The case for photonic interconnects starts with a number that the datacenter industry calls "energy per bit" — the amount of electrical energy required to move one bit of data across a link. For a modern high-speed copper cable operating at 400Gbps over 3 meters, that number is roughly 6-8 picojoules per bit. For a silicon photonics transceiver at the same data rate covering 300 meters of single-mode fiber, it's closer to 2-3 picojoules per bit. At 800Gbps, the copper number climbs steeply because skin effect and dielectric losses increase with frequency. The photonic number improves, because higher data rates can be achieved by adding wavelength channels (WDM) without increasing the optical power per wavelength.
That asymmetry — copper gets worse with bandwidth, photonics gets relatively better — is the core physics argument. It's not about photonics being magical. It's about the fundamental electromagnetic properties of copper versus glass at the frequencies that matter for data communication.
Why Copper Loses at High Data Rates
Electrical signals in copper conductors lose power through two mechanisms that both scale unfavorably with frequency. The skin effect concentrates alternating current in a thin layer near the conductor surface, effectively reducing the conductor's cross-section and increasing resistance. At 28 GHz (corresponding to roughly 56 Gbps per lane with standard NRZ encoding), the skin depth in copper is about 0.4 micrometers. At 56 GHz (112 Gbps per lane), it's 0.28 micrometers. The resistance — and the resistive heating — increases accordingly.
Dielectric losses in the PCB material surrounding the conductor also increase with frequency. The dielectric loss tangent of standard FR-4 circuit board material is high enough at 56 GHz and above that it dominates the total link loss. Even with low-loss laminates designed for high-speed signaling, insertion loss at 56 GHz over a 15cm trace on a PCB is in the range of 15-20 dB — a significant fraction of the entire loss budget for a 400Gbps link.
These losses have to be compensated with equalizers — circuits that consume additional power to reverse the frequency-dependent distortion of the transmitted signal. At 400Gbps and above, the equalizer power consumption becomes a significant fraction of the total link power. Active electrical cables (AECs) — which embed equalization and retiming chips in the cable assembly itself — are increasingly common at 400Gbps precisely because the passive copper channel can no longer support the link without active compensation.
What Photons Do Differently
Light in a single-mode optical fiber has losses of approximately 0.2 dB/km at 1550nm wavelength — the telecom C-band. That loss is entirely wavelength-dependent, not bandwidth-dependent. A fiber carrying a single wavelength at 10 Gbps has the same per-unit-length loss as the same fiber carrying a single wavelength at 400 Gbps. The data rate scales by modulating the optical signal more aggressively, not by reducing the wavelength or increasing the optical power in a way that increases loss.
Wavelength division multiplexing (WDM) extends this further. A single-mode fiber can carry dozens of wavelengths simultaneously, each at 100-400 Gbps, for a total fiber capacity measured in terabits per second. The fiber infrastructure doesn't change — you add wavelength channels by adding transceivers at the endpoints and appropriate multiplexers/demultiplexers. The marginal cost of bandwidth is in the optical components, not in the fiber itself.
For intra-datacenter interconnects — the case where the distance advantage of optics is less pronounced — the argument shifts from per-bit energy to physical density and thermal management. Optical cables are thinner than electrical cables at equivalent bandwidth, which matters when you're routing 100,000 cables through a datacenter. Optical cables don't emit heat along their length the way copper cables do at high data rates, which simplifies thermal management in high-density rack environments.
The Modulation Format Question
Scaling optical data rates requires more sophisticated modulation formats. Simple on-off keying (OOK) — the original optical modulation scheme — transmits one bit per symbol. Pulse amplitude modulation with 4 levels (PAM4) transmits 2 bits per symbol by using four distinct optical power levels. PAM4 at 50 Gbaud gives 100 Gbps per wavelength. Currently deployed 400G coherent systems use polarization-multiplexed 16-QAM, which transmits 8 bits per symbol on each of two polarizations — 16 bits per symbol total. That's how you get 400 Gbps on a single wavelength while the symbol rate stays at 32 Gbaud.
More complex modulation requires more signal processing at both ends. The digital signal processor (DSP) for 400G coherent optical systems is a significant chip — tens of watts of power consumption, hundreds of millions of transistors. For long-haul links, this is clearly worth it. For short datacenter links below 500 meters, coherent optics are power-overkill. The right solution at shorter distances is intensity-modulated direct detection (IMDD) with PAM4 or PAM8 — simpler optics, less DSP, lower power, adequate reach for the application.
The industry is in the middle of a modulation format transition right now. 400G datacenter optics are predominantly PAM4 IMDD. 800G products are using PAM4 with higher symbol rates (100 Gbaud versus 50 Gbaud). 1.6T products will likely use PAM4 at 200 Gbaud or move to PAM8, depending on how well optical components scale versus digital equalization. The companies that understand this transition — and build components optimized for the specific modulation format and reach that dominates in 2026-2028 procurement — have a significant timing advantage.
Lumenwire and the Silicon Photonics Manufacturing Argument
Lumenwire, our San Jose portfolio company, is building silicon photonics interconnects specifically targeting the AI datacenter market. Their bet is on co-packaged optics at 800G and 1.6T — moving the optical engine from the front panel to directly adjacent to the switch ASIC, which reduces the electrical-optical interface losses that dominate pluggable transceiver power budgets.
The manufacturing advantage of silicon photonics for this application is that standard CMOS foundries can process silicon photonics wafers. The waveguides, modulators, and photodetectors are patterned using the same lithography and etch tools as standard chip manufacturing. You get the scale economics of semiconductor manufacturing applied to what had previously been a discrete component assembly process. At high volumes, the cost per port for a silicon photonics-based interconnect is fundamentally lower than for InP-based or traditional pluggable optics — not because of design elegance, but because of manufacturing economics.
The silicon photonics thesis isn't new. What's new is that the market timing has finally arrived. The AI infrastructure buildout is generating procurement demand at the scale that makes silicon photonics economics compelling. Lumenwire's current design wins are at two hyperscale operators whose next-generation AI clusters are specifying CPO requirements. Those wins represent volume production starting in 2026. The physics has been right for a decade. The market is right now.
Building photonic components, optical interconnects, or silicon photonics manufacturing tools? Reach the Coexin team.
