molecular scale, and there is probably a second way to do it. This sets out what the second way actually requires — and the one requirement no molecular film has yet met.
The requirement set is short. It is also unforgiving, and one item on it has never been demonstrated in a molecular film by anyone.
Start with a thin film: carrier molecules dispersed at random through a host polymer. One phase, no lattice, no repeating unit. Infrared light at the right frequency is absorbed, because that is what chemical bonds do. So far this is ordinary spectroscopy — a single absorption band on a flat baseline, and nothing about it computes anything.
Turning that into a computing medium takes three things. Each is a property of the material rather than of a circuit built on top of it, and each can be measured and can fail. We are going to be specific about which ones have been shown and which one has not.
One · it has to bend
A material that responds in exact proportion to its input is useless for computing, and the reason is worth stating precisely rather than waving at. Feed a linear medium two inputs and the output is the sum of the two outputs. Stack a hundred linear media and the result is still one linear map. Nothing you can build out of linearity produces a function that linearity did not already contain.
What breaks that is saturation. Drive the film harder and molecules already excited cannot absorb again, so the response to twice the amplitude falls short of twice the response. That shortfall is not a defect. It is the whole basis on which combinations of inputs can produce something no single input produces.
This is the first thing our campaign measures, and it is the gate everything else waits behind: a reproducible input–output relation at the operating mode that departs from linear, in a real film, at a workable loading, surviving its controls. It has a pre-registered answer of “no” available to it.
Two · it has to remember — and that word is doing four jobs
Take two identical films. Give each of them the same two drives, carrying the same total energy, and change only the order in which they arrive. Then send an identical test pulse into both, last. The answer that comes back is different.
That is memory in its loosest sense: the film is responding to the history of what it has seen and not only to the most recent thing. And it is worth noticing what selected the state. Not which wire was energised — the light itself did.
But after a drive ends, the fast response decays and the film returns to baseline. A material that computes on sequences rather than on single inputs needs something that does not — a second variable that outlasts the input.
Here is where the field talks past itself, and we have done it too. “Memory” in a computing material means at least four different things, and conflating them makes claims look larger than they are.
We want to be fair about this, because the temptation in a piece like this one is to make the published work sound like a stepping stone. It is not. Building a non-volatile element that reliably holds trained weights, in hardware, at room temperature, is a genuinely difficult achievement and several groups have done it. The third and fourth senses are further out — and in a molecular film neither has been shown at all. That is where we are working.
The distinction the whole thing turns on
A persistent state that does not change the transform is storage. A persistent state that changes the transform is context.
The value of storage is the information it holds. The value of context is the change it makes to the computation. In the first case you have a computer with a memory attached to it. In the second, the memory is part of the computation — there is no point in the cycle where storage ends and processing begins. That is the difference between a component and a substrate, and it is the difference the company is built on.
Something close to this exists and it is not ours. A programmable memtransistor has been shown in which a non-volatile charge-trap state modulates the dynamics of a volatile reservoir; and inorganic phase-change films — GST, GSST, Sb2Se3 — hold an optically written non-volatile state at many levels and use it to reprogram an optical element. Both are real, both are in hardware, and one of those materials is transparent right across the band we work in.
What the first reconfigures is a wired electronic device, and what the second reconfigures is a linear optical element. Nobody has shown a persistent state that changes the nonlinear input–output map of a molecular film computing on light in its own band. Not us. It is an open question, and it is the one our current work exists to answer.
What would settle it is a single experiment: write one state and measure the input–output curve; reset; write a different state and measure the same curve — after the fast response has died away, so that what is being tested is the retained state and not a fading trace. If the curves differ in shape, the stored state changed the transform. If they do not, the question closes.
Three · you have to be able to read it out
A material that computes beautifully and cannot be interrogated is a curiosity. The readout is committed and it is deliberately modest: one weak probe pulse in, resolved across frequency and across time delay. What comes back is not a number but a surface — a vector of values.
