The constraint is an operation rather than a result: a Markov chain does not divide across parallel lanes, and over the same period the idea kept developing in the parameterisation whose gradient costs a single forward pass, which is what a diffusion model is. Physical AI is where that cost is most visible, because the dominant visuomotor policy runs Langevin dynamics at inference for 10 to 100 network evaluations per action, and its own literature names that as the barrier to real-time deployment. The crossover against a measured 100 ms embedded sampler already falls inside that range. The two cases ask different things of the same physical property: an in-memory matmul treats device fluctuation as error to be suppressed, and a sampling fabric treats it as the source of randomness to be characterised.
One of eight, and only one of them is physics. How we read a frontier →
The algorithm you already run is an energy-based model.
In 2020 Sara Hooker named a mechanism she called the hardware lottery: an idea wins because it suits the available hardware, not because it is better, and lotteries can cast successful ideas as failures. Energy-based models are such a case, and the argument is mechanical rather than rhetorical. Fitting one needs samples from an unnormalised distribution, which needs a Markov chain, which is sequential, and a graphics processor accelerates work that divides into independent pieces. Backpropagation divides perfectly; a chain does not divide at all, so lane count does not enter its wall-clock time. Over the same period the idea kept developing in the parameterisation whose gradient costs a single forward pass.
A robot can go left or right around an obstacle and both are correct. A policy trained by minimising squared error against demonstrations of both returns their mean, and the mean of left and right is straight ahead, into the obstacle. More demonstrations make the mean more exactly wrong. Our own bench reproduced it cold: on a genuinely multivalued system a fairly supervised feed-forward network scored 0% at every model size tried. An energy landscape holds each route as a separate minimum and has to choose one, which is the whole difference.
The dominant visuomotor policy class runs stochastic Langevin dynamics at inference, at a published cost of 10 to 100 network evaluations per action, and its own literature names that as the key barrier to real-time deployment. So Physical AI's dominant algorithm is a sampler running on hardware that cannot sample. Against a measured 100 ms end-to-end latency on an embedded 2,048-spin sampler, the crossover falls near 33 evaluations, which is inside the range the field already uses.
A second workload is arriving on the same silicon: transformer attention is being mapped into memory arrays, with large reported gains. The two cases use the device differently, and the difference is physical. An in-memory matrix multiply is acceleration and treats device fluctuation as a source of error to suppress. A fabric that settles into a Boltzmann distribution is realisation, and treats the same fluctuation as the source of randomness to characterise. The variability and drift that make analog arithmetic difficult therefore enter the two programmes on different terms.
Efficiency here is stated in three denominators, and a figure is comparable only within one. Energy per operation describes a device. Energy per completed task on a body includes actuation, which does not move when compute improves, so it carries a limit of 1.6 to 2.4 times. Time per action is a latency against a control period, and that is the denominator the embodied case is developed in, together with duty cycle and thermal envelope.