Enabling THz bandwidth on-chip with plasmonic modulators

Plasmonic electro-optic devices are emerging as a route to sub-THz and THz-class bandwidths, offering a compact, low-power alternative to traditional silicon photonic modulators for next-generation data center interconnects.

By Stephan Koch, VP Marketing & Sales at Polariton Technologies in Zurich, Switzerland. Polariton is now part of Marvell Semiconductors with a plan of bringing plasmonic modulators to data center applications

Growing demand for higher data rates in intensive data centre operations is accelerating the development of advanced system components, including electro-optic modulators designed to support THz-regime operation beyond conventional communications applications.

Next-generation modulation is increasingly a competition among material platforms, several of which have credible paths to adoption. Traditional III-V components, together with emerging thin-film lithium niobate (TFLN), barium titanate (BTO), and polymer-powered devices, are all expected to play a role. In an environment constrained by production capacity, multiple platforms are likely to coexist, with a clear advantage for solutions that combine cost efficiency with a scalable supply chain.

Plasmonic modulators use polymers as the non-linear material that produces the desired phase shift of coherent light. Their micrometer-scale dimensions are central to device performance, particularly in relation to bandwidth and power dissipation.

Plasmonics is silicon photonics
The modulator’s non-linear material is arranged coplanar to the silicon photonics waveguide within a slot bounded by two plates fabricated from silicon-compatible metals. Gold is commonly used, although other wafer-fab-compatible alternatives are also feasible. The plasmonic modulator is added during post-processing of standard silicon photonics wafers, which can be sourced from multiple silicon foundries [1], and requires only a limited number of lithographic steps. This approach enables electro-optic devices that provide:

• bandwidth suitable for sub-THz operation;
• reduced signal processing for equalization, or alternatively greater tolerance to noise in the communication channel;
• micrometer-scale active devices that support efficient use of chip area;
• lower switching power.

Within the active device, the electrical electrodes also function as the optical waveguide. This compact design is effective because the optical and electrical driving fields strongly overlap, resulting in high modulation efficiency and enabling short modulator lengths. Electrically, the modulator can be represented as an RC circuit with a corner frequency in the THz range, determined solely by the RC time constant.

From a physical perspective, the device width is smaller than the diffraction limit of the light used for communication, typically 1.3 or 1.55 µm. As a result, the relevant operating principle is plasmonic rather than purely photonic, allowing operation at smaller dimensions. The photonic mode is tapered into a plasmonic mode that propagates along the metal-dielectric interface of the slot before being converted back into a photonic mode.

These structures are sometimes referred to in the industry as plasmonic organic hybrid (POH) devices. In this article, however, we use the term plasmonic electro-optic modulator because the underlying physics relies on surface plasmon polaritons (SPPs). In 2025, researchers at ETH Zurich reported a record for an electro-optic modulator, achieving a 3 dB bandwidth of 997 GHz [2].

The prime application in the AI era
Plasmonic modulators have long been developed for fiber-optic communication, among other applications, because of their substantial bandwidth. The current increase in demand from data centers has placed silicon photonics at the center of efforts to address segments in which optical links can replace copper interconnects. Copper will nevertheless continue to coexist with optical technologies within servers and racks because of its cost and power-dissipation advantages, while optical links are expected to gain share wherever greater reach and bandwidth are required.

Within data centers, system-component dimensions, integration density, power dissipation, and communication protocols are increasingly important design considerations. As a result, equipment is becoming more specialized by application, and standards and implementation approaches are becoming more segmented. The industry commonly describes these segments as scale-in, scale-up, scale-out, and scale-across. At short reach, most communication remains copper-based; optical links become more prevalent as distances increase across racks and clusters. Hyperscalers have a clear interest in extending copper connectivity at the rack level for as long as practical, although the transition toward optical interconnects is becoming increasingly difficult to avoid.

Scale-across is a relatively recent segment definition, introduced approximately one year ago, and its precise boundaries remain under discussion across the industry. Historically, coherent communication was used primarily to connect hubs and data centers across regions. More recently, it has begun to move inside data-center premises, particularly in implementations that require lower signal-processing overhead and can benefit from the shorter link distances.

