Broadcom Thor 2: 400G to Connect AI Brains

Published on May 18, 2026 | Translated from Spanish

The race to train larger artificial intelligence models demands moving data at breakneck speeds. Broadcom presents the Thor 2, a 400G Ethernet adapter designed to minimize latency in AI clusters. It's not an invention for watching Netflix faster; it is the highway through which machine thoughts travel.

Ethernet switch cluster backplane illuminated by pulsing data streams, optical transceivers glowing blue-green while transferring 400G signals between GPU server racks, nanosecond latency visualized as lightning-fast photon paths crossing silicon photonics chips, cooling fins vibrating from high-frequency processing, technical engineering visualization with cutaway view of fiber optic cables connecting AI brain nodes, motion blur on data packets, dramatic dark server room lighting with cyan accents, photorealistic industrial render

The architecture behind the Thor 2 🚀

The BCM57608 integrates a high-performance network controller supporting 400 Gbps Ethernet connectivity per port. Its architecture focuses on reducing end-to-end latency through an optimized packet processing engine and support for technologies like RoCEv2. It is designed to work with high-density switches, eliminating bottlenecks in GPU communication during distributed training.

When your home router begs for mercy 😅

While the Thor 2 moves 400 gigabits without breaking a sweat, your home router likely chokes on three YouTube streams and a video call. It's like comparing an F1 car to a rental scooter. Next time your connection drops, remember: there are network cards that move more data in one second than your provider will give you in a year.