Light moves through glass without resistance, carries no heat, and can transmit billions of bits per second down a fiber thinner than a human hair. So why are we still doing the most computationally intensive work in human history using electrons squeezed through nanometer-scale silicon transistors, generating enormous heat, hitting physical walls that Moore’s Law can no longer paper over? The answer, for most of the last decade, has been integration: photonics was magnificent for moving data between machines, but getting it to do math inside a chip was another problem entirely. That problem is now cracking open.
The core idea behind optical neural network accelerators is straightforward and beautiful. Matrix multiplication, which is the overwhelming majority of what a large model actually does during inference, maps naturally onto the physics of light. When coherent optical signals pass through a programmable mesh of Mach-Zehnder interferometers, the interference patterns they produce perform the equivalent of a matrix-vector multiply in the time it takes light to cross the chip, which is picoseconds. No fetch-decode-execute cycle. No memory bandwidth bottleneck. The computation happens at the speed of propagation.
Lightmatter has been building toward this for several years, and their Passage interconnect fabric already ships to hyperscaler customers for moving data between conventional accelerators at dramatic bandwidth improvements over copper. But the more radical bet is Envise, their photonic processor aimed directly at inference workloads. The architecture combines photonic matrix engines with electronic control logic, a hybrid approach that sidesteps the trickiest unsolved problems (nonlinear activations, weight storage, precision) by letting photons do the linear algebra and silicon handle everything else. It’s not a pure optical computer, but that’s exactly the point: you don’t need purity, you need throughput per watt.
The energy argument is where this gets genuinely staggering. A conventional GPU doing large-batch transformer inference is moving data back and forth between HBM and compute cores thousands of times per forward pass, and that memory traffic dominates the energy budget. Optical matrix multiplication sidesteps a large chunk of that. Early published numbers from photonic accelerator research suggest potential efficiency gains of 10x to 100x for specific workloads compared to current-generation electronic accelerators, though achieving that at production scale and precision remains an active engineering challenge. Given that AI inference is on track to consume a significant fraction of global electricity within a few years, even a 5x improvement would reshape the economics of the entire industry.
The hard problems are real and worth taking seriously. Thermal stability matters enormously for photonic devices because tiny temperature fluctuations shift the interference conditions, and chips run hot. Precision is another frontier: optical systems naturally operate at lower numerical precision than the 8-bit or 16-bit arithmetic that modern inference engines take for granted, though research groups at MIT, Stanford, and several well-funded startups are making steady progress on calibration and error correction schemes that close the gap. Manufacturing yield for complex integrated photonic circuits has historically lagged electronic silicon, though the foundry ecosystem has matured substantially as telecoms and datacom applications have driven volume.
What makes this moment different from previous photonic computing hype cycles is the convergence of three things: foundry maturity from silicon photonics processes at TSMC and GlobalFoundries, genuine commercial urgency from inference cost pressure at scale, and a new generation of architectures explicitly designed around the constraints and strengths of optical hardware rather than trying to emulate what GPUs do. The design philosophy has flipped from imitation to native expression.
When the first photonic inference accelerators hit hyperscaler racks at meaningful scale, the story of AI hardware will have turned a page that’s been waiting to turn for a long time. The physics has always been pointing this direction. The engineering is finally catching up.