Silicon photonics implementations
Optical communication can be implemented using a range of modulators that are more or less native to silicon photonics. This arcticle compares Mach-Zehnder modulator (MZM) implementations in terms of their fitness for current and future communication speeds. The next generation of optical communication is expected to support 400G per lane, with data transmitted using four-level pulse-amplitude modulation (PAM4) and SerDes operating at 224 GBd, corresponding to 448 Gbit/s transmitted per lane. The principal advantages of plasmonic modulators arise from their micrometer-scale implementation. Smaller active devices can support very high operating speeds while reducing the energy required for charging and discharging. Their compact footprint also provides an integration advantage when many lanes must be placed in close proximity, as in co-packaged optics for scale-up applications. In addition, the flat transmission spectrum contributes to improved signal quality by reducing distortion along the communication path.

Exploring THz domain applications
The pull from AI-driven innovation is benefiting more applications. Several entities are pushing the operation into the THz regime. At Leapwave in Madrid, they are developing dielectric wires for communication and test & measurement, hence enabling boundaries to 500 GHz and more. Interfaces between the components are essential, and they have developed ultra-wideband approaches for that.

Another company, AttoTude in Menlo Park, is combining THz radio and the innovation of THz interconnect. The playground is the Terahertz gap, which is the part of the spectrum between what has traditionally been addressed by RF engineering and what is served with photonics.

Further, the EU project ECO‑eNET explores ultra‑high‑speed wireless links to bridge gaps between fiber networks using THz frequencies. Instead of converting signals to electrical, the approach keeps them in the optical domain to minimize losses. The goal is to achieve 200–300 GHz carrier frequencies.

Finally, at the Max-Planck-Institute for Plasma Physics in Greifswald, in collaboration with the Chalmers University of Technology in Gothenburg, they are pursuing higher magnetic fields in confinement fusion experiments by driving the electron cyclotron emission experiments well beyond the 100 GHz range. This is the result of the availability of cost-effective telecommunication technology for sub-THz-to-optical up-conversion.

Polymer development progresses fast
Electro-optic material manufacturers are also gaining increasing relevance, as their materials provide the essential non-linear response required for modulation. These materials are primarily responsible for converting electrical signals into optical signals, and vice versa, with new classes of polymers emerging as particularly well suited to this function.

Polymers are organic materials with strong optical and electrical properties, developed over more than two decades and now progressing toward commercial adoption. Earlier development focused primarily on dielectric performance, followed by substantial advances in easy of manufacturing and thermal stability at elevated temperatures.

Moreover the pace of polymer development exceeds that of new crystal engineering, creating significant potential for continued performance improvements over the coming years.

Enough bandwidth for generations

Silicon photonics has historically faced scrutiny in high-speed optical communication because it lacks native high-speed modulators. However, continued advances in both silicon photonics and integrated material platforms now enable 400G-per-lane implementations and position silicon as a strong foundation for the next several generations of transceivers. While multiple material platforms are competing for market adoption, silicon-based approaches retain important advantages in cost structure and supply-chain scalability. The shift toward smaller photonic integrated circuits (PICs), including ring resonator modulators (RRMs) and micrometer-scale plasmonic modulators, can also reduce wafer demand, supporting more efficient resource use in an industry constrained by manufacturing capacity.

More compact PICs also support higher levels of integration, including 32 or 64 lane architectures, which are becoming important for co-packaged optics (CPO) and near-packaged optics (NPO) implementations.

Polymers are also important enablers for high-speed modulators, offering strong dielectric and electro-optic properties that extend into the THz region. The rapid development of new polymer generations, together with the emergence of multiple material providers, is contributing to a broader reshaping of the industry.

Micrometer-scale plasmonic PICs provide the bandwidth required for 400G, 800G, and 1.6T transmission per lane, supporting multiple future generations of photonic transmitters. When combined with application-specific drivers and digital signal processors, they can also deliver advantages in switching power.