Dot-com boom déjà vu: What the nineties Internet predictions got right, wrong, and why AI timelines are the real risk

Summary

The dot-com boom offers a partial template for understanding AI hype cycles. The Internet was genuinely transformative, but predictions failed on timeline rather than direction. Retail didn't vanish by 2009 as forecasted. Instead it declined unevenly over two decades. Walmart remains a trillion-dollar company. Nike is worth $90 billion. Shopping malls struggle but persist, particularly in Los Angeles where developers like Rick Caruso found new formats. Nearly every dot-com claim contained a directional truth, but regulatory, market-driven, and social friction slowed everything down.

Predictions assumed civilizational phase changes equivalent to the printing press or electricity would compress into five years rather than fifty. Voluntary and involuntary brakes applied. Media companies sued file-sharing platforms. Financial markets retreated. The Battle of Seattle in 1999 mobilized 40,000 people over four days against World Trade Organization talks, raising political salience around labor displacement and corporate consolidation. That momentum didn't halt Internet adoption but created political permission structures that shaped policy, including tariffs and domestic procurement rules, over the following decade.

AI sentiment differs sharply. More than 30% of Americans believe AI could end human life on Earth, a vastly higher proportion than those concerned about Y2K or 2012 apocalypse narratives. Yet actual adoption remains shallow. An NBER survey of American firms found only 20% reporting productivity or employment benefits from AI, while 80% reported no measurable impact. The gap reflects real barriers such as HIPAA compliance, IT department blockers, and sector-specific constraints, but also points to a widespread implementation problem.

Protests are already materializing. New Jersey canceled a planned 25,000-square-foot data center after community pushback, a fraction of Meta's typical 500,000-square-foot campus footprint. The facility was likely intended for content delivery rather than frontier AI training, but it will function as a political datapoint for policymakers considering AI rollout friction.

The core risk is whether unmanaged negative externalities harden into political constraints before adoption scales. Energy demand was foreseeable. Hyperscalers could have subsidized electrical buildouts years ago. The next inflection points likely arrive in 2027 or 2028. Forecasting specific problems and mitigation pathways before they metastasize may prevent the friction that slowed the Internet from becoming roadblocks for AI.