Cisco has launched the Silicon One G300, a 102.4Tbps networking chip built on TSMC’s 3nm process. The chip supports 64 1.6Tb Ethernet ports, doubling the capacity of its predecessor. According to Martin Lund, Cisco’s EVP of Common Hardware, only three companies can produce chips at this level: Cisco, NVIDIA, and Broadcom.

Pushing physical limits comes with a cost—liquid cooling is required. Lund noted that the chip is near the maximum size manufacturable today, and that power density now demands liquid cooling for switching gear.

Programmability sets the G300 apart. The chip can be reconfigured after deployment to adapt to evolving AI workloads and protocols. For AI factories running tens of thousands of chips, this flexibility translates directly into operational and financial advantages.

Ethernet has won the AI networking standard war. InfiniBand’s address space tops out at 65,000 nodes, far short of the hundreds of thousands of GPUs in large AI clusters. Ethernet delivers the scalability, interoperability, and ecosystem support that large-scale deployments require.

Lund revealed that five of the six largest hyperscalers are already using Silicon One technology, with the sixth in progress. The market is expanding beyond traditional hyperscalers to include neocloud and sovereign cloud projects.

Looking ahead, silicon photonics and co-packaged optics (CPO) could cut power consumption by up to 70%. But true all-optical packet switching remains years away.

ICgoodFind : Cisco’s 102.4T G300 packs programmability, liquid cooling, and 3nm into one chip. Five hyperscalers are already on board—the sixth probably won’t hold out much longer.