Physical AI starts from an idea that sounds like a category error: there is no neural network running on the hardware, because the hardware is the neural network. No weights stored in memory, no model file, no separation between the thing that computes and the thing being computed. A tangle of nanoscale wires reorganises itself in response to the voltages you apply, and that reorganisation is the learning.
That argument was set out in detail on 18 September 2026, when Nature Reviews Physics published a Perspective titled “Self-organizing memristive networks as physical learning systems” — a 165-reference survey from six researchers across the United States, Italy, New Zealand and Australia. UCLA publicised it under a blunter headline: when the hardware becomes the neural network.
The motivation is not elegance. The paper’s own framing is that the impetus for physical AI comes “largely from the unsustainability of artificial neural network software implemented on conventional transistor-based hardware.” This article works through what the researchers are actually claiming, what has been measured in the lab, where the honest limits sit, and why anyone running AI workloads on ordinary silicon should care about a materials-science result.
Table of contents
- What Physical AI Actually Means
- Why Physical AI Is Being Taken Seriously Now
- The Two Device Families Behind Physical AI
- What Physical AI Has Actually Been Measured Doing
- The Theory That Makes Physical AI Tractable
- Where Physical AI Is Not Ready Yet
- What Would Have to Happen for Physical AI to Ship
- Who Is Doing This Physical AI Work
- What Physical AI Means for Everyone Else
- References
What Physical AI Actually Means
The term is doing more work than it looks, and it is worth separating from the several other things “physical AI” is used to describe in industry marketing.
The hardware is not a platform, it is the model
Adam Stieg, a research scientist at UCLA and associate director of the California NanoSystems Institute, put the distinction plainly: “Most AI treats the hardware as a passive platform for running software.” The physical AI proposition inverts that. In these systems, as Stieg describes it, “the model evolves in the physical network itself. It adapts and changes.”
Physical AI hardware self-organises rather than being designed
The networks at the centre of this work are not laid out by a designer. Silver nanowires or metal nanoparticles are deposited and allowed to settle into a disordered mesh. Under electrical stimulation they spontaneously acquire new structural and functional properties, which is the formal sense in which they are self-organising: the wiring diagram emerges rather than being specified.
In physical AI the memristive junctions are the synapses
Where two nanowires cross, the junction changes its conductance in response to the history of voltage across it — resistive memory switching. That history dependence is the closest physical analogue to a synaptic weight, except that nothing writes it. The paper’s shorthand for the whole class is SOMNs: self-organising memristive networks.
Why this is not just neuromorphic computing
Neuromorphic chips imitate the brain’s architecture in conventional silicon, with designed circuits standing in for neurons and synapses. Physical AI goes a step further and gives up the design. The collective behaviour arises from interactions between nanoscale elements, in the spirit of the “More is Different” argument the authors explicitly invoke.
Why Physical AI Is Being Taken Seriously Now
Interest in exotic computing substrates tends to track the cost of the conventional one, and that cost is now the story.
The energy argument is the whole physical AI argument
Data centres consumed around 415 TWh of electricity in 2024, roughly 1.5% of world consumption, according to the International Energy Agency. Its projection has that figure reaching about 945 TWh by 2030, with electricity demand from AI-optimised data centres more than quadrupling over the same period.
Moving data is what costs the money, and physical AI removes the move
In a conventional accelerator, most of the energy goes on shuttling weights and activations between memory and arithmetic units rather than on the arithmetic itself. Physical AI attacks that directly: if the weights are the material, there is nothing to fetch. Computation happens where the information already is.
Learning at the edge, not in a data centre
The applications the authors name are all places where a round trip to a data centre is impractical: satellites, autonomous vehicles, industrial robots, smart devices, distributed sensing and personalised healthcare. The pitch is not to replace cloud training but to complement it, handling sensor data locally in hardware that adapts on site.
Continual adaptation instead of retraining
Because the substrate changes as it is stimulated, these systems are pitched at continual learning rather than train-once-deploy-everywhere. The paper draws explicit parallels between adaptive dynamics in these networks and plasticity in biological neuronal networks, and argues that is the route to brain-like continual learning at low power.
The Two Device Families Behind Physical AI
Two experimental lineages dominate the literature, and they fail and succeed in different ways.
Physical AI device family one: nanowire networks, from 2011
UCLA introduced the nanowire approach in 2011. Silver nanowires, produced by electrodeposition, form a dense disordered mesh whose cross-point junctions switch. Under voltage sweeps the network shows a collective hysteresis response characteristic of memristive systems — not a property of any single junction but of the mesh as a whole.
Physical AI device family two: nanoparticle networks, from 2013
The nanoparticle route followed in 2013. Here the switching lives in tunnelling gaps between particles, and because junction conductance depends exponentially on the instantaneous tunnelling barrier width, even a modest history-dependent change in gap size produces strong nonlinear potentiation.
