Will the AI infrastructure boom end in stranded datacenters? No, former Cisco executive Mike Volpi told a16z, the capacity will get used but the current owners may not survive to run it.
Volpi ran much of Cisco’s 1990s acquisition spree and he said history rhymes here. Both eras imagined big applications before the network was ready, both drew a flood of capital, and both built ahead of demand. The tech differs in one key way. Wavelength division multiplexing, a technique that puts many colors of light on one fiber, multiplied bandwidth 40 to 80 times at once, Volpi said, while compute still tracks closer to Moore’s Law. Demand proof is stronger this time, with OpenAI, Anthropic, Fireworks and Together growing faster than 1990s bandwidth sellers did, Volpi said.
Financing is cleaner now but the asset is trickier. In the 1990s vendors funded much of the build on their own balance sheets, Volpi said, while today’s build is driven by private debt and public equity plus some Nvidia vendor style funding. Fiber in the ground stays useful for 20 to 30 years, Volpi said, while a GPU’s life was assumed to be about five years. That assumption broke. H100s priced around $1.82 per GPU hour four years ago now rent for $2.50 to $3, Volpi said, because scarcity keeps older chips working and AI software keeps them relevant.
The risk sits with neoclouds, which are cloud providers built specifically to rent GPU capacity to others. Volpi said he heard there are 200 of them and most have weak balance sheets. That warning already has teeth. CoreWeave shares were down 17% over five days, Nebius down 15% and IREN down nearly 10% as of Friday morning after the election, per Fierce Network. For startup teams renting by the hour, that means budget for volatile pricing and have a second source. For neocloud operators, it means debt that assumed full utilization can turn fast.
Volpi’s advice to Nvidia draws straight from Cisco. His team did about 70 deals, kept each under 4 to 5% of market cap, paid almost all in equity, and by 2000 about half of Cisco’s roughly $30 billion in revenue came from acquired sources, Volpi said. Open protocols made that work because a bought switch talked to Cisco routers on day one. He called Mellanox genius, said he does not see the logic for Hugging Face, and said it does not matter because the market has given Nvidia optionality to take misses. He called Nvidia’s funding of many small clouds noble but futile, since scale businesses consolidate to three hyperscalers or three airlines no matter how many startups get funded.
Watch token revenue per watt, which is the dollars a datacenter squeezes from each unit of power. Blackwell beats Hopper and Vera Rubin will beat Blackwell on that math, Volpi said, so H100s only survive while semiconductors stay scarce. That scarcity in chip manufacturing shows no quick fix, Volpi said, so buyers should track power contracts and chip supply before they track model releases.
