Cisco's new AI chip can connect 128,000 GPUs
Cisco just announced its Silicon One G300 chip, built for massive AI clusters and ready to handle training, inference, and real-time tasks at lightning speed.
With a whopping 102.4 Tbps switching capacity, it powers Cisco's new liquid-cooled Ethernet systems—delivering almost 70% better energy efficiency.
The G300 will hit the market in late 2026.
The G300 packs 512 ultra-fast 200 Gbps SerDes
The G300 packs 512 ultra-fast 200 Gbps SerDes that can be aggregated (8x200Gbps) to support port speeds up to 1.6 Tbps, yielding 64 1.6 Tbps ports per chip.
That means you can connect up to 128,000 GPUs using only 750 switches (way down from the old 2,500).
It also features smart networking tricks like fast hardware-based load balancing and shared packet buffers that help keep things running smoothly—even when data traffic spikes.
Cisco claims its chip is faster than the competition
With new 800 Gbps linear pluggable optics that eliminate onboard DSP/retimers to reduce power and slash switch energy needs by about 30%, the G300 is designed to be greener too.
While its raw speed matches chips from Broadcom and NVIDIA, Cisco says their design gets jobs done faster thanks to better buffering and smarter hardware-driven traffic management.