The a16z Show · 2026-11-11 · 56:05
Presentadores: Raghu Raghuram · Guido Appenzeller · Mike Volpi
Former Cisco M&A chief Mike Volpi — who helped build the internet in the late 1990s and now runs Hanabi Capital — sits down with a16z's Raghu Raghuram and Guido Appenzeller to compare the two great infrastructure buildouts. His core argument: the fiber laid in the 1990s stayed useful for decades even after its builders went bankrupt, and AI compute will likely follow the same pattern — fully utilized, but not necessarily by today's owners. Along the way he covers why H100s are outliving their five-year depreciation, what NVIDIA should learn from Cisco's acquisition machine, why only NVIDIA and Google can currently ship a rack as a single system, and why AI data centers may force networking back toward circuit-like architectures.
Volpi opens with the Twain line — history doesn't repeat, but it rhymes. In the 1990s, wavelength-division multiplexing let one fiber carry 40 to 80 times the bandwidth it used to, while compute improves only at something like Moore's-law pace. The offsetting difference: AI's practical applicability today is higher than bandwidth's was then — Anthropic, OpenAI and open-model providers like Fireworks, Together and Modal are growing far faster than the 1990s infrastructure businesses did.
Volpi is emphatic: today's AI buildout will be fully utilized, 'if not more, 100%'. Whether today's owners keep the assets is a different question — it depends on macro and geopolitical forces outside the builders' control. He counts roughly 200 neoclouds, many with weak balance sheets, and expects an economic correction that reshuffles ownership while the underlying assets stay useful for a long time. He also notes the sheer scale: AI infrastructure is now absorbing around 2% of GDP, versus a small fraction of a percent for the internet buildout.
Four years ago H100s were assumed to have a five-year useful life and rented for about $1.80–$2 per GPU-hour; by that math they should be nearly defunct now. Instead they rent for $2.50–$3 — AI's own software progress keeps older chips useful longer, which makes the assets more financeable. Volpi's key metric is token revenue per watt: Blackwell and Vera Rubin extract more dollars from each watt than Hopper, so H100s only stop making sense once Vera Rubin supply is sufficient. Until then, scarcity — rooted in semiconductor manufacturing, which he says won't change materially for two to three years, maybe five — keeps even older generations valuable.
Cisco's M&A worked because it did enough deals for statistics to kick in: roughly a third worked very well, a third were mediocre, a third failed — and the team wasn't afraid to fail, so it caught outliers like the companies of Andy Bechtolsheim and Mario Mazzola. By 2000, about 50% of Cisco's revenue came from acquired sources, across roughly 70 deals, with almost no single deal costing more than 4–5% of market cap. His advice to NVIDIA: Mellanox was genius; he scratches his head at Hugging Face — but with the market granting NVIDIA enormous optionality, this is the moment to take risks. He is more skeptical of NVIDIA financing neoclouds: an industry built on scale wants to consolidate to three players, like airlines, no matter how many small ones you fund.
Internet networking was built for random, bursty traffic — packet switching with statistical multiplexing. AI data centers look nothing like that: links are consistently fully utilized and traffic is far more predictable, especially in training. Volpi expects a new data-center network architecture built on predictability and full utilization — scheduled or circuit-like switching, possibly optical — which could make switches much cheaper by shedding buffers and lookup complexity. It's a classic innovator's dilemma: incumbents like Broadcom have no incentive to disrupt their own packet-switching products, so the change may have to come from GPU vendors shipping networking as part of the system — today only NVIDIA and Google can ship a rack as a single system.
“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.”
— Mike Volpi · 0:00
Escuchar el episodio original ↗
Esta columna es un resumen y comentario. Todos los derechos pertenecen al programa original.