Scale: three orders of magnitude in one figure
The Perspective’s own micrographs make the scale range concrete. The scale bars on its scanning electron microscope images of nanowire and nanoparticle networks run 25 µm, 5 µm and 200 nm — a span of 125 to 1 between the widest and narrowest view.
Joule heating as a mechanism, not a fault
One detail captures how differently this field thinks about hardware. When large biases are applied, Joule heating can break down individual nanowires, creating gaps where metal ions migrate and form filaments. In a processor that is a failure. In a physical AI substrate it is a new memristive junction — the device rewiring itself.
| Property | Nanowire networks | Nanoparticle networks |
|---|---|---|
| Introduced | 2011 (UCLA) | 2013 |
| Typical material | Silver nanowires, electrodeposited | Metal and metal-oxide particles |
| Switching site | Wire-to-wire cross-point junctions | Tunnelling gaps between particles |
| Nonlinearity source | Filament formation and relaxation | Exponential dependence on gap width |
| Imaged scale in the paper | 25 µm and 5 µm fields | 200 nm field |
What Physical AI Has Actually Been Measured Doing
This is where the field earns or loses its credibility, and the honest answer is that the benchmarks are real but modest.
93.4% on MNIST, learned online, on the device
The strongest published result of this type came from a nanowire network device performing online dynamical learning. Applied to MNIST handwritten digit classification, it reached an overall accuracy of 93.4%, and the same device demonstrated recall of an embedded spatiotemporal sequence pattern. That work also found a correlation between per-class accuracy and mutual information.
Reservoir computing is the usual physical AI harness
The standard way to extract useful work from one of these meshes is reservoir computing: the network’s nonlinear dynamics map a spatiotemporal input into a high-dimensional feature space, and only a small readout layer is trained. A fully memristive implementation of this, with a resistive switching memory readout, has been demonstrated on MNIST and on Mackey–Glass time-series prediction.
The training-cost claim is structural, not marginal
The efficiency claim follows from that architecture rather than from any clever optimisation. If the reservoir is fixed physics and only the readout is fitted, the number of trained parameters collapses. That is the sense in which these systems are described as having a reduced training cost — the substrate does the expensive transformation for free.
In physical AI, criticality is a measurable property rather than a metaphor
The theoretical half of the Perspective is about transitions between conductance states, where the authors report that “criticality and other dynamical phase transitions emerge” in both experiments and models. Networks poised near a critical point are the ones that compute well, which gives the field an actual tuning target rather than a vague appeal to complexity.
How this differs from the learning you already run
It is worth stating what physical AI is not doing. There is no gradient descent through the substrate, no backpropagation, and nothing resembling reinforcement learning running inside the material. The adaptation is a physical process with learning-like statistics, harnessed by a trained readout — which is why the comparison to conventional training curves is not straightforward.
The Theory That Makes Physical AI Tractable
A disordered mesh of millions of junctions is not obviously analysable, and for years that was the objection.
Physical AI borrows its tools from statistical physics
The Perspective’s theoretical contribution is to assemble approaches that do work: mean-field theory, graph theory and concepts from disordered systems. These give a way to reason about collective behaviour without simulating every junction, which is what makes the substrate engineerable rather than merely interesting.
Why disorder helps rather than hurts
In conventional manufacturing, device-to-device variation is a defect to be minimised. In a reservoir it is the source of the rich dynamics that make the feature mapping useful. The field’s central bet is that variability can be exploited rather than suppressed — a genuinely different relationship with fabrication tolerance.
The plasticity parallel
The paper argues that adaptive dynamics in these networks parallel plasticity in biological neuronal networks. That is a claim about dynamics, not about consciousness or cognition, and the authors keep it that way: the interest is in whether physical adaptation supports continual learning without the catastrophic forgetting that plagues sequential training in software.
Where Physical AI Is Not Ready Yet
The Perspective is candid about this, and the limitations are the parts to read twice.
Physical AI variability and reproducibility
Nanofabrication of self-organised networks does not produce identical devices. Two meshes made the same way behave differently, which matters enormously if the plan is to ship a product rather than publish a paper.
Long-term stability of a physical AI substrate under cycling
Memristive states drift. Repeated cycling changes device behaviour over time, and a substrate that is the model cannot be reloaded from a checkpoint when it does.
Scaling while keeping control
Larger networks bring richer dynamics and less controllability. Getting both at once is an unsolved problem, and it is the one that most directly limits the size of task these systems can take on.
