Compute Grew 25x. Networking Grew 8x. The Fix | Silicon Data, Upscale AI

IGNITE STUDIO Recorded live at NeoCloud Summit 2026, Hotel Kabuki, San Francisco. Steven Hou, Head of Research at Silicon Data, interviews Aravind Srikumar of Upscale AI on the day the company announced Token Fabric. Aravind explains why networking has become the bottleneck for AI: over the past six years compute has scaled roughly 25x while network bandwidth has grown only about 8x, leaving GPUs waiting on the wire. He walks through Token Fabric, Upscale's architecture that unifies scale-up and scale-out into one network, built on purpose-built silicon, open standards, SONiC-based software and the Sky Command orchestration layer. The conversation also covers how Upscale complements NVIDIA, why heterogeneous compute needs open networking, and why being an AI-only networking company matters. Topics include: • Why networking is now the AI bottleneck • Compute grew 25x, networking grew 8x • Introducing Token Fabric • Unifying scale-up and scale-out • Open standards: Ethernet, UALink and SONiC • Complementing NVIDIA and NVLink • Networking for heterogeneous accelerators Explore more conversations from Ignite Studio: https://www.ignitegtm.com/studio.html 📣 Join us at AI INFRA SUMMIT 6 on December 9, 2026, in San Francisco. Register now → https://luma.com/aiinfra6