TLDR
Former Cisco executive Mike Volpe compares the current AI infrastructure build-out to the 1990s internet fiber boom, arguing that while an economic correction may reshuffle ownership, the underlying GPU and data center assets will remain valuable for decades. He shares lessons from Cisco's high-volume M&A strategy, discusses NVIDIA's dominant position, and predicts a fundamental shift in networking architecture from statistical multiplexing to predictable, circuit-like systems for AI workloads. Volpe believes the AI infrastructure opportunity could ultimately exceed the internet boom in scale and value.
For you
PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.
Sign in to unlockSurprising Insights
GPU asset life is extending longer than expected: H100s rented at higher rates years later because AI-driven software improvements increase hardware efficiency, making older chips more financeable.
Networking may revert to circuit-switching (optical/TDM) because AI workloads are fully utilized and predictable, making traditional statistical multiplexing obsolete — a return to telecom's roots.
NVIDIA financing NeoClouds to diversify its customer base may be futile because scale economics naturally consolidate the industry to a few hyperscalers regardless of funding.
The "minimum unit" of AI infrastructure has shifted from a GPU card to a rack, and potentially to an entire data center row, as systems become tightly integrated.
Key takeaways
AI infrastructure build-out parallels the 1990s fiber boom but with higher practical applicability and faster revenue growth from model providers like OpenAI and Anthropic.
Underlying assets (GPUs, data centers) will remain useful long-term even if ownership changes due to macroeconomic corrections, similar to fiber optic cables after the dot-com bust.
NVIDIA dominates the current scarce GPU supply chain ("Jensen's world"), with fab capacity as the key bottleneck unlikely to ease for 3-5 years.
Networking architecture must evolve from statistical multiplexing (designed for variable internet traffic) to deterministic, high-utilization circuit-like systems (optical/TDM) for predictable AI workloads.
Cisco's M&A success came from high acquisition volume, tolerance for failure (1/3 hits, 1/3 misses, 1/3 failures), and repeatable integration processes — not from avoiding risk.
Startups should proactively build relationships with potential acquirers and competitors early; CEOs should maintain regular dialogue with key industry players.
Chip diversity will increase as model architectures diversify (e.g., 3D DRAM, hybrid copper bonding, optical interconnects) to match specific workload needs like KV-cache-intensive reasoning.
The current era is uniquely intellectually demanding because it requires systems-level thinking across physics, semiconductors, networking, and model architecture to identify bottlenecks and opportunities.
Notable quotes
“We live in a world where demand for AI capacity tokens is near infinite and supply is incredibly constrained. So another way to put it is we all just live in Jensen's world.”
▸ Notable quotes · 0:00“Do I think the build-out that's happening now is useful and will be fully utilized, if not more? 100%. Will it be owned by the same people that own it today? A lot of interesting dependencies.”
▸ Notable quotes · 6:59“The retro on the M&A process that I think worked well for Cisco was we were not afraid to fail. And we did it with high enough frequency that we got a few outliers. You bought this thing for $500 million and, like, it's doing a billion in earnings two years later.”
▸ Notable quotes · 14:53“The fundamental tenet of the workload has changed. Yet, if you look at the dominant networking architecture today... it's a stat mux packet switching architecture. So I think that at some point for data center networking, we're going to have a new kind of architecture that assumes a high degree of predictability and a full utilization.”
▸ Notable quotes · 48:21“I actually think this is the most fascinating time to be doing our job and to be in our industry... because it requires more intellectual exertion to understand what next. You have to think at a systems level.”
▸ Notable quotes · 53:58Chapters
1:06 — Introduction: AI Build-out vs. Internet Boom: Hosts introduce Mike Volpe, former Cisco M&A lead during the 1990s internet build-out, to compare that era with today's AI infrastructure expansion.
3:02 — Parallels and Differences: 1990s Fiber vs. AI Compute: Volpe outlines similarities (surge in demand, capital influx) and key differences: compute improves via Moore's Law (slower than WDM fiber), AI has higher practical applicability and revenue growth, and assets will remain useful even if ownership changes.
10:00 — Capital Structure and Asset Longevity: Discussion of capital sources (vendor financing vs. capital markets), GPU asset life extending beyond expectations due to software efficiency gains, and the critical metric of token revenue per watt.
15:06 — Cisco's M&A Playbook and Lessons for NVIDIA: Volpe details Cisco's high-volume, equity-based acquisition strategy (70 deals, 50% revenue from acquisitions), tolerance for failure, repeatable integration processes, and advice for NVIDIA's current M&A approach.
46:10 — Networking Architecture Shift for AI Workloads: Analysis of why traditional statistical multiplexing (Ethernet/IP) is mismatched for predictable, fully-utilized AI traffic, and the coming shift to deterministic optical/TDM circuit-like architectures pioneered by Google.
38:03 — Chip Diversity, Future Outlook, and Intellectual Challenge: Exploration of emerging chip architectures (3D DRAM, optical interconnects) matched to model diversity, Volpe's view that this era tops the 1990s in dynamism and intellectual demand, and the massive prize for solving systemic bottlenecks.
Speakers
Mike Volpe — General Partner, Hanabi Capital; former Cisco SVP Corporate Development
Raghu Raghuram — General Partner, Andreessen Horowitz (a16z)
Guido Appenzeller — General Partner, Andreessen Horowitz (a16z)