Datacenters are, in one sense, very large machines for moving data. Processors, memory, and accelerators generate enormous volumes of information that must travel between chips, between servers, and between buildings. For short distances, copper traces and cables carry that traffic. For longer distances, fiber optics already dominate. The frontier is the middle: the links inside servers, between racks, and between chips, where copper is running out of headroom and optics is moving in.
The problem with copper is physics. As data rates rise, copper traces suffer increasing loss, crosstalk, and power consumption. Equalization and error correction can compensate, but at the cost of power and latency. Optical links avoid many of these losses because photons do not suffer resistive heating in the same way, and fiber has vastly more bandwidth than copper. The cost is conversion: data must be converted from electrical to optical and back again, which requires lasers, modulators, waveguides, and detectors, each with its own power, loss, and packaging costs.
Excitons enter this story because they offer a route to more compact, more efficient, and more integrated versions of each of those components. An excitonic laser could be smaller and lower-threshold. An excitonic modulator could be shorter and lower-power. An excitonic detector could be thinner and more easily integrated on silicon. None of these is yet a commercial product for datacenter interconnects, but the direction of research is clear: photonics is moving closer to the chip, and excitonic materials are one of the tools for making that move.
Why copper is struggling
The limits of copper are not a single cliff but a gradual erosion of margin. At data rates now common in datacenters, copper traces on a printed circuit board attenuate high-frequency signals, requiring equalization, retimers, and forward error correction. Each of those adds power and latency. Cables between racks are limited in length and must be replaced with active optical cables or fiber for longer reaches. As accelerator clusters grow, the fraction of system power spent on moving data — rather than computing — grows with it.
Optical interconnects solve the loss problem by moving data as light through waveguides or fiber, where attenuation is far lower and bandwidth is far higher. The challenge is that the optical components must be efficient enough that the savings in transmission are not consumed by conversion losses. A laser that wastes most of its electrical power as heat, a modulator with high insertion loss, or a detector with poor sensitivity can erase the advantage of optics.
This is why interconnect research is as much about energy per bit as about bandwidth. A link that moves data at high speed but at high energy per bit is not a solution for a datacenter that must move exabytes under a power cap. The goal is to reduce the total energy per bit — including laser, modulator driver, detector, and electronics — while increasing density and bandwidth.
The interconnect stack: lasers, modulators, waveguides, detectors
An optical interconnect has four core elements. The light source provides a continuous-wave laser, or in some architectures a directly modulated laser. The modulator encodes data onto the light. The waveguide or fiber carries it. The detector converts it back to an electrical signal. In a wavelength-division multiplexed link, multiple wavelengths travel through the same waveguide, multiplying bandwidth without multiplying waveguides.
Each element has its own requirements. The laser must be efficient, wavelength-stable, and capable of operating at datacenter temperatures without active cooling if possible. The modulator must be fast, low-loss, and low-power. The waveguide must have low loss and be compatible with packaging. The detector must be sensitive, fast, and low-noise.
Silicon photonics has made enormous progress on all four, using silicon waveguides, germanium detectors, and hybrid integration of III-V lasers. It is the incumbent platform for datacenter optics. Excitonic devices are not competing with silicon photonics as a whole; they are candidate components that could be integrated onto silicon photonics to improve specific elements.
Where excitons could help: sources
A laser for interconnects must meet stringent specifications: sufficient output power to overcome link losses, wavelength stability over temperature, low relative intensity noise, and the ability to be manufactured at scale. Vertical-cavity surface-emitting lasers, or VCSELs, already serve short-reach links. Edge-emitting lasers serve longer reaches. Directly modulated lasers simplify the link by eliminating a separate modulator.
Excitonic and polariton lasers offer the possibility of lower threshold and smaller footprint, which could reduce power consumption and enable denser integration. A polariton laser that operates without population inversion could in principle have a lower threshold than a conventional laser, reducing the electrical power required to produce a given optical output. A TMD or quantum dot laser integrated directly on silicon could enable on-chip sources without hybrid bonding of III-V chips.
The gap is manufacturing and reliability. Interconnect lasers must operate continuously for years at elevated temperatures, with stable wavelength and power. Laboratory polariton lasers often require cryogenic temperatures, optical pumping, or precise cavity alignment. Electrically pumped, room-temperature, wafer-scale excitonic lasers with the required stability remain a research goal, not a shipping product. For the near term, the more plausible impact is in improving existing III-V lasers through excitonic design — for example, using quantum dots or quantum wells with excitonic effects to improve efficiency or temperature stability — rather than replacing them outright.
Where excitons could help: modulators and detectors
Modulators and detectors are where excitonic materials may have the nearer-term impact, because the requirements are somewhat less stringent than for lasers and because integration on silicon is more straightforward. A TMD modulator on a silicon waveguide could be far shorter than a silicon Mach-Zehnder modulator, saving footprint and potentially reducing drive power. A quantum dot or TMD detector could be integrated on silicon without germanium epitaxy, potentially simplifying fabrication.
The key metrics are the same as for any interconnect component: insertion loss, extinction ratio, speed, and power. A modulator that is compact but lossy may not improve the link budget. A detector that is sensitive but slow may not meet the required data rate. The advantage of excitonic materials is that they offer strong light-matter interaction in a thin layer, which can in principle achieve the required modulation or absorption in a short device. The challenge is to achieve that without introducing excessive loss, variability, or temperature sensitivity.
