A photon travels through silicon at the speed of light and consumes almost no energy doing it. That sounds obvious — it’s a photon, after all — but the implications for AI inference are staggering. Every matrix multiplication your GPU executes burns energy moving electrons through resistive metal. A photonic processor does the same multiplication by routing light through waveguides and directional couplers. The physics are so different that the efficiency comparisons start to feel almost unfair.

This is not a distant concept. Companies like Lightmatter and Luminous Computing have been building photonic interconnect and compute silicon for years, and the architecture is maturing fast. Lightmatter’s Passage interconnect fabric — already sampling with hyperscale customers — uses photonics not for the matrix math itself but for chip-to-chip communication, where bandwidth and latency are brutal bottlenecks at scale. That alone is significant. Moving data between thousands of accelerator dies is where enormous energy budgets go to die in conventional AI clusters. Replacing copper with light cuts both the power and the latency in ways that compound across the system.

But the more radical ambition is doing the matrix multiplications optically. This is where the physics gets genuinely exciting. An optical matrix-vector multiplier encodes weight values as the transmittance of Mach-Zehnder interferometers — tiny tunable waveguide structures — and inputs as the amplitude of incoming light pulses. The dot product happens passively, at the speed of light propagation through the device, consuming energy only for the tuning and for the photodetectors that read the result. For the actual multiply-accumulate operations that dominate transformer inference, you’re not burning switching energy at all. You’re using diffraction and interference, processes that would happen anyway.

The honest caveat is that analog photonic compute trades precision for speed and efficiency. Current integrated photonic processors typically operate at 4-8 bit effective precision, which is fine for inference on quantized models but not trivially compatible with the FP16 or BF16 training workflows the rest of the stack assumes. Thermal drift and fabrication variation in waveguide dimensions introduce noise that digital chips simply don’t have. The field is working through these problems with calibration schemes, error-correction, and hybrid digital-analog designs that offload only the most compute-intensive operations to optical hardware.

What makes the moment interesting is that the model side and the hardware side are converging toward each other. Quantization-aware training, which major labs now apply routinely, produces models that are increasingly tolerant of low-bit inference. Techniques like SmoothQuant and GPTQ have pushed high-quality inference down to 4-bit integer representations. A photonic processor that operates cleanly at 6-bit effective precision starts to look entirely sufficient for production inference on models that were explicitly trained with that budget in mind. The gap is closing from both directions simultaneously.

The energy numbers are what drive the urgency. Running a large language model at scale consumes megawatts. Inference datacenters are straining power grids. Any architecture that can deliver equivalent throughput at a fraction of the joules-per-token is not merely interesting to researchers — it changes the economics of everything built on top of AI. Photonic inference could mean running frontier models in edge devices with real battery budgets, or packing ten times the inference capacity into the same datacenter envelope.

The engineering challenges that remain — efficient light sources, low-loss fiber coupling, dense wavelength-division multiplexing for parallelism — are hard but they are not mysteries. They are problems with known physics and motivated teams. Photonic AI processors are already in the hands of customers for interconnect. The step to full optical matrix compute is the next arc of the same trajectory, and it is moving faster than most of the industry is watching.