Silicon Photonics: Why AI is Forcing Data Centers to Abandon Copper

Building a 100,000-GPU data center isn’t just a software problem anymore; it’s a plumbing crisis. We are trying to push terabytes of data per second across rooms the size of football fields, and the physical medium we have relied on for a century—copper—is finally tapping out.

Before you read: We previously covered how AI’s massive physical footprint is reviving dormant nuclear reactors. This article explores the other physical limit breaking inside those facilities: the cables connecting the GPUs.

Table of contents

• The Copper Wall
• Firing Lasers at Silicon
• Co-Packaged Optics (CPO)
• The Manufacturing Nightmare
• Fun facts
• Sources

The Copper Wall

Copper has a fatal flaw at high frequencies: signal attenuation. To push 800 Gigabits per second (Gbps) over copper, the cables need to be incredibly thick, highly shielded, and strictly limited to lengths under three meters. It also requires massive amounts of power just to force the electrical signals through the metal’s natural resistance.

A modern AI server rack is choked by hundreds of pounds of Twinax copper cables. It restricts airflow, traps heat, and requires brute-force cooling systems just to keep the metal from melting the insulation. Anyone who has ever tried to neatly route a bundle of 400G DAC cables in a 42U rack knows it feels less like IT work and more like wrestling a nest of extremely rigid, spiteful pythons.

Firing Lasers at Silicon

The solution is Silicon Photonics. Instead of pushing electrons through metal, engineers use tiny lasers to shoot photons through microscopic glass waveguides etched directly into the silicon chips.

Light doesn’t suffer from electrical resistance. You can multiplex different colors of light on the exact same fiber, pushing terabytes of data over kilometers with nearly zero latency and a fraction of the power footprint. If copper is a physical freight train hauling coal across the country, silicon photonics is beaming the exact same payload via a laser pointer.

Co-Packaged Optics (CPO)

Historically, we used pluggable optical transceivers at the edge of the server board. But at the scale of modern AI, even the few inches of copper traces traveling from the GPU die to the edge of the motherboard are becoming a bottleneck.

The new frontier is Co-Packaged Optics (CPO). The lasers and optical engines are moving directly onto the GPU package itself. Light goes straight into the processor. Companies like NVIDIA, Intel, and Broadcom are pouring billions into this architecture because it is the only mathematically viable way to connect the next generation of AI supercomputers.

The Manufacturing Nightmare

Here is the grounded reality: making this work at scale is a manufacturing hellscape. Silicon is brilliant at computing, but it is physically terrible at emitting light.

To make this work, fabs have to bond Indium Phosphide lasers to silicon wafers with sub-micron precision. If a microscopic laser dies inside a co-packaged chip, you can’t just unplug a cable and swap a $50 transceiver; the whole $40,000 GPU package is compromised. The industry is effectively trying to mass-produce microscopic sniper rifles that never miss and never break.

Fun facts

• Did you know that a fully populated modern AI cluster can contain over 40 tons of copper cabling? The structural integrity of the raised floor becomes as much of an engineering concern as the network latency.
• Photons traveling through a silicon waveguide actually move at about two-thirds the speed of light in a vacuum, but they still generate virtually zero thermal footprint compared to high-speed electrical signals.

The AI infrastructure war isn’t just about who designs the fastest tensor cores anymore. The bottleneck has moved to the interconnects. The next trillion-dollar moat belongs to whoever can master the physics of light.

Sources

• IEEE Spectrum: The Silicon Photonics Revolution
• Light Reading: Co-Packaged Optics and the Data Center Power Crisis
• Yole Intelligence: Silicon Photonics Market Monitor 2024
• SemiAnalysis: AI Data Center Networking Bottlenecks