Interfacing with conventional electronics
Whatever the substrate does has to be read out, conditioned and delivered to ordinary digital systems. That interface is both an engineering cost and a place where the energy advantage can quietly be given back.
| Dimension | Conventional neural network stack | Physical AI substrate |
|---|---|---|
| Where the model lives | Weights in memory, separate from compute | In the material’s own structure |
| What gets trained | Every layer, by gradient descent | A small readout layer only |
| Dominant energy cost | Data movement between memory and compute | Stimulation and readout of the network |
| Device variation | A defect to be engineered out | A resource that enriches the dynamics |
| Reproducibility | Bit-exact across identical hardware | Device-specific; an open problem |
| Recovery after failure | Reload the checkpoint | No checkpoint exists |
| Demonstrated scale | Frontier models, billions of parameters | MNIST-class tasks on single devices |
What Would Have to Happen for Physical AI to Ship
Between a Perspective in a review journal and a component you can buy there is a long list of unglamorous steps, and it is worth being concrete about them.
A datasheet is the real milestone
Every commercial component is sold against specified tolerances. A physical AI device whose behaviour varies from unit to unit cannot be specified that way, so either fabrication has to tighten dramatically or the calibration has to move into the readout layer, fitted per device after manufacture. The second route is more plausible and it is already implicit in how reservoir computing is used.
Ageing has to be characterised, not just observed
Because a physical AI substrate holds the model in its own conductance states, drift is not a background nuisance — it is model decay. Any serious product needs a measured picture of how accuracy degrades over months of cycling, and a defined procedure for recalibrating the readout when it does.
The readout electronics decide the energy story
The efficiency claim for physical AI is made about the substrate. The figure that matters commercially is the whole module: stimulation circuitry, analogue-to-digital conversion, the trained readout and the interface to a conventional processor. Until someone publishes an end-to-end energy number, the comparison with a low-power accelerator is not settled.
Sensors are the obvious first socket
The most credible near-term use of physical AI is not replacing an accelerator but sitting next to a sensor: compressing a noisy time series into a handful of classifications at the point of measurement, so that far less data ever needs transmitting. Vibration monitoring, structural health sensing and satellite instrument data all have that shape.
Who Is Doing This Physical AI Work
The author list is a useful map of where physical AI research is concentrated, and it is unusually distributed.
Six authors, four countries
The Perspective’s corresponding author is Francesco Caravelli, a theoretical physicist at Los Alamos National Laboratory’s Condensed Matter and Complex Systems division who also holds a position at the University of Pisa. His co-authors are Gianluca Milano at INRiM in Turin, Adam Stieg at UCLA, Carlo Ricciardi at Politecnico di Torino, Simon A. Brown at the University of Canterbury’s MacDiarmid Institute, and Zdenka Kuncic at the University of Sydney.
A national-lab theorist leading a materials review
That the corresponding author is a condensed-matter theorist at a national laboratory, rather than a device engineer, says something about where the field’s bottleneck currently sits. The experiments have been running for over a decade; the framework for reasoning about them is the newer contribution.
Public funding on four continents
The work behind it is supported by the US Department of Energy, New Zealand’s MacDiarmid Institute and Marsden Fund, the European Union’s Next Generation EU programme and the European Research Council. There is no dominant corporate sponsor, which is worth noting when comparing the pace here with commercially funded accelerator development.
What Physical AI Means for Everyone Else
Nothing in this research changes what you should deploy next quarter. It does change how to read some of the claims you will be hearing.
It is a complement to cloud computing, not a replacement
The explicit framing is complementary: handle sensor data locally, in hardware, and leave the heavy training where it already happens. Anyone selling physical AI as a substitute for data-centre training is going beyond what the researchers claim.
The vocabulary is about to be borrowed
“Physical AI” is already used in industry to mean robots and embodied systems — a usage we have covered in reporting on embodied AI data standards. The materials-science sense in this paper is narrower and more literal, and the two will be conflated in marketing within months. Ask which one a vendor means.
The near-term pressure is still on small models
The practical version of the same energy argument is happening in software right now, with vendors squeezing capable models onto devices — the direction we examined in our coverage of tiny on-device language models. That is where the efficiency gains available this year are, and physical AI is the longer bet behind it.
What would count as real progress
Three things would move this from promising to practical: a device whose behaviour is reproducible enough to specify in a datasheet, a demonstration well beyond MNIST-class tasks, and a measured end-to-end energy figure including readout electronics. None exist publicly yet. When one does, it will be a much bigger story than this Perspective.
References
Toward physical AI: When the hardware becomes the neural network — UCLA Newsroom
Self-organizing memristive networks as physical learning systems — Nature Reviews Physics
Self-Organising Memristive Networks as Physical Learning Systems (preprint)
California NanoSystems Institute: Toward physical AI
Energy and AI, Executive Summary — International Energy Agency
AI is set to drive surging electricity demand from data centres — IEA
The MacDiarmid Institute for Advanced Materials and Nanotechnology
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.