What the Internet Boom Teaches Us About the AI Buildout | Mike Volpi

Mike Volpi opens by saying history does not repeat, but it rhymes. In the late 1990s, Cisco and the broader industry built internet and web infrastructure around expectations that pictures, text, audio, video, commerce, and social networking would become transformative applications. Connectivity and bandwidth demand surged alongside large inflows of equity and debt capital. Volpi’s central comparison is that the earlier boom built communication, whereas today’s boom is building intelligence—and intelligence may ultimately have greater economic value.

Volpi explains that fiber capacity expanded unusually quickly because wavelength-division multiplexing allowed many colors of light to travel through the same fiber. A single fiber that once carried one or 10 gigabits could eventually carry 40 or 80 times as much. When internet businesses later disappointed investors, companies with heavy debt failed, but the fiber they had installed remained useful and was operated by new owners. He expects a similar outcome for AI infrastructure: a correction may change ownership without making the underlying compute useless.

Compared with the 1990s bandwidth build-out, Volpi sees much stronger practical applicability in AI. Anthropic, OpenAI, and providers such as Fireworks, Together, and Modal have shown revenue growth rates far above those of the earlier infrastructure businesses. He predicts an economic correction because roughly 200 neoclouds exist and many have weak balance sheets, but he still believes the build-out will be fully utilized, or more than fully utilized. Whether today’s owners retain the assets will depend on macroeconomic and geopolitical forces.

The speakers contrast fiber, which can remain useful for 20 or 30 years, with GPUs whose useful life was initially assumed to be about five years. Yet H100s, once rented for roughly $1.80 to $2 per GPU-hour, are still operating and reportedly renting for $2.50 to $3. Volpi says the key metric is token revenue per watt: newer Blackwell and Vera Rubin systems should extract more dollars from each watt, but extreme scarcity means older GPUs remain valuable. If supply ever becomes abundant, their economics could deteriorate quickly.

Raghu Raghuram notes that AI infrastructure investment is much larger—around 2% of GDP—than the internet investment of the late 1990s. Volpi says today’s build-out is more driven by capital markets, including private debt and public equity, whereas more 1990s financing came from equipment vendors’ balance sheets. NVIDIA’s arrangements resemble vendor financing even if they are not called that. The earlier boom had the advantage that fiber went into the ground as a durable asset; AI financing depends more heavily on how long installed compute remains economically useful.

Asked what NVIDIA can learn from Cisco, Volpi describes a high-frequency, portfolio-based M&A strategy. Cisco made enough acquisitions for statistically meaningful outcomes: roughly one-third worked very well, one-third were mediocre, and one-third failed. The company was willing to fail because outliers such as Andy Bechtolsheim’s and Mario Mazzola’s companies produced venture-like returns; during Volpi’s tenure, about 50% of Cisco’s revenue came from acquisitions, with most deals costing no more than 4% or 5% of market capitalization. He says NVIDIA has similar optionality today.

Volpi praises NVIDIA’s acquisition of Mellanox but is less certain about the rationale for some other deals, such as Hugging Face. The hosts debate NVIDIA’s financing of neoclouds: it gives startups access to scarce capacity and may help shape the industry, but Volpi argues that scale economics will eventually consolidate the market, much as U.S. airlines consolidated. NVIDIA may want 50 customers instead of three hyperscalers, but funding weaker companies cannot permanently overcome an industry structure that favors scale. The strategy can still produce returns if NVIDIA exits at the right time.

Volpi says Cisco created a small deal team and a larger integration project-management group after discovering that companies are good at repeatable processes but bad at one-time events. Integration managers worked with colleagues across business and functional units, giving each acquisition a defined process. Cisco’s open networking protocols also helped: an acquired switch could immediately communicate with Cisco equipment, sometimes allowing the company to change the badge and use its existing sales force. Later, larger acquisitions such as Scientific Atlanta became difficult, especially when Cisco attempted partial rather than full integration.

For a company deciding whether to sell, Volpi identifies three cases: a strong business that should not sell, a struggling business where a sale may help everyone, and a “tweener” requiring judgment. The founder’s energy and belief in reaching the next phase matter more than anything else; the board must also ask whether the company can become 10 times larger, not merely double. Volpi advises founders to maintain relationships with NVIDIA, Amazon, OpenAI, competitors, and other potential buyers. Regular conversations are useful intelligence gathering, not necessarily “shopping” the company.

Volpi describes the current infrastructure landscape as constrained mainly by semiconductor supply, with fabrication capacity unlikely to change materially for two or three years, perhaps five. He expects model routers and more economical models to make token demand price-sensitive. Unlike traditional networking, AI systems often keep GPUs continuously connected and heavily utilized, so the best architectures treat the rack—or even the data center—as one system. He favors more chip diversity because model designs differ, and highlights 3D memory and hybrid copper bonding for reasoning workloads with large KV caches.

Traditional networks were designed for stochastic traffic and statistical multiplexing, but Volpi says AI workloads are more predictable: training links are consistently busy, and inference can become more predictable through batching. He expects future data-center networks to use architectures based on scheduled or circuit-like traffic rather than heavily buffered packet switching. That could make switches cheaper by reducing buffers and lookup complexity, but incumbent vendors have little incentive to disrupt their existing products. The change may require GPU vendors to ship networking as part of the full system.

Volpi says this is the most intellectually demanding cycle of his career because investors must reason across the entire system, from EUV equipment and semiconductor packaging to networking, models, and applications. He joined Hanabi Capital after retiring in 2023 because the opportunity was too compelling for the yacht-and-golf plan to survive. He believes the potential end prize is larger than the internet boom: one company could become worth a trillion dollars and another $500 billion. The hosts agree that this cycle surpasses 1999 in dynamism and importance.