In a data center dedicated to artificial intelligence, computing is only part of the job. Processors also need to be fed, their results synchronized, and torrents of information moved between machines. These journeys consume energy and generate heat. Light promises to reduce that energy bill. But behind “photonic chips,” two stories intersect: the already industrialized world of optical communications and the more experimental field of computing with light. Looking ahead to September 2026, distinguishing between them remains essential to understanding what can actually change.
The problem is not just computing faster
Imagine a kitchen equipped with hundreds of very fast chefs but served by congested corridors. Adding more cooks is no longer enough: the flow of ingredients needs to speed up. Large AI infrastructures face a similar challenge. Their accelerators exchange parameters, activations, and intermediate results, while memory must feed the compute units.
On electrical links, increasing data rates and distance makes transmission more complex. Losses, signal distortion, and the circuitry needed to restore signals drive up power consumption. Yet copper remains remarkably effective over short distances: it is well understood, economical, and easy to integrate. The question, then, is not how to eliminate it everywhere, but where light becomes the better option.
Optical interconnects: bringing fiber closer to the processor
Optics has long carried data through telecommunications networks and between computing devices. The new development is bringing optical components closer to the chips that produce and receive data. A modulator encodes information onto a light wave; a photodetector converts it back into an electrical signal at the destination. Between the two, a fiber or waveguide carries it.
In a conventional architecture, the electrical signal still travels across part of the circuit board before reaching a pluggable optical module. With co-packaged optics, optical engines are integrated into the same package as the electronic chip, or placed as close to it as possible. The aim is to shorten the most costly electrical segments. Other approaches use optical chiplets paired with processors.
The milestones are tangible. In 2024, Intel presented an optical interconnect chiplet co-integrated with a processor in a public demonstration. That same year, Lightmatter unveiled Passage, a photonic interconnect platform designed for computing infrastructure. Ayar Labs is also developing on-chip optical input/output. These announcements demonstrate real industrial activity; on their own, they do not prove widespread, cost-effective deployment.
A benefit that must be measured end to end
Light notably allows several wavelengths to travel through the same medium, like multiple lanes on a highway. This can increase available bandwidth without a proportional increase in the number of connections. But an optical link is not energy-free: the laser must be powered, modulators driven, signals detected, and sometimes the temperature stabilized.
Comparing only the energy consumed in an optical waveguide with that of an electrical cable would therefore be misleading. The right measure covers the entire link, at comparable data rates, distances, and transmission quality. It must also account for utilization. A technology that performs exceptionally well at full load may become less attractive when infrastructure operates below capacity for long periods.
Photonic computing: when light performs the operation
Photonic computing pursues a different goal. It is no longer simply about moving bits, but about using the properties of light to perform certain operations. Networks of interferometers or other optical components can carry out mathematical transformations, particularly matrix multiplications. These operations are ubiquitous in neural networks.
The promise is compelling: processing information in parallel, with very rapid propagation and potentially low energy consumption at the core of the device. Companies such as Lightelligence, along with many research laboratories, have demonstrated optical computing architectures. However, successfully performing an operation in a laboratory is not the same as replacing a GPU in a data center.
Many of these approaches involve analog computing. Information is represented by a physical quantity, such as the intensity or phase of light. Noise, manufacturing variations, and thermal drift then limit precision. Obtaining a sufficiently reliable result may require calibration, correction, or multiple passes, reducing the initial advantage.
The conversion and memory trap
Data generally resides in electronic memory. It must therefore be read, encoded for the photonic circuit, and the result then retrieved. Conversions between digital and analog signals, as well as between electricity and light, can consume a significant share of the total energy. And the nonlinear operations required by neural networks are still often handled by electronics.
The most credible scenario is therefore not an entirely light-based computer, but a hybrid system. A photonic block would accelerate a carefully chosen task, while electronics would handle memory, control, and other computations. Its value would depend on how much work is accomplished between conversions: a few isolated operations offer less opportunity to offset their cost than large-scale processing.
Two markets, two timelines
For interconnects, the need is already clear: connecting more accelerators with high bandwidth without sending network power consumption soaring. The debate centers on integration, cost, reliability, and the point at which optics becomes preferable. For photonic computing, it remains to be established which applications can derive lasting benefits from a specialized accelerator compared with electronic chips, which are also improving.
In both cases, manufacturing matters as much as physics. Aligning fibers, assembling components, testing links, and ensuring they withstand a hot environment are industrial challenges. The proximity of electronics and optics also raises a maintenance question: replacing an accessible module does not have the same implications as repairing a highly integrated assembly.
When evaluating an announcement, three questions help move beyond the spectacle:
- What is the scope? Does the improvement apply to a component, a complete link, or an application running end to end?
- What is the comparison? Does the electronic baseline deliver equivalent precision and service?
- How ready is it for manufacturing? Is this a demonstration, customer samples, or repeatable production?
Lower energy use is possible, not automatic
Reducing the energy needed to transport a bit or perform an operation does not guarantee a fall in overall consumption. If the savings make it possible to build larger models or multiply the number of queries, demand can absorb the gains. Photonics is a tool for improving efficiency, not an excuse to avoid thinking about how infrastructure is used and sized.
What comes next? For September 2026 and beyond, the most solid prospect is that light will gradually move closer to processors, starting where communications become a bottleneck. Photonic computing could find specialized applications, provided it can demonstrate an advantage that includes conversions, memory, and software. The revolution may be less visible than a “light-based processor”: more data transported usefully, with less energy consumed.
