Broadcom Doubles AI Data Centre Speeds with Industry-First 400G Optical Chip

Broadcom Pushes AI Networking Into the Fast Lane

Broadcom has announced a sweeping expansion of its AI infrastructure portfolio, centred around a breakthrough optical chip that doubles the speed at which data moves through AI data centres.

The star of the show is Taurus, a tiny processor built on cutting-edge three-nanometre technology that can push four hundred gigabits of data per second through a single optical lane. That is double what current technology can manage. When assembled into a full switch, a single unit the size of a pizza box can handle over one hundred terabits per second of total switching capacity.

An Open Standard for Optical AI Connections

Perhaps equally significant is Broadcom's role in co-founding the Optical Compute Interconnect Multi-Source Agreement alongside Nvidia, AMD, Microsoft, Meta, and OpenAI. This consortium aims to create a universal optical standard that works across competing connection protocols, allowing data centres to swap copper cables for optical links between AI accelerators regardless of the vendor.

The specification will start at two hundred gigabits per second per fibre and scale toward three point two terabits, supporting multiple form factors from pluggable modules to co-packaged optics.

A Broader Portfolio Push

Beyond optics, Broadcom is showcasing its three point five D modular chip platform for custom AI accelerators, an eight hundred gigabit network interface card called Thor Ultra, next-generation Ethernet retimers, and PCIe Gen six switching solutions.

The announcements arrive just ahead of OFC twenty twenty-six in Los Angeles, where Broadcom and over thirty partners plan live demonstrations. With a sold-out exhibition floor and sixteen thousand expected attendees, the conference reflects the enormous demand for faster AI networking infrastructure.