Building Real-World Infrastructure for AI: Google, Cisco & a16z
In this a16z Show panel, Google's Amin Vahdat, Cisco's Jeetu Patel, and a16z's Raghu Raghuram discuss the unprecedented scale of the current AI infrastructure buildout, which they describe as comparable to the internet buildout, the space race, and the Manhattan Project combined — and roughly 100x larger in pace than the late-1990s internet cycle. They argue the industry is grossly underestimating demand, constrained primarily by power, land, permitting, and supply chain bottlenecks that will persist for 3-5 years. A major theme is the shift from general-purpose to specialized silicon. Vahdat notes Google's TPUs are 10-100x more efficient per watt than CPUs for certain workloads, and that even older generations remain at 100% utilization. Both speakers see a 'golden age of specialization,' with inference workloads (prefill vs. decode) driving differentiated architectures. Networking is being reinvented across three axes — scale up, scale out, and scale across (data centers up to 800-900km apart acting as one logical facility) — with Cisco positioning itself as an alternative to a Broadcom-dominated ecosystem. The geopolitical dimension looms large: panel members observe that China's reliance on 7nm chips combined with abundant power and engineering talent may push them toward radically optimized, power-efficient architectures, potentially diverging from the U.S. path of pushing toward 2nm. On inference economics, both note that 10-100x cost reductions keep getting absorbed by users demanding longer reasoning chains (from 20-minute sessions to 30-hour autonomous coding runs), making intelligence-per-dollar a never-ending optimization target. They advise founders against building thin model wrappers and instead recommend tight model-product integration with feedback loops.
Original sources
Podcast The a16z Show
Episode:Building the Real-World Infrastructure for AI, with Google, Cisco & a16z ↗
YouTube The a16z Show
Video:Building the Real-World Infrastructure for AI, with Google, Cisco & a16z ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.