Storage for the Server

The Optical Rewiring of AI is this layer's defining trend. As GPU clusters scale from thousands to hundreds of thousands of GPUs, electrical copper interconnects hit fundamental physics limits: signal loss, heat generation, and crosstalk all increase with cable length and data rate. The solution is optical — converting electrical signals to photons at the chip edge and transmitting over fiber.

NVIDIA's GB300 NVL72 rack requires 1.6T optical transceivers at every port. Each rack needs dozens of transceivers. A 100,000-GPU cluster needs hundreds of thousands. Coherent (formerly II-VI) and Lumentum make the indium phosphide (InP) laser chips that sit at the core of every transceiver. POET Technologies and Broadcom are working on silicon photonics alternatives that integrate optical components directly onto the chip — potentially cheaper at scale but not yet proven at the data rates required.

The supply chain stretch reveals the hidden constraint: InP substrates. Indium phosphide wafers are grown by a handful of companies — Sumitomo Electric (Japan), AXT Inc (US), and Wafer Technology (UK). The InP substrate market is tiny compared to silicon — perhaps $500M globally — but without it, no 1.6T transceiver ships. This is a classic chokepoint where a $500M input constrains a $50B network buildout.

Ciena and Semtech provide the DSP (digital signal processing) chips that clean up optical signals at the receiving end. As data rates push to 1.6T and beyond, DSP complexity increases quadratically — more computation per bit to compensate for signal degradation. This creates its own power and cooling burden, which feeds back into the 800V Power Architecture and Liquid Cooling Mandate trends.