Photonic chips that never forget

A team at the University of Oxford built a photonic switch that remembers where it sent the light — even after you turn the power off. That sounds like a minor engineering detail. It is not. Every competing optical switch loses its configuration the moment power is removed; keeping light pointed the right way required a constant drain, which made large-scale reconfigurable optical systems impractical. Oxford's device, using a phase-change material called antimony triselenide (Sb₂Se₃), holds its setting indefinitely with zero standby power. The active switching region is more than 15 times smaller than competing approaches. It blocks unwanted light by more than 20 decibels — a measure of how completely the switch shuts off the path it isn't supposed to take — operates across a wide wavelength band used in datacenter communications, and weakens the passing signal by less than 2 decibels, according to Semiconductor Engineering. The result, published in Science Advances, is the combination the field has been waiting for.

The comparison to field-programmable gate arrays is earned this way: FPGAs gave chip designers a way to reconfigure hardware after it left the fab — closing the gap between fixed silicon and the flexibility of software. Photonics never had that moment. Optical switches configured at tape-out stayed that way until you physically rewired the system, because every alternative approach required continuous power to hold its state. The Oxford result does not merely improve an existing approach — it removes the static power requirement entirely, which is the specific barrier that kept programmable photonics from scaling beyond laboratory demos.

A companion preprint posted to arXiv in April 2026 by researchers at multiple institutions — under review at Nature Photonics — demonstrates Sb₂Se₃ optical switches on an 8-inch wafer-scale silicon nitride process, with extinction ratios of 25 dB, endurance exceeding 140 million switching cycles, and multi-level operation above 6 bits per switching element. That is where the fab-compatibility argument becomes concrete rather than aspirational: the integration path is documented, the process flow is described, and the numbers are reproducible in a format wafer fabs already work with. An earlier demonstration from Southampton, posted in November 2025, showed the approach scaling from simple 2×2 switches to 5×5 coupler arrays on standard silicon-on-insulator, using direct laser programming of phase-change pixels — confirming the technique is not unique to one lab.

The practical consequence is most concrete for AI accelerator architectures. Optical interconnects are already inside and between datacenter racks — co-packaged optics is the industry's chosen answer to bandwidth and power bottlenecks in copper traces at short reaches. What has not been practical is dynamic reconfiguration of those paths: routing light differently based on workload, adapting topology on the fly, implementing photonic meshes that behave more like programmable logic than fixed wiring. The static power requirement was the reason. Sb₂Se₃ removes it.

None of this is production-ready. The papers describe cleanroom devices, not commercial modules. The 140 million cycle endurance figure comes from a preprint not yet peer-reviewed. "CMOS integration ongoing" is not the same as "taped out at TSMC." The Oxford team's devices were measured in controlled conditions; how they perform across the full temperature range of a datacenter rack, over years of operation, with thermal cycling and aging effects has not been established at system level. No semiconductor foundry has announced a process node offering Sb₂Se₃ photonic switching as a standard option. There are no named industry partners in either paper.

The honest version of this story is: a material that worked in a cleanroom is showing the combination of properties the field has been waiting for, and it is being worked on at multiple institutions simultaneously — with enough momentum that the papers have shifted from "could this work" to "how do we integrate it." That is a different kind of signal than a single lab result. Whether it becomes the photonic FPGA moment depends on whether the fab integration actually closes, and that is a question for the next round of papers, not this one.