Foxconn's research arm and National Yang Ming Chiao Tung University have demonstrated a silicon photonics transmitter capable of carrying 34.132 terabits of data per second over a single optical fiber, according to results published July 25 in the international optics journal *Optics Express*.
The work comes from a collaboration between the Foxconn Research Institute (FRI) — the in-house R&D division of Foxconn Technology Group (TWSE: 2317), the world's largest electronics contract manufacturer — and NYCU, a leading engineering university based in Hsinchu, Taiwan. Researchers say the result points toward a new technical approach for handling the data movement demands of next-generation AI infrastructure.
Data Movement, Not Compute Power, Now Bottlenecks AI Data Centers
As generative AI and large language models have driven rapid expansion of GPU clusters, a quieter constraint has become harder to ignore: the interconnects linking chips, servers, and racks increasingly cannot move data fast enough to keep those chips busy.
The problem compounds as deployments grow. The more accelerators a data center installs, the more those chips must communicate — and any lag in data delivery can leave even the fastest processors waiting. Bandwidth, not raw compute, has become the limiting factor in many large-scale AI workloads.
Traditional copper electrical interconnects have reached practical limits in speed, power consumption, and heat generation as throughput requirements climb. The industry has responded by accelerating adoption of silicon photonics, which encodes data onto light signals rather than electrical current, enabling higher bandwidth at lower energy cost per bit.
The FRI-NYCU project addresses a specific challenge within that shift: how to maximize the data capacity of each fiber and each light source.
A Single Comb Laser Emits 23 Separate Wavelength Channels
The system's most distinctive component is an O-band ultra-wideband quantum dot comb laser — a single device that simultaneously produces 23 evenly spaced wavelength channels. Conventional high-speed optical transmitters typically require one dedicated laser per channel, an approach that multiplies packaging complexity, power draw, and thermal management demands.
By consolidating light generation into one source, the team reduced the number of discrete components without sacrificing channel count. The researchers selected the 23 comb channels with sufficient signal-to-noise ratios to serve as independent transmission lanes.
Each of those 23 channels is encoded using PAM4 (four-level pulse amplitude modulation), a format that carries two bits per symbol by using four distinct amplitude levels rather than two. At 212 Gbit/s per channel, PAM4 substantially improves spectral efficiency compared to simpler two-state signaling schemes.
To convert high-speed electrical signals into optical form, the team integrated four Mach-Zehnder modulators (MZMs) fabricated on a silicon photonics platform. High-frequency GSSG electrode structures were incorporated to suppress electromagnetic crosstalk between adjacent channels — a prerequisite for reliable operation when many parallel high-speed lanes share the same chip.
Combining Wavelength and Spatial Multiplexing Reaches 34.132 Tbit/s
Beyond stacking wavelength channels, the research team introduced a second dimension of capacity: a seven-core multicore fiber, in which a single physical cable houses seven independent spatial waveguides that carry data streams simultaneously.
Multiplying 23 wavelength channels by seven spatial cores, each running at 212 Gbit/s, produces a theoretical aggregate throughput of 34.132 Tbit/s. The paper reports that after signals traveled two kilometers of fiber, average transmitter and dispersion eye closure quaternary (TDECQ) values held within 2.02 dB — a signal-integrity benchmark indicating the system maintained acceptable quality over that distance.
The researchers position this as a proof-of-concept demonstration. The 34.132 Tbit/s headline figure represents the combined upper-bound capacity across all 23 wavelengths and all seven cores, not a throughput rate that has been deployed or validated in a live commercial data center.
Foxconn Eyes Co-Packaged Optics, but Commercial Hurdles Remain
FRI described the result as laying groundwork for co-packaged optics (CPO) — an emerging design approach that physically relocates optical components to sit alongside switch ASICs or compute chips, shortening the high-speed electrical path and reducing the power loss and heat associated with conventional pluggable modules.
As AI clusters scale up, CPO has drawn growing industry interest as a way to raise bandwidth density without proportional increases in energy consumption. Consolidating multiple wavelengths into one comb laser, and substituting multicore fiber for multiple single-core cables, could shrink both the physical footprint and energy profile of optical interconnect systems if the approach can be scaled.
The distance from laboratory result to commercial product, however, spans a distinct set of engineering challenges: integrating the laser and silicon photonics die into manufacturable packages, achieving precision fiber coupling, sustaining packaging yield, managing long-term thermal stability and reliability, and reaching competitive unit economics at volume. FRI frames the research as validating capacity potential, not announcing product readiness.
The project was led by Kuo Hao-chung, director of FRI's semiconductor research division; group leader Hong Yu-heng; and researcher Chang Yun-han, alongside NYCU Distinguished Professor Tsou Chih-wei. Support came from Taiwan's National Science and Technology Council, the Industrial Technology Research Institute (ITRI), and a research team at National Chung Hsing University led by Professor Cheng Mu-hai.
FRI was established in 2020 with a mandate to pursue technologies expected to mature within three to seven years. The silicon photonics result signals Foxconn's intent to extend its presence beyond hardware manufacturing into foundational research on high-speed interconnects and semiconductor components. （Related： TSMC's Kumamoto Fab Gradually Restarts After Magnitude-7.1 Earthquake ｜ Latest ）