That vector meets a small grid of weights and produces a label. You do not program the film. You train the readout. The division of labour is the point: the expensive, hard-to-build part — nonlinear mixing of a high-dimensional input — is done by the physics, and what is left for the electronics is a linear operation they are already extremely good at.
Usable computation is bounded by how many independent things you can observe, not by how many molecules are in the film. A single probe reduced to a single number would cap the measured capacity almost regardless of how rich the material’s behaviour is. Recovering rank by resolving that one probe more finely, rather than by adding probes, is a design commitment we have made and not yet tested.
Where the energy comes from
There is an objection that a physicist raises in the first five minutes, and it is a good one. A room-temperature scene is a very faint infrared source. Run the numbers on what arrives across a single absorption line, through fast optics, and compare it to what it takes to push a condensed-phase vibrational transition anywhere near saturation, and the gap is not close. A passive thermal scene cannot appreciably drive that transition. It is not a near miss.
So the architecture does not ask it to. Two drives reach the material: a strong, featureless bias that carries the energy and no information, and the weak signal that carries the information and steers what the bias does. The bias makes a state change possible; the signal decides which state, or how fast. The supply rail carries the power, and the gate decides what happens to it.
One boundary on that, because the calculation is narrower than the sentence. What the arithmetic closes is the fast route — a scene driving the vibrational transition directly, on picoseconds. It does not close a slower one, in which scene energy accumulates into a thermal or structural state over milliseconds. That route is not excluded, it is not the same architecture, and we measure it separately rather than assuming it either way.
Moving the energy source onto a bias is not free. The bias enters the system’s power budget, and any claim that the architecture saves energy end to end now has a term on the other side of the ledger that has to be measured rather than argued.
Bandwidth survives this unchanged — a classification that never becomes a frame is still a frame nobody transmits. So does privacy: a scene that is never digitised is not a scene image that can leak. The power claim has to be re-earned, and we would rather say so here than be shown it in diligence.
Why the substrate is the argument
The reason to do any of this at the molecular scale, rather than building the same functions out of silicon, is what the elementary events cost and how many of them there are.
One vibrational quantum in the operating band costs about 20 zeptojoules — roughly seven times the thermodynamic floor for erasing a bit at room temperature. A CMOS switching event is a band rather than a point, depending on process node and on whether you count interconnect, and it sits well above. And at around ten weight per cent loading there are on the order of 1020 carrier molecules in a cubic centimetre of film.
These are native physical scales and design parameters. They are not a device energy per operation and they are not a demonstrated rate. A write drives many molecules at once; readout, reset and the surrounding electronics sit on top of the substrate figure and are unmeasured. The opportunity is the gap between these scales and today’s system-level cost. How much of that gap a real device captures is engineering, and measuring it is what the programme is for.
Where this stands
Our first fabrication and characterization campaign has been running with our university partners since August 2026, funded by the founder rather than by investors. It makes films, not devices.
It answers the first requirement before it answers any others: is there a resolvable, repeatable nonlinear response in a real film at a workable loading? Then whether the material holds a state at all, and whether that state changes the transform.
Every one of those has a pre-registered result that would falsify it, agreed before the measurement runs. A negative is a real outcome and we have written down in advance what each one would mean.
The same argument, built as something you can move through — see the interactive illustration.
InteractiveThe three requirements, one at a time: a response that bends, a state that persists, and a readout that can see it. Conceptual throughout — no measured result is shown.
Open the illustration →Also from this work
Six fields in which the functional behaviour of a material turns out to be carried by how its components are arranged rather than by what those components are. The case that arrangement is the operative variable — and that the material is closer to a substrate the arrangement runs on.
Read more →Biology is the one existence proof that matter computes at molecular scale. What that demonstrates, and what it does not: the molecular components are conserved and older than the architectures built from them, and the architectures arose independently more than once. Arrangement is the operative variable again, one scale below where the first piece found it.
Read more →That is not a reason to avoid saying it. It is the reason the next experiment is worth running.
Start a conversation →