Silicon Photonics

Silicon photonics is a technology that uses silicon-based photonic integrated circuits to generate, modulate, transmit, and detect light, enabling data to be carried optically instead of electrically. In the context of artificial intelligence, silicon photonics has become a critical enabler of AI data centre interconnects: as GPU clusters scale, copper electrical links struggle to deliver the bandwidth and power efficiency required, and optics — increasingly co-packaged directly with switch and compute chips — is emerging as a structural requirement for next-generation AI infrastructure.^[1][2]

History and Background

Silicon photonics research dates to the 1980s, with the goal of fabricating optical components using standard silicon manufacturing processes. The technology first found commercial success in data centre optical transceivers during the 2010s, replacing slower electrical interconnects in high-speed networks as Ethernet speeds moved from 100G toward 400G and 800G.^[2]

The AI boom sharply accelerated the technology's importance. At its GTC conference on 18 March 2025, NVIDIA announced NVIDIA Photonics, a silicon photonics platform that replaces traditional pluggable transceivers with co-packaged optical devices (CPO), enabling direct connection between fibre optic cables and switch chips and reducing data centre power consumption by approximately 40 megawatts while improving network efficiency for AI clusters.^[3] In 2026 NVIDIA deepened the bet: on 2 March 2026 it announced multiyear strategic partnerships with Coherent Corp. and Lumentum Holdings, investing US$2 billion in each company alongside multibillion-dollar purchase commitments to expand US-based manufacturing of lasers and optical networking products — a signal that optics had become a gating factor on AI infrastructure scaling.^[4]

Key Concepts and Technology

Photonic integrated circuits on silicon: Silicon photonics builds optical components — waveguides, modulators, photodetectors, and micro-ring resonators — on standard silicon wafers using CMOS-compatible fabrication, enabling mass production at semiconductor scale.^[2]

Co-packaged optics (CPO): CPO integrates optical engines directly into the same package as the switch or compute ASIC, eliminating separate pluggable transceiver modules, the DSP chips they contain, and long power-hungry electrical traces. Vendors including Broadcom and Cisco claim CPO can cut switch power consumption by around 30% in 51.2T switches with 64 800GbE ports, with power savings of 30–50% reported relative to pluggable optics.^[3][5]

Micro-ring modulators (MRMs): NVIDIA's approach, developed with TSMC using the COUPE 3D packaging technology, integrates 200G PAM4 micro-ring modulators onto the silicon photonics die to maximise optical engine bandwidth density while keeping footprint and power low.^[1][6]

Adoption trajectory: TrendForce estimated that CPO would account for only about 0.5% of optical transceiver modules used in AI data centres in 2026, with adoption expected to accelerate in NVIDIA's Rubin generation for scale-out inter-rack transmission, and forecasts that silicon-photonics-based CPO solutions could reach approximately 35% penetration of AI data centres by around 2030.^[1]

Applications and Impact

Silicon photonics is used in data centre optical transceivers (200G, 400G, 800G, and emerging 1.6T modules), InfiniBand and Ethernet AI cluster networking, and emerging scale-up architectures that replace electrical NVLink with optical interconnects between GPUs. Beyond data centres, the technology is found in wearables, short-wave infrared (SWIR) sensors for health and industrial monitoring, and optical I/O research.^[2][5] NVIDIA positions CPO not as an optional enhancement but as a structural requirement for future AI data centres, making silicon photonics one of the defining infrastructure technologies of the AI era.^[6]

>See Also

• NVIDIA
• GPU Cluster
• TSMC
• High-Bandwidth Memory
• Liquid Cooling
• AI Data Centres Malaysia
• National Semiconductor Strategy Malaysia

References