What is optical interconnect in AI data centres?

The use of light (laser + fibre + photodetector) rather than electricity (copper cable + connector) to move data between AI accelerators and switches. Optical dominates at longer distances and higher bandwidths where copper's power-per-bit-per-metre curve becomes prohibitive.

Why is copper being displaced?

As AI accelerator power and per-link bandwidth scale (400G to 800G to 1.6T per link), the distance over which copper can carry the signal with acceptable power budget compresses. At 800G, copper's economical reach compresses to about 2-3 metres. At 1.6T, copper struggles beyond 1-2 metres.

What is CPO (co-packaged optics)?

Co-packaged optics integrate the laser, modulator and photodetector directly onto the switch or accelerator ASIC substrate, eliminating the pluggable transceiver housing. Lower power-per-bit (typically 30-50% claimed vs pluggable at equivalent bandwidth), higher port density, better thermal integration.

Which vendors matter for optical interconnect exposure?

Broadcom and Marvell for multi-layer exposure. Coherent and Lumentum for pluggable optical. TSMC and GlobalFoundries Fotonix for silicon photonics fabs. Ayar Labs, Lightmatter, Ranovus, Celestial AI for CPO specialists. InnoLight, Accelink, Eoptolink for Chinese pluggable cohort.

What is the TSMC + Nvidia CPO Photonics engagement?

Announced March 2026. Establishes TSMC as the silicon photonics foundry partner for Nvidia's CPO switch designs. Reference commercial engagement for the CPO commercial-scale ramp path.

Grid-to-Chip

Copper hits the wall in AI racks: the migration to optical, CPO and silicon photonics

2 September 2026 · 31 min read

More on

Battery role migration: from R1 backup to R6 grid-forming

Mount Diablo: The Open Rack Architecture That Forced NVIDIA Up the Stack

How firm AI load becomes a flexible grid asset

Get the next essay in your inbox. Free.