The artificial intelligence infrastructure boom of 2024-2026 bears striking resemblance to the telecommunications bubble that consumed $2 trillion in capital between 1998 and 2002, leaving 95% of installed fibre optic capacity unused and triggering spectacular corporate collapses.

The Parallel Investment Mania

Just as the Telecommunications Act of 1996 unleashed a wave of competitive network builders, today's AI revolution has spawned dozens of companies racing to build data centres, GPU clusters, and specialised computing infrastructure. The scale is breathtaking: industry estimates suggest over $500 billion in AI infrastructure investment globally since 2023, with hyperscale data centre construction accelerating at unprecedented rates.

The echoes are unmistakable. In 1999, telecom executives justified massive fibre deployments with claims that internet traffic doubled every 100 days — a statistic later debunked by University of Minnesota researchers who found actual growth was merely doubling annually. Today's AI infrastructure builders make similarly bold projections about compute demand, often citing exponential growth curves that assume universal AI adoption across all industries simultaneously.

The Overcapacity Warning Signs

The telecom bubble's fatal flaw was building for projected demand rather than actual usage. Companies like Global Crossing spent $12 billion on networks that carried a fraction of anticipated traffic before filing for bankruptcy. Level 3 Communications and Williams Communications built redundant coast-to-coast networks, creating an overcapacity factor of 20 times actual demand.

Current AI infrastructure exhibits similar patterns. Multiple cloud providers are simultaneously building massive GPU clusters, often in the same geographic regions. The assumption that every enterprise will immediately adopt AI workloads requiring enormous compute resources mirrors the telecom bubble's assumption that every business would instantly embrace high-bandwidth internet applications.

The Valuation Disconnect

Perhaps most concerning is the valuation methodology. During the telecom bubble, companies were valued on projected network capacity and potential customer bases rather than actual revenue or profitability. WorldCom's $11 billion accounting fraud exemplified how financial engineering masked fundamental unprofitability.

Today's AI infrastructure companies often trade on similar metrics: total GPU count, theoretical compute capacity, and projected AI workload growth. Many have yet to demonstrate sustainable unit economics or clear paths to profitability, yet command valuations that assume rapid monetisation of massive capital investments.

The Inevitable Correction

The telecom bubble's collapse came not from technological failure — the internet did indeed transform commerce — but from massive overinvestment relative to adoption timelines. The infrastructure built during the mania eventually proved valuable, but not for the companies that built it or the investors who funded it.

Similarly, AI will likely transform industries over the coming decade. However, the current infrastructure build-out may be creating capacity that won't be fully utilised for years, if ever. The companies surviving this cycle will likely be those with strong balance sheets and realistic deployment timelines, not those racing to build the most capacity fastest.

Read the full Telecom Bubble and Fibre Optic Mania (1998-2002) entry in the Bubblepedia Encyclopedia

The Capital Allocation Warning Signs

The most alarming parallel between today's AI infrastructure boom and the telecom bubble lies in the systematic misallocation of capital driven by competitive paranoia rather than demand fundamentals. During the telecom frenzy, companies like Global Crossing and WorldCom raised over $20 billion each, building redundant networks based on the fear of being left behind rather than rigorous demand analysis. The result was catastrophic: by 2002, installed fibre capacity utilization rates averaged just 2-3%.

Today's data centre builders exhibit identical behavioral patterns. Nvidia's H100 GPU clusters, costing $10-20 million each, are being deployed at rates that far exceed current enterprise AI adoption. Microsoft alone committed $50 billion to AI infrastructure in 2024, while Amazon Web Services announced plans for 100 new data centres globally. Yet enterprise surveys consistently show that only 15-20% of companies have moved beyond pilot AI projects to production deployments.

The warning signs are crystallizing in real-time metrics. Data centre vacancy rates in key markets like Northern Virginia have dropped to historic lows of 1.5%, driving speculative construction reminiscent of the telecom bubble's build-first-ask-questions-later mentality. Power grid constraints are forcing developers to secure capacity years in advance, creating artificial scarcity that inflates asset values. Most tellingly, the gap between announced AI infrastructure spending ($500+ billion) and actual enterprise AI software revenue ($50 billion in 2024) mirrors the telecom era's disconnect between network investment and revenue generation that ultimately triggered the sector's collapse.