Wavelength-division multiplexing adds another dimension. If excitonic modulators can be tuned electrically — by gating a TMD to shift its exciton resonance — they could provide tunable filters or modulators that select different wavelengths without requiring separate lasers. This is attractive for dense multiplexing, where many wavelengths must be managed in a small footprint. Tunability must be balanced against stability: a tunable device that drifts with temperature or bias will require active control, adding complexity.
Co-packaged optics and on-chip photonics
The most important trend in datacenter optics is moving optics closer to the compute. In pluggable optics, transceivers sit at the edge of a switch or server, and electrical signals travel over copper to reach them. In co-packaged optics, the optical engines are placed in the same package as the switch ASIC, shortening the electrical path and reducing power. In the longer-term vision of on-chip photonics, optical links connect chiplets or cores directly.
Each step closer to the chip reduces electrical losses but increases thermal and integration challenges. The optical components must operate in the thermal environment of a high-power ASIC, survive packaging stresses, and be manufactured with the yield and reliability required for datacenter deployment. This is why co-packaged optics is proceeding incrementally, with careful attention to thermal management, fiber attach, and known-good-die testing.
Excitonic materials are relevant here because they can in principle be integrated at the back end of line, at low temperature, on top of CMOS or on silicon photonics interposers. Van der Waals transfer, solution processing of quantum dots or perovskites, and low-temperature deposition of organics all avoid the high-temperature epitaxy required for III-V integration. The trade-off is that these integration methods are less mature, with lower yields and less uniformity than established silicon photonics processes.
Energy per bit: the metric that matters
For datacenter operators, the decisive metric is energy per bit, including all components of the link. A link that achieves high bandwidth but at high energy per bit will not be deployed at scale because the power budget is fixed. Reducing energy per bit requires improving laser efficiency, modulator drive power, detector sensitivity, and the electronics that drive them.
Excitonic devices could in principle reduce energy per bit in several ways: lower-threshold lasers reduce laser power, compact modulators reduce drive power, and sensitive detectors reduce the required optical power. In practice, the total energy per bit is dominated by the electronics — drivers, transimpedance amplifiers, serializers, and digital signal processing — as much as by the photonics. A photonic improvement that requires more complex electronics may not reduce total system power.
This is why system-level demonstrations matter. A component-level result showing low switching energy or high responsivity is necessary but not sufficient. A link-level demonstration showing a complete transmitter and receiver operating at the required data rate, with measured bit error rate and total power consumption, is the evidence that matters for datacenter deployment.
The economics of adoption
Datacenter optics is a high-volume, cost-sensitive market. Transceivers are manufactured in large quantities, with tight cost targets and stringent reliability requirements. A new material or device architecture must not only perform better but be manufacturable at scale, with yields and costs that compete with incumbent technologies.
This creates a high barrier for excitonic devices. Silicon photonics and III-V lasers benefit from decades of manufacturing investment, established supply chains, and proven reliability data. A new excitonic component must offer a compelling advantage — a large reduction in power, a large increase in density, or a capability that incumbents cannot provide — to justify the investment in new manufacturing processes and qualification.
The most plausible path is not wholesale replacement but incremental insertion. An excitonic modulator or detector that can be added to an existing silicon photonics process with minimal changes, and that improves a specific metric, is more likely to be adopted than a completely new platform. Similarly, an excitonic material that improves an existing III-V laser or detector is more likely to be adopted than a new laser architecture.
What to watch
For readers tracking this space, the signals to watch are not physics milestones alone but system and manufacturing milestones. Has an excitonic modulator or detector been demonstrated in a complete link at datacenter data rates, with measured bit error rate and power? Has it been integrated on silicon photonics at wafer scale with acceptable yield? Has it operated continuously at datacenter temperatures for an extended period? Has a commercial vendor announced a product or qualification effort?
Physics milestones — lower threshold, narrower linewidth, faster switching — are important, but they are steps toward these system milestones, not substitutes for them. The field is moving from single-device demonstrations toward integrated, system-level demonstrations, and that transition is the best indicator of whether excitonic interconnects will move from research to deployment.
The long-term bet is that photonics will continue to move closer to compute, and that the materials that enable that move will be those that can be integrated compactly, efficiently, and reliably. Excitons offer a physics toolkit for that integration. Whether that toolkit becomes a product depends on the unglamorous work of manufacturing, packaging, and qualification — work that is less visible than a physics breakthrough but decisive for datacenter adoption.
Sources and further reading
This article draws on literature on silicon photonics, datacenter interconnects, and co-packaged optics, as well as reviews of 2D material photonics and polariton devices. Useful sources include review articles in Nature Photonics, Nature Reviews Materials, Journal of Lightwave Technology, and IEEE Journal of Selected Topics in Quantum Electronics on silicon photonics, optical interconnects, and co-packaged optics, and industry white papers and roadmaps from datacenter and photonics consortia on energy per bit and link budgets. Specific claims about data rates, energy per bit, and link budgets change rapidly and should be checked against current industry roadmaps and product specifications; this explainer avoids citing single-vendor specifications as general facts.