Silicon Photonics Modulators Hit the Diffraction Limit — Marvell Breaks Through with Plasmonics: 10 µm, 1 THz, and a Lifeline Beyond 3.2T
Silicon photonics (SiPh) optical engines carried the entire pluggable-module era, but their core component — the modulator — is running into a hard physical wall: the diffraction limit. Today's mainstream silicon modulators are 3,000–5,000 µm long and run at only 60 GHz, a real space hog in a package where every square millimeter is precious. Through its acquisition of Polariton, Marvell has put a route around that wall on the table: plasmonics. By squeezing light below the diffraction limit at a metal-dielectric interface, it has built a modulator prototype about 10 µm long with operating frequencies approaching 1 THz — 1/300 to 1/500 the length and more than 10x the speed. Crucially, it is built directly on existing SiPh platforms and can be manufactured with foundry processes, with no need to switch to an entirely new material system. This isn't a lab toy; it's a realistic lifeline for silicon photonics modulators to survive beyond 3.2T.
1. AI Needs 3.2T, but Silicon Modulators Are Long and Slow
SiPh optical engines are the bridge between the electrical and optical worlds. By integrating hundreds of formerly discrete components onto a single piece of silicon, they dramatically cut the cost, power and size of optical systems and enabled the rise of pluggable optical networking. LightCounting forecasts optical module revenue will reach $64 billion by 2031, more than three times the industry's total revenue in 2025. But to feed AI infrastructure, optical engines must keep climbing to 3.2T and beyond.
The bottleneck is the modulator — the component that writes data onto the phase, amplitude and frequency of light. Its size is locked in by the diffraction limit: light cannot be confined in a waveguide narrower than about half a wavelength. As a result, today's silicon modulators are typically 3,000–5,000 µm long and run at only 60 GHz. Inside a module or chip package, that is an extremely expensive piece of real estate.
The problem isn't that lasers aren't bright enough or DSPs aren't powerful enough — it's that light itself is locked in by the diffraction limit, so no matter how much silicon modulators are optimized, they struggle to scale efficiently to higher bandwidth. Without a change, next-generation SiPh devices won't deliver the bandwidth AI infrastructure demands. We saw this earlier in The 300mm Silicon Photonics Comeback: A 500 µm Ultra-Compact MZM Targets 400 Gbps per Lane: even pushed to the limit of compactness, silicon modulators are still on the order of 500 µm. That is the ceiling the diffraction limit places on silicon.
2. Plasmonics Isn't a Platform Switch — It Squeezes Light Below the Diffraction Limit
Rather than starting over with a new material system, it performs an "interface surgery" on the silicon photonics platform.
Plasmonics relies on surface plasmon polaritons (SPPs). The device contains a metal-dielectric channel filled with an electro-optically active Pockels material, and the entire section is integrated directly into the SiPh device. When laser light hits the metal-dielectric interface, free electrons in the metal couple strongly with the incoming light to form SPPs.
This strong coupling does two things at once: it squeezes light into dimensions far smaller than its own wavelength, and it amplifies the local electric field. The direct result of a smaller optical mode and a stronger field is that the device responds faster to electrical signals and modulates more efficiently — and higher efficiency naturally means the device can be shorter.
Notably, the Pockels material filling the channel belongs to the electro-optic (EO) materials family. We covered the EO polymer track in Earnings Highlights: Lightwave Logic — AI Pushes Silicon Photonics to Its Limits, and EO Polymers Go from Decade-Long Backup to Lead Role; EO polymers are exactly the kind of Pockels material that can fill a plasmonic channel.
3. 10 µm and 1 THz: Just How Extreme Are These Numbers?
Marvell's plasmonic modulator prototype is about 10 µm long — 300 to 500 times shorter than existing SiPh devices — with operating frequencies approaching 1 THz, more than 10x faster than 60 GHz (Marvell's own estimate).
This shortness isn't a design compromise; it is the natural result of high modulation efficiency: the higher the efficiency, the shorter the interaction length needed for the same modulation depth. And the bandwidth is backed by measurement — tests show the plasmonic modulator's electro-optic bandwidth reaches 1 THz.
4. Bandwidth Density Soars — but Don't Ignore Channel Loss
The leap in modulation speed drives an exponential rise in bandwidth density (how much data can be moved per second per square millimeter) — exactly what packaging needs most.
To be balanced, there is real signal loss in the channel. But because the modulator is so short, the impact of that loss on signal quality and bit error rate is kept very small. That's the core trade of plasmonics: accept a little channel loss in exchange for extremely short length. The net result is THz-class speed in a micron-scale footprint, while inheriting the scale and cost economics of existing silicon photonics — letting designers extend the life of SiPh platforms without jumping to a new material system.
