The telecom supply chain built the dot-com bubble. The AI chip supply chain is building the next one. Here is what leaders need to understand before it pops.
Most people remember the dot-com bubble as a story about bad websites. Pets.com. Webvan. Startups with no revenue and Super Bowl ads. The narrative is tidy. A bunch of companies with silly names raised a bunch of money, burned through it, and collapsed.
That narrative is wrong. Or at least, it is dangerously incomplete.
The dot-com crash was not primarily a story about overvalued software companies. It was a story about supply chains. About physical infrastructure built on projections that never materialized. About hundreds of billions of dollars in fiber optic cable buried in the ground based on a single, intoxicating assumption: that internet traffic would double every 100 days, forever.
It did not.
And now, a quarter century later, we are watching the same pattern play out with AI. Not with glass cables this time. With silicon chips.
Eighty Million Miles of Glass That Nobody Needed
In the five years following the Telecommunications Act of 1996, telecom companies invested more than $500 billion into laying fiber optic cable, building switches, and expanding wireless networks. Most of that money was borrowed. The logic at the time felt airtight: the internet was growing, bandwidth would be the commodity of the century, and whoever controlled the most glass in the ground would win.
Companies like Global Crossing, WorldCom, Qwest, and Level 3 Communications raced to bury fiber across every available right-of-way. Qwest, which had evolved from a railroad construction company, strung cable along rail corridors from Denver to Seattle. Williams Communications threaded it through natural gas pipelines. Even the Bay Area Rapid Transit system signed a $40 million deal to run fiber through its tunnels under the San Francisco Bay.
Wall Street cheered. Equipment makers like Lucent and Nortel sweetened the deal with vendor financing, essentially lending carriers the money to buy more of their own gear. Bond markets poured in billions. Everyone was doing the math, and the math said: infinite demand.
The math was wrong.
By the time the dust settled, companies had laid more than 80 million miles of fiber optic cable across the United States. Even four years after the bubble burst, 85% to 95% of that fiber remained dark. Unused. Silent glass sitting in the dirt. Corning, the world's largest fiber producer, watched its stock fall from nearly $100 to about $1. Ciena's revenue dropped from $1.6 billion to $300 million. Level 3's stock fell 95%.
The technology was real. The internet did change the world. It just did not change the world as fast as the supply chain had been built to serve.
The Pattern Has a Name. It Is Called Collective Hallucination.
Andrew Odlyzko, an emeritus mathematics professor at the University of Minnesota who studies financial bubbles, has a term for what happened. He calls it collective hallucination. That is the moment when investors, executives, the press, and entire industries stop seeing risk because the upside story is too compelling to question.
Here is how the hallucination worked in telecom. It started with a real technology breakthrough. Wavelength Division Multiplexing allowed a single fiber strand to carry multiple data streams instead of one. That was genuine innovation. Capacity exploded. Costs fell. The internet became viable for real commercial use.
But then the projection machine took over. Industry voices predicted that total bandwidth would triple every year for 25 years. WorldCom claimed internet traffic was doubling every 100 days. The real doubling rate was closer to once per year, but that number was not exciting enough to sustain the investment thesis.
So executives kept spending. Banks kept lending. And supply grew 186,000 times faster than demand over seven years.
Read that number again. 186,000 times.
The cautionary signs were there. They were disregarded. As one fiber-era executive later reflected, it was an absolute gold rush. His company was spending about $110 million a week on network buildout. When reality caught up, the stock dropped 95%.
Now Replace Fiber with Silicon and Read the Story Again
In 2026, the four largest hyperscalers—Amazon, Alphabet, Meta, and Microsoft—plan to spend a combined $650 billion to $700 billion on capital expenditures. That is a roughly 67% increase over 2025. The vast majority of that money is going to AI chips, servers, and data center infrastructure.
Amazon alone has committed to $200 billion. Alphabet is projecting $175 billion to $185 billion. Meta is spending up to $135 billion. Microsoft is on pace for $120 billion or more. These are not round numbers pulled from press releases. These are earnings guidance figures, given to investors, backed by contracts already signed with chip manufacturers.
And the supply chain is straining under the weight.
NVIDIA's GPU lead times have stretched to a year. SK Hynix has already sold its entire 2026 high-bandwidth memory output. TSMC's advanced packaging capacity is fully allocated through at least mid-2027. Microsoft, Google, Meta, and Amazon placed multi-billion-dollar forward orders in 2025 that consumed most of NVIDIA's available allocation through the end of 2026 and into 2027.
The bottleneck is not even the GPU die itself anymore. It is the memory. Every major AI chipmaker—NVIDIA, AMD, and Intel—is competing for high-bandwidth memory allocation from the same three suppliers: SK Hynix, Samsung, and Micron. The result is predictable. Less supply for the installed base while demand from new architectures increases.
This is what a supply chain frenzy looks like. We have seen this before.
The $800 Billion Question Nobody Wants to Answer
Bain & Company released a number in their 2025 Global Technology Report that should be tattooed on the forehead of every AI investor. By 2030, AI companies will need $2 trillion in combined annual revenue to fund the computing power required to meet projected demand. Even with aggressive assumptions about cost savings and enterprise adoption, Bain projects that revenue will fall $800 billion short.
Eight hundred billion dollars. That is not a rounding error. That is a structural gap between what is being built and what anyone can pay for.
