TSMC COUPE silicon photonics platform expected to ramp in 2028 for NVIDIA Feynman products
News

TSMC COUPE silicon photonics platform expected to ramp in 2028 for NVIDIA Feynman products

Tuesday, March 10, 2026 at 02:18 AM

TSMC is developing its Compact Universal Photonic Engine (COUPE) technology to integrate silicon photonics into advanced packaging using SoIC bonding. This co-packaged optics (CPO) solution stacks electronic ICs on photonic ICs to reduce power consumption and latency compared to traditional copper interconnects. The technology is expected to be paired with CoWoS for high-density AI accelerator integration. Analysts anticipate a high-volume ramp by 2028, coinciding with NVIDIA's Feynman series, including NVLink 8 and Spectrum 6/7 switches, as system bandwidth requirements exceed 200T.

Context

Recent reports from Citi Research indicate that TSMC’s COUPE (Compact Universal Photonic Engine) platform is expected to reach high-volume production in 2028. This timeline aligns with the launch of Nvidia’s Feynman product series, which is anticipated to feature NVLink 8 and Spectrum 6/7 networking technologies. The COUPE platform represents a shift from traditional copper interconnects to co-packaged optics (CPO), utilizing SoIC 3D stacking to integrate electronic and photonic circuits directly. This transition is designed to address the physical limits of copper by reducing power consumption and cutting latency by an order of magnitude. While Nvidia showcased its first-generation COUPE optical engines at GTC 2025 for the Rubin architecture, the 2028 ramp-up for Feynman marks a critical evolution toward "optical-compute systems." By moving optical engines from pluggable modules to the interposer level, TSMC and Nvidia aim to support system bandwidth exceeding 200T. Industry analysts note that while pluggable transceivers currently remain the standard, the aggressive scaling of AI accelerator bandwidth to 400G per channel and beyond makes CPO an essential enabler for future intra-rack and intra-package data movement.

Related Companies

TSMC
TSMC
2330
TW
Nvidia
Nvidia
NVDA
US
Meta
Meta
META
US
Apple
Apple
AAPL
AMD
AMD
AMD
US
Amazon
Amazon
AMZN
US