5. Can It Be Mass-Produced? The Metal-Dielectric Is Added at the Back End, So Foundries Can Build It
The single most important point about plasmonics for the industry: it is built on SiPh platforms and doesn't require moving to a new material system. All components can be produced with modified silicon processes, with the metal-dielectric added near the end of the flow. In other words, it's an incremental upgrade on the existing silicon photonics foundation rather than a teardown — good news for yield, cost and supply chain continuity.
It addresses both ends of the application spectrum. Inward (scale-inside/scale-up), co-packaged optics (CPO) can deliver the same or higher bandwidth in less area, in environments that are tightly constrained in space and power and sensitive to latency. Outward (scale-across), it enables smaller, more power-efficient ZR/ZR+-class modules that move more data over longer distances. Entirely new module form factors built around massively parallel plasmonic arrays are also cited as a possibility. Overall, this can mean higher utilization, less network congestion and lower total cost of ownership. For details on the scale-inside/scale-up/scale-across framework, see Related Reading below.
6. From Life Sciences to Optical Communications: Why Marvell Bought Polariton
Plasmons aren't new — they are used to accelerate drug development and boost sensor sensitivity, and even Roman-era artisans used them to make glass cups that change color with the light. Bringing them into optical communications was the work of Polariton and ETH Zurich, which have been pursuing this topic for more than 15 years.
The timeline is worth remembering: in 2022, they showed that plasmonic modulators can operate at the cryogenic temperatures required for quantum computing, territory where conventional optics struggles; last year (2025), they set a world record with a 1.1 THz plasmonic modulator; and in 2026, high-speed SiPh devices containing plasmonic transponders capable of 400G/lane were delivered to customers. Now this line of work belongs to Marvell.
This follows the same playbook as Marvell's string of moves to buy optical capabilities. We broke down that acquisition logic in Celestial AI Explained: Marvell's $3.25 Billion Bet on Bringing Light into the Chip: rather than waiting for technology to mature on its own, buy the key IP and team outright and fold them into its own scale-up roadmap. Polariton is the modulator version of the same move.
7. Conclusion
What plasmonics offers isn't "a bit faster" — it's a way around the diffraction-limit wall, something silicon modulators can't achieve through optimization alone. It shrinks the modulator from millimeters to microns and pushes bandwidth from GHz to THz, and because it can be added at the back end of existing silicon photonics processes, it preserves manufacturability and cost viability. For a module market projected to reach $64 billion by 2031, this is the most important piece for extending SiPh beyond 3.2T.
But let's not overstate it. Marvell itself says the technology is still at a very early stage of commercialization: the next phase focuses on solidifying device reliability, performance and high-volume manufacturing methods. The prototype is impressive, but whether it can be produced in volume at stable yields is the real test for plasmonics on the path from world record to shipments. For Taiwan's supply chain, the question to track now isn't whether plasmonics will win, but rather — as modulator materials branch out from pure silicon to Pockels materials, TFLN and plasmonics in parallel — where each company can secure a position across epitaxy, process, packaging and test.
This article is for technology and industry trend analysis only and does not constitute investment advice.
References
Marvell Blog, "Plasmonics: A Path to Higher Bandwidth in Optics in the AI Era," Claudia Hoessbacher, Wolfgang Heni, 2026-06-17. Original article
LightCounting Market Forecast, April 2026 (forecast of $64 billion in optical module revenue by 2031)
D. Moor et al., IEEE, November 2024 (measured 1 THz electro-optic bandwidth of a plasmonic modulator)
Ueli Koch, ETH Zurich Research Collection, 2019 (plasmonic device with integrated driver)
Polariton, March 2025 (1.1 THz plasmonic modulator world record)
ETH Zurich, Polariton and NLM Photonics, September 2022 (cryogenic electro-optic modulation record)
Related Reading
TFLN Breaks Another Record, but Don't Misread the Signal: Thin-Film Lithium Niobate's Opportunity and Three Walls in Optical Communications/CPO: beyond plasmonics, another material route around silicon modulator limits — and where its three walls lie.
The Great Optical Packaging Transition (Part 2): CPO's Three-Stage Evolution — Why OBO Died, Scale-Out Moves First, and Scale-Up Is the Endgame: understand the scale-inside/scale-up/scale-across framework to see where plasmonic engines fit.
Earnings Highlights: Marvell (MRVL) | FY27 Q1 — From Following CapEx to Defining the Shape of Scale-Up: the financial context behind Marvell's string of acquisitions and scale-up strategy.
Comments