Show and guest
An a16z podcast conversation featuring a16z partner Mike Volpi (former Cisco executive who led M&A in the 90s); hosted by a16z partners Raghu Raghuram and Guido Appenzeller. Original English title: Mike Volpi on Why AI Infrastructure Will Outlast the Hype Cycle.
Key takeaways
- Today’s AI infrastructure is “100% useful, and will be fully utilized—even overutilized,” but it may not belong to today’s owners; the balance sheets of the 200 new clouds Nvidia talks about are unhealthy, and an economic adjustment and industry shakeout are expected 6:48 7:15
- Compared with fiber in the 90s: WDM multiplied the capacity of a single fiber by 40–80 times; computing power is constrained by Moore’s law, so improvements are more modest. But AI has greater practical applicability than bandwidth did back then, and Anthropic and OpenAI are growing revenue far faster than companies did then 4:20
- In terms of scale, AI infrastructure investment today is about 2% of GDP, whereas internet infrastructure investment then was only a few tenths of a percentage point of GDP. Back then, what was being built was “communications”; today, it is “intelligence,” which has greater economic value 7:51
- GPUs are lasting longer than expected: H100 was initially expected to have a 5-year lifespan and rent for 1.82 dollars per hour; four years later, the rental price had instead risen to 2.5–3 dollars. The key metric is how much token revenue can be generated per watt. The scarcity lies in semiconductor manufacturing itself, which is unlikely to change in the next 2–3 years; whoever owns wafer-fab capacity is the kingmaker 11:23 34:21
- Lessons from Cisco M&A: during his tenure, the company made about 70 acquisitions—roughly a third were major successes, a third were mediocre, and a third failed. No single deal exceeded 4–5% of the company’s market capitalization. In 2000, about 50% of revenue came from acquisitions. His advice to Nvidia: Mellanox was a brilliant acquisition; he doesn’t understand Hugging Face, but it is worth a try 15:18 18:09
- The lesson from Cisco’s decline: customers overordered, then orders fell to a quarter of their previous level. Quarterly results missed by 35–40% (Cisco had not missed in a decade), and its share price fell from more than 80 dollars to 9 dollars, dropping by about nine-tenths over six months. AI data-center networking will also need to change: traffic is predictable and links remain fully loaded, so statistical multiplexing and packet switching are no longer a good fit. The future will move toward circuit-like/optical switching and cheaper architectures 45:41 46:48
Background and notes
An a16z podcast conversation with a16z partner Mike Volpi, drawing on his firsthand experience at Cisco in the 90s to compare that era with today’s AI infrastructure. Subtitle coverage note: the video subtitles contain 1740 timestamped segments, covering 0:00–55:53 (total length 56:06), covering essentially the entire video. The views and timestamps in this article are based on the original subtitles; each timestamp marks the start of the corresponding spoken segment. Publication-time note: the publication time was approximately 00:00 on 2026-11-12 Hong Kong time, converted from the relative time displayed by YouTube when the content was captured, with hour-level precision.