David Bombal #549: The Apple of Networking? Cisco’s 100T Full Stack (Connects 128,000 GPUs)
Feb 21, 2026
Martin Lund, chip architect behind Cisco’s Silicon One, builds ultra‑fast networking silicon for massive AI clusters. He explains the 102.4Tbps G300, on‑chip intelligent agents that reroute traffic in real time, and how to link upwards of 128,000 GPUs without stalls. He also covers the 1.6T linear pluggable optics and why Ethernet is winning in large AI data centers.
AI Snips
Chapters
Transcript
Episode notes
Market: Multiple Players, Few Shipments
- Other vendors like NVIDIA and Broadcom are announcing similar class solutions, but shipping is limited today.
- Cisco emphasizes system integration (silicon, hardware, software) as its differentiator.
Design For Programmability And Lifecycle
- Build systems, not only chips, to extend device lifecycle through programmability.
- Use programmable engines to adapt to new AI model requirements without hardware replacement.
Intelligent Collective Networking Boosts Performance
- Cisco's