AI Infrastructure Boom Mirrors Dot-Com Fiber Glut, But With Key Differences
TREE NEWS reports: Former Cisco senior vice president Mike Volpi, who led roughly 70 acquisitions during the networking giant’s rise to the world’s most valuable company, compared the current AI infrastructure buildout to the late-1990s internet buildout in a recent a16z podcast. Volpi, joined by a16z partners Raghu Raghuram and Guido Appenzeller, argued that while AI infrastructure assets will ultimately be useful, they may not remain in the hands of today’s owners—much like the fiber optic cables that were laid during the dot-com boom but later changed hands at bargain prices after debt-laden operators went bankrupt.
The discussion comes as CoreWeave reported better-than-expected third-quarter results, sparking a 12% single-day rally in the NeoCloud sector and offering temporary relief from investor anxiety over AI returns. Volpi’s remarks carry weight given his experience: Cisco’s stock plummeted from $80 to $9 in a single quarter after the telecom bubble burst, erasing 90% of its market value.
Fiber’s Fate as a Template for GPUs
Volpi sees a parallel between fiber and GPUs: both are essential infrastructure assets that retain value even when their original owners fail. However, he highlighted a crucial difference—GPU asset lifespans may be longer than initially expected. “Fiber buried in the ground can last 20 to 30 years,” he noted, while computing assets are typically assumed to depreciate faster due to Moore’s Law. Yet the H100 GPU, expected to have a five-year lifespan with rental rates of $1.8-$2 per GPU-hour four years ago, is still in high demand and now commands $2.5-$3 per hour. Volpi attributes this to an unintended consequence: AI software capabilities are extending the useful life of underlying hardware, making it easier to finance.
Appenzeller countered that current high H100 rents reflect overall supply scarcity, not the chip’s competitive edge. Once next-generation architectures like Vera Rubin scale, older chips will lose economic viability as token revenue per watt improves dramatically. Volpi acknowledged this but argued that scarcity will persist because fab capacity—particularly ASML lithography equipment output—is a hard constraint that won’t change for at least two to three years, possibly five.
Bigger, More Resilient Buildout
The scale of AI infrastructure investment is roughly an order of magnitude larger than the internet buildout, representing about 2% of US GDP versus a fraction of a percent in the 1990s. Volpi reframed the difference: “At 100,000 feet, what are you building? Back then it was communications. Today, we’re building intelligence. The economic value of intelligence is likely higher than that of communications.”
He also pointed to faster application-layer revenue growth. Companies like Anthropic, OpenAI, and inference clouds such as Fireworks, Together, and Modal are growing far faster than their dot-com predecessors, even if margins are thinner. Additionally, the demand side is broader: “Seven billion people have a supercomputer in their pocket,” and software has penetrated nearly every industry, providing a wider consumer base for AI than the internet had in the 1990s.
Still, Volpi remains cautious. “I’m more optimistic about this cycle overall, but I also believe there will be corrections and industry consolidation.” He singled out the fragile balance sheets of NeoCloud providers—”Last time I spoke with Nvidia, there were 200 NeoClouds, and many have ugly balance sheets.” He predicts a significant number will fail in an economic downturn, but their underlying assets will be acquired and put to use, mirroring the fiber story.
Living in Jensen’s World
Volpi described the current supply-demand dynamic as “living in Jensen’s world,” where AI compute demand is nearly infinite and supply is severely constrained. He offered advice for Nvidia, which he said is in a moment of “maximum optionality” and should take risks. He praised the Mellanox acquisition as “genius” and said the rumored Hugging Face deal is not a big deal. However, he was skeptical of Nvidia’s strategy to finance NeoClouds to expand its customer base, calling it “futile” because the industry naturally concentrates around a few hyperscalers. “You can fund 100 US airlines, but you’ll still end up with three,” he said.
Network Architecture Ripe for Disruption
Volpi also predicted a fundamental shift in data center networking. Current switches are designed for random, uneven traffic patterns requiring statistical multiplexing. But AI data centers have highly predictable, continuously saturated traffic—especially in training—making circuit switching or time-division multiplexing more suitable. “In a relatively static world, you don’t need as much intelligence and buffering. Just connect and send,” he said. This is bad news for incumbents like Broadcom and Cisco, as simpler architectures command lower prices. He sees Google’s data center network, including optical circuit switching, as a pioneer, and believes the revolution may have to be driven by “rack-scale” players like Nvidia or Google, since networking is increasingly integral to system architecture.
Market Implications
Volpi’s analysis suggests that while the AI infrastructure theme remains powerful, investors should distinguish between the durability of the assets and the durability of the companies that own them. The potential for GPU assets to outlive expectations could support valuations for lessors and cloud providers, but the coming NeoCloud shakeout poses risks. Nvidia’s dominance may continue, but its attempts to broaden its customer base could face structural limits. The networking sector, particularly incumbent switch makers, faces long-term disruption risk as AI workloads demand simpler, more deterministic architectures.
Key Takeaways for Investors
- Asset longevity matters: GPU useful lives may exceed initial estimates, supporting cash flows for infrastructure owners, but technological obsolescence remains a threat when next-gen chips scale.
- Ownership risk: History suggests infrastructure assets often change hands after a boom-bust cycle; today’s NeoCloud leaders may not be tomorrow’s owners.
- Scale and demand are larger: AI infrastructure investment is an order of magnitude bigger than the dot-com buildout, with faster application revenue growth and broader adoption, potentially cushioning the downturn.
- Consolidation ahead: The NeoCloud sector is overcrowded; expect failures and M&A, with assets recycled to stronger players.
- Network disruption: Incumbent switch vendors face innovation dilemma as AI data centers shift toward circuit-switched architectures; watch for moves by Nvidia and Google.