Meanwhile, Big Tech is increasingly leaning on debt markets to bridge the gap. Alphabet held a $25 billion bond sale and quadrupled its long-term debt in 2025. Amazon is now projected to run negative free cash flow in 2026. According to Bank of America, consensus estimates of AI capex among the five major tech spenders could climb to 94% of operating cash flows in 2025 and 2026, up from 76% in 2024.
In 2025, investment in tech equipment and software reached 4.4% of GDP. That is nearly as high as the peak of the dot-com bubble.
If you are a business leader reading those numbers and thinking this does not affect me because I am not a hyperscaler, think again. When the telecom bubble popped, it did not just take down Global Crossing and WorldCom. It sent shockwaves through the entire technology ecosystem. Venture funding evaporated. IT budgets got slashed. Companies that had nothing to do with fiber optic cable lost access to capital and customers because the broader market cratered.
Supply chain bubbles do not stay contained. They never have.
The Enterprise ROI Gap Is the Tell
Here is where the telecom parallel gets uncomfortable. The fiber glut happened because supply outran demand. The question for AI is whether the same thing is happening, just measured differently.
The numbers are mixed, and that is exactly the problem.
Deloitte's 2026 State of AI in the Enterprise report surveyed more than 3,200 leaders and found that while 66% of organizations report productivity and efficiency gains from AI, revenue growth remains largely aspirational. Only 20% of organizations are actually generating revenue through their AI initiatives. Seventy-four percent are hoping to, but hoping is not the same as doing.
A separate study found that 97% of executives deployed AI agents in the past year, but only 29% are seeing significant ROI. BCG's research identifies just 5% of enterprises achieving substantial value from AI. And 42% of companies abandoned most of their AI projects in 2025.
Let me restate that. Nearly half of companies abandoned most of their AI projects last year.
This is not a sign that AI does not work. AI works. I see it work every day in the businesses I consult with. But there is a massive gap between AI delivering real value at the operational level and AI justifying $700 billion in annual infrastructure spending. The fiber optic cables worked too. The technology was sound. The economics were not.
What the Telecom Bubble Got Right (Eventually)
Here is the part of the story that gets lost in the wreckage. All of that dark fiber? It eventually got used. The oversupply of bandwidth crashed prices by 90%, and that cheap capacity became the foundation for YouTube, Netflix, cloud computing, and the entire modern internet economy.
YouTube launched in 2005, built on top of bandwidth that was essentially free because of the dot-com overbuild. Google bought it a year later for $1.65 billion. In 2023, YouTube generated $31 billion in revenue for Alphabet.
The companies that built the infrastructure mostly went bankrupt. The companies that built on top of the cheap infrastructure that survived became the most valuable businesses in history.
That is the pattern. The overbuild destroys the builders. The overbuild creates the platform. Someone else captures the value.
If this pattern holds for AI, and I believe it will, the excess GPU capacity and data center infrastructure being built today will eventually make running large language models cheap enough to enable applications we cannot imagine yet. The question is who survives the correction, and who gets buried under the infrastructure they built.
What This Means If You Are Not Sitting in a Big Tech Boardroom
If you are a business leader running an organization under 100 people, this might feel abstract. You are not placing billion-dollar chip orders. You are not building data centers. You are trying to figure out whether to invest in AI for your team, and how much, and which tools, and whether the whole thing is going to look different in 18 months.
It is going to look different in 18 months. That is actually the point.
Here is what the supply chain dynamics mean for you right now:
AI costs are likely to stay elevated in the short term, then fall dramatically. The chip shortage, the memory bottleneck, the packaging constraints. These are real, and they are keeping the cost of AI infrastructure high. But supply always reacts to demand. New fabs are being built. Alternative accelerators are gaining traction. NVIDIA has pushed its product release cadence from every two years to every year. When the supply catches up—and it will—prices will drop. Potentially fast.
Open source is your insurance policy. While the supply chain frenzy plays out at the infrastructure level, open-source AI models continue to advance in parallel with proprietary ones. Meta's Llama, Mistral, and other open-source models are giving organizations real options that do not depend on a single vendor's pricing or capacity constraints. The leaders who will weather this best are the ones building on flexible tech stacks, not locking themselves into a single provider during a supply crunch.
Now is the time to build capability, not just buy tools. The smartest move in a supply chain bubble is not to bet on the infrastructure. It is to build the organizational muscle to use whatever infrastructure becomes available at whatever price point emerges. Train your team. Build your workflows. Understand what AI can and cannot do for your specific operations. When costs come down—and they will—the organizations that already know how to use AI will move fastest.
Do not wait for the bubble to pop. Do not rush because of the hype. Both reactions are wrong. The telecom bubble taught us that the technology was real but the timeline was not. Leaders who pulled back entirely missed the internet. Leaders who overextended got destroyed. The right answer was measured, intentional adoption with an eye on the long game. That is still the right answer.
The Same Movie, Different Screen
We have been here before. A transformative technology arrives. The supply chain races to build the infrastructure. Capital floods in based on projections that assume infinite, exponential demand. Supply outruns reality. The correction is painful. And then, years later, the technology fulfills its promise anyway, just on a different timeline and for different beneficiaries than the ones who built the rails.
AI will transform how we work, make decisions, and build businesses. I believe that deeply. I see it every day in the organizations I work with.
But $700 billion in annual infrastructure spending, an $800 billion revenue gap, 85% to 95% of fiber still dark four years after the last time we did this. These are not numbers you ignore because the technology is exciting.
The internet was exciting too. The fiber was still dark.
