What is photonic computing?
Photonic computing harnesses photons instead of electrons, delivering ultra-high bandwidth via wavelength multiplexing, near-zero latency and heat for data transfer, and major energy savings—especially ideal for AI matrix multiplications and data-centre interconnects.
In an era where artificial intelligence models demand ever-greater computational power while traditional electronic chips struggle with energy consumption, heat, and interconnect bottlenecks, a promising alternative is gaining serious momentum: photonic computing. Also known as optical computing, photonic computing represents a paradigm shift in how we process information. Instead of relying on electrons flowing through transistors—as in every conventional CPU, GPU, and AI accelerator today—photonic computing harnesses photons (particles of light) to perform computations, move data, or both. The Core Idea: Computing with Light At its heart, photonic computing replaces or augments electrical signals with optical ones. Light travels at roughly 300,000 km/s in a vacuum (and close to that speed in waveguides on chips), experiences far less resistance than electrons in wires, and can carry vastly more data per channel thanks to wavelength multiplexing (using different colours of light simultaneously). Key advantages include: Ultra-high bandwidth — Multiple wavelengths (colours) of light can travel the same waveguide without interference, enabling massive parallelism. Extremely low latency for data movement — Photons don't generate resistive heat the way electrons do. Dramatically lower energy use — Especially for data transfer and certain operations like matrix multiplications central to AI. Reduced heat generation — Critical for dense AI data centres facing thermal walls. These benefits make photonic approaches particularly attractive for AI accelerators, high-performance computing (HPC), and next-generation data-centre interconnects. A Bit of History The dream isn't new. Researchers explored optical computing as early as the 1960s, using lenses to perform Fourier transforms passively at the speed of light with zero energy for the computation itself. Through the 1980s and 1990s, free-space optical systems and early photonic logic gates showed promise, but integration challenges, lack of efficient optical transistors (nonlinear elements for logic), and the relentless progress of Moore's Law in silicon electronics kept photonics on the research fringes. The resurgence began around 2015–2020 with advances in silicon photonics — the ability to fabricate waveguides, modulators, detectors, and lasers on standard CMOS-compatible silicon wafers. Companies like Intel, Cisco, and Broadcom commercialised silicon-photonic transceivers for data centres, proving the manufacturing ecosystem exists. AI's explosive growth since 2022 supercharged interest. Training and inference for large models burn enormous power and create interconnect bottlenecks (the "memory wall" and "data-movement wall"). Photonics excels exactly where electronics struggle: moving huge volumes of data quickly and efficiently. How Modern Photonic Computing Works (2026 Perspective) Today's photonic computing is rarely "all-optical" in the pure sense. Most practical systems are hybrid — photonic elements handle what they do best while electronic circuits manage control, memory access, and precision logic. Common architectures include: Photonic interconnects / co-packaged optics (CPO) Light replaces electrical traces between chips, chiplets, or racks. This slashes power and latency for massive GPU clusters. NVIDIA has signalled plans to make silicon photonics and CPO mandatory for next-gen AI data centres by 2026. Photonic accelerators for AI Specialised chips perform matrix-vector multiplications (core to neural networks) optically via interference in waveguides or phase-change materials. Examples use time-wavelength multiplexing or free-space optics for analogue compute. Demonstrated systems achieve orders-of-magnitude better energy efficiency for certain workloads. All-optical or analogue photonic processors Emerging prototypes solve optimisation problems (e.g., NP-hard tasks) or run neural networks entirely in the optical domain, often with lasers, spatial light modulators, or integrated photonic circuits. Photonic quantum computing (a related but distinct branch) Photons serve as qubits for true quantum advantage. Companies like PsiQuantum, Xanadu, and Quandela pursue this path, but it targets fundamentally different problems than classical photonic computing. Leading Players in 2026 The field has moved from academic labs to serious commercialisation: Lightmatter — A frontrunner with its Passage 3D photonic superchip and interconnects. Valued at $4.4B+ after massive funding rounds, targeting AI infrastructure with production deployments expected soon. Neurophos, LightSolver, Celestial AI, Ayar Labs, Salience Labs — Focused on photonic AI accelerators or optical compute for specific workloads. Established giants — TSMC advancing silicon photonics manufacturing, Broadcom and NVIDIA integrating CPO. Startups raised hundreds of millions in 2024–2025, reflecting investor belief that photonics can help break through AI's energy and scaling walls. Challenges…
Read on MPC Markets Mosaic
