A million optical neurons on one chip

Generative AI keeps growing hungrier. Diffusion models that turn text into images need ever larger matrices multiplied over and over, and those multiplications currently run on silicon that dissipates heat as fast as it computes. A team at Shanghai Jiao Tong University led by Yitong Chen argues that the work belongs in light instead. On 19 December 2025 they published, in Science, a chip called LightGen that performs large-scale semantic image generation optically, without shuttling the data back and forth into electronics for the heavy arithmetic.

The ambition is easy to understate. Optical computing has existed for decades, but demonstrations tended to handle one modest layer or a fixed classification task, which is why the chips stayed in papers and not in products. LightGen claims to cross three thresholds at once. It integrates about a million optical neurons on a single chip, converts data between optical dimensions entirely in the optical domain, and trains the network without the ground-truth labels that electronic back-propagation normally requires. Together these are the bottlenecks the field names whenever it explains why optical accelerators have not yet left the laboratory.

What the chip actually does is concrete. The team wired the optical processor to a diffusion model and generated images at resolutions of at least 512 by 512 pixels from semantic input, and extended the same optical pipeline to three-dimensional generation. The point is not that a photonics chip draws pictures prettily; the point is that the generative model, with its repeated, parallel matrix operations, was thought to be the kind of workload that only dense electronic hardware could serve. Showing that the whole synthesis can run in the optical domain reframes the workload rather than merely accelerating a slice of it.

The design borrows from the same group’s earlier ACCEL chip, reported in Nature in 2023, which handled image processing on a mostly electro-optical substrate. LightGen pushes further toward fully optical operation. Rather than convert signals to electrons for each layer, it keeps the light propagating and reconfigured on chip, which is what makes the million-neuron scale conceivable. Optical multipliers are passive; they do not consume dynamic power the way transistors do, and they operate at the speed of light through the medium. The trade-off, which the authors address directly, is that programming and training an optical network are awkward because light is hard to steer precisely.

That training problem is where the no-ground-truth result matters most. Optical neural networks cannot easily run the labelled-data back-propagation that electronics use, so the group developed an optical training scheme that adapts the chip without an explicit target. The method is the kind of unglamorous contribution that determines whether a demonstration becomes a platform. A million neurons are impressive once; a way to keep them useful over successive tasks is what survives contact with real workloads.

The honest caveats are also clear. A single chip generating small images is a proof of principle, not a data centre. Optical components still need supporting electronics to read out results and to control the phase shifters that configure the network, and the energy accounting at system level is not yet settled. Generative models also demand feedback loops that optics handle less gracefully than feed-forward inference.

What LightGen establishes, though, is a direction. Optical computing stops being a curiosity for small, fixed tasks and becomes a route for the parallel, matrix-heavy workloads that are currently straining electronics. If the training methods hold up at larger scale, the heat-bound arithmetic of generative AI may be exactly the problem that photons were waiting to solve.