Strategist Tom Lee is advising investors to buy AI infrastructure stocks on dips, drawing a direct parallel to Cisco Systems' sustained growth during the dot-com infrastructure buildout of the late 1990s.
Lee's thesis rests on a simple premise: companies providing foundational infrastructure for significant technologies—whether networking equipment then or computational power now—benefit from multi-decade demand cycles regardless of cyclical volatility. Under this framework, current pullbacks in AI-adjacent equities represent entry points rather than structural deterioration.
The Cisco precedent is instructive. The networking-equipment giant captured dominant market share as enterprises built out internet connectivity. Cisco's revenue grew from $1.3 billion in 1995 to $12.2 billion by 2000, and the company sustained profitability and pricing power for nearly two decades after the dot-com crash.
But the analogy breaks on several material points. First, Cisco operated in a low-competition environment with high switching costs; customers who standardized on Cisco's routers and switches faced significant replatforming expenses. Current AI infrastructure—whether chips, cloud compute, or software frameworks—faces accelerating commoditization and shorter technology cycles. A dominant training chip today can face displacement within 18-24 months as competitors iterate. Cisco's networking equipment had 5-10 year competitive windows.
Second, Cisco benefited from pricing power. As demand grew, customers had limited alternatives and could not easily negotiate down costs. Today's AI infrastructure providers face downward pricing pressure from hyperscalers building proprietary chips and from competition driving gross margins lower. Nvidia's gross margins, while elevated, are beginning to compress as AMD and Intel scale alternatives.
Third, the regulatory and geopolitical backdrop is different. Cisco faced no export restrictions, no domestic industrial policy favoring competitors, and no threat of forced technology transfer. Current AI infrastructure providers operate under these constraints, which could fragment markets and compress addressable opportunity.
Lee's framework—that essential infrastructure compounds value over decades—is not wrong in principle. But the burden is on investors to identify which current AI players will retain Cisco-like structural advantages. The speed of technological obsolescence in semiconductors and AI algorithms suggests the answer is: very few, and only those with defensible moats in specific bottleneck inputs.
