The rule was straightforward for sixty years: get twice as many transistors if you wait two years. The modern world, including phones, search engines, and the notion that computing becomes more affordable and efficient on a timetable you could practically set your watch to, was made possible by Moore’s Law. It is no longer really holding. Nowadays, chipmakers are etching features that are only a few atoms wide, and there isn’t much more space for further shrinkage. There is no negotiation in physics.
In the meantime, the demand curve completely reversed. Climate models, real-time fraud detection, and AI training runs all require more processing power right away, ideally without melting the building. Regional power grids are already under stress from data centers. According to engineers I’ve read, AI’s demand for computing is doubling about every few months, a rate that no silicon roadmap has ever been designed to match.
At this point, light begins to appear intriguing—not as a trick, but as a truly unique physical substrate. Because they are charged, electrons push against obstacles and release heat as they travel. That is not what photons do. They don’t really interact with one another the way electrons do inside a transistor, and they lack mass and charge. That may seem like a drawback, and for many years it was, but it also means that light-based signals can travel at a fraction of the heat cost and with nearly no energy loss.
Years ago, fiber optics demonstrated this point by replacing copper internet lines with glass and gaining efficiency and speed in the process. Getting that advantage past the cable and onto the chip, where the real computing takes place, has always been a more difficult task. Companies like Lightmatter and Lightelligence can help with that.
Both have developed chips utilizing Mach-Zehnder interferometers, which split and recombine light waves to perform matrix multiplication, the unglamorous but crucial math that underpins the majority of contemporary neural networks. These devices are the result of MIT research. According to reports, Lightmatter has performed AI workloads several times faster and with significantly less power than comparable GPU systems.
However, it’s important to be truthful about the catch. Since there is currently no true photonic equivalent of RAM, optical chips are still unable to store memory as light. As a result, the majority of these systems continuously convert signals between optical and electrical, which reduces the efficiency gains that initially make light appealing.
Manufacturing is also harsh: the industry still lacks the kind of standardized design guidelines that made silicon so inexpensive to mass-produce, and alignment must be exact to a fraction of a wavelength. A startup called Akhetonics is pursuing fully optical logic, which uses materials like graphene to directly control light. According to its own roadmap, a functional all-optical processor won’t be available until around 2027. This isn’t coming next year, just based on that timeline.
Something messier and more gradual than a smooth transition from silicon to light seems more likely, at least for the time being. Data centers are already using optical interconnects to transfer data between racks and, increasingly, between chips. Full optical processors, which could eventually completely replace a CPU, are still more of a research project than a finished product. Before anything that resembles a true replacement for silicon actually appears on a shelf, it’s likely that the future will appear hybrid for a considerable amount of time, with electronics handling the logic and light handling the movement.
