Photonic Computing: Technology, Commercial Case and the Future of Traditional Compute
Photonic computing uses light to transmit or process information. Its attraction is not that photons make every calculation faster. It is that optical systems can carry many high-bandwidth channels in parallel while reducing the electrical losses associated with moving data…
What photonic computing is, why it matters now, and the central commercial conclusion Photonic computing uses light to transmit or process information. Its attraction is not that photons make every calculation faster. It is that optical systems can carry many high-bandwidth channels in parallel while reducing the electrical losses associated with moving data across increasingly large AI systems. The investment case divides into two distinct markets. Optical interconnects replace electrical links between chips, packages, racks or memory pools. Optical computation uses photonic circuits to perform mathematical operations, especially matrix multiplication. The former is closer to broad commercial adoption because communications is already photonics' strongest domain. The latter remains workload-specific and must overcome precision, conversion and software constraints. The central commercial conclusion for MPC Markets clients is that photonics is more likely to complement conventional processors than replace them. Near-term value should accrue around optical I/O, lasers, packaging, foundry integration and hyperscale deployment. Photonic accelerators may become valuable for large, regular linear-algebra workloads, but optical-core efficiency should not be confused with full-system efficiency. Key Takeaway: The highest-confidence thesis is connect first, compute selectively. Photonics can expand the effective scale of electronic computing, while CPUs, GPUs and memory remain responsible for control, storage, nonlinear functions and general-purpose execution. The Compute Bottleneck How AI workloads, data movement, memory bandwidth, power consumption, and slowing CMOS scaling create an opening for photonics AI infrastructure is increasingly limited by the cost of feeding accelerators rather than by arithmetic capability alone. Parameters and intermediate results must move among high-bandwidth memory, accelerator packages and networked servers. As electrical link rates rise, signal conditioning, retiming and heat removal consume more power and package area. The International Energy Agency estimates that data centres consumed about 415 TWh in 2024, approximately 1.5% of global electricity use, after growing at around 12% annually over five years. It also notes that a typical AI-focused data centre can consume as much electricity as 100,000 households, while the largest facilities under construction may consume substantially more. These comparisons are indicative rather than universal, but they establish power availability as a constraint on deployment. CMOS scaling continues, but improvements in transistor density no longer translate automatically into proportional system-level gains. Memory access, communication and cooling increasingly determine usable performance. Photonics creates an opening because wavelength division multiplexing can place multiple data channels on one waveguide, while optical links can maintain high bandwidth over distances that become difficult for copper. This does not mean optics eliminates energy use. Lasers, modulators, receivers and electro-optical conversion all consume power. The relevant comparison is the energy and cost of the complete link or computation at the required reach, precision and utilisation. How Photonic Computing Works Encoding information in light, integrated photonic circuits, optical matrix multiplication, wavelength division multiplexing, and the continuing role of electronics Hybrid photonic computing flow from electronic data through wavelength-encoded optical matrix multiplication and back to electronic control and nonlinear processing. Values can be encoded in a light wave's amplitude, phase, wavelength, polarisation or spatial path. Modulators translate electronic inputs into optical signals. Waveguides route those signals through interferometers, resonators or other weighting structures. Photodetectors then convert the resulting light back into electrical current. For matrix-vector multiplication, input values are encoded onto optical channels and combined with programmed weights. Interference and photodetection perform weighted accumulation using the physical propagation of light. Multiple wavelengths can process several channels concurrently, creating substantial parallel throughput. The system remains hybrid. Electronic memory stores weights and activations. Digital-to-analogue converters drive modulators, while analogue-to-digital converters may digitise detector outputs. Electronics provide scheduling, accumulation, calibration and nonlinear activation functions. These boundaries matter because an extremely efficient optical multiply can sit inside a system dominated by conversion, memory and laser power. Optical computation is therefore best understood as a specialised dataflow engine, not a drop-in optical CPU. Quantum photonics is an adjacent field based on different physical and commercial assumptions and is not interchangeable with the classical…
Read on MPC Markets Mosaic
