Let's travel back to 1999, when the internet was going to change everything (it did) and fiber optic cables were digital gold (they weren't). Companies like Global Crossing and WorldCom looked at early internet traffic growth and did what seemed like perfectly rational math: more people online equals more data flowing equals more physical infrastructure needed. Simple.
These companies proceeded to lay fiber optic cables across every continent with the religious fervor of 19th-century railroad barons. The vision was intoxicating—infinite bandwidth for an infinitely connected world. Stock prices soared. Investors threw money at anything with fiber or optical in the name. The infrastructure buildout was truly epic.
Then Cisco—bless their engineering hearts—came along and casually announced they'd figured out how to compress vastly more information through existing fiber networks using dramatically less physical infrastructure. It's the same pattern we've seen with mobile networks—the jump from 3G to 4G to 5G didn't require more frequency spectrum, just better technology to squeeze exponentially more data through the same bandwidth.
Suddenly, all that expensively laid cable became what economists politely call stranded assets and what everyone else called worthless junk buried in the ground.
The collapse was swift and merciless. Billions in capital expenditures vanished. Companies that were supposed to own the internet's backbone became cautionary tales. The same infrastructure that was "insufficient" for projected demand became massive overcapacity overnight.
But here's the twist that should give pause to anyone paying attention to today's AI spending spree: the internet did change everything. The technology was revolutionary. The applications were transformative. The only problem was the infrastructure requirements were grossly, catastrophically overestimated.
The AI Industrial Complex: History's Most Expensive Remix
Fast forward to 2025, and we're witnessing what can only be described as the fiber optic bubble's bigger, more expensive, electricity-guzzling cousin. Meta is burning through $40 billion annually on AI infrastructure—that's more than the GDP of most countries spent on building digital brains. The AI infrastructure buildout makes the dot-com era's cable-laying look quaint by comparison. Data centers are sprouting across landscapes like technological mushrooms after rain, each one consuming enough power to run a medium-sized city. The intensive buildout to connect computers into supercomputers has reached levels that would make the most enthusiastic fiber-layer from 1999 blush.
And just like their fiber-laying predecessors, today's AI infrastructure builders are making the same fundamental error: assuming that current inefficiencies represent permanent requirements rather than temporary limitations waiting to be solved.
Every AI query today gets the Ferrari treatment. Your simple question about the weather triggers the same massive computational resources as complex scientific modeling. It's like using a Formula 1 race car to deliver pizza—technically impressive, wildly effective, and completely unnecessary.
The DeepSeek Moment: When Efficiency Crashes the Party
In the fiber bubble, Cisco's efficiency breakthrough was the moment reality crashed into fantasy. Today's equivalent might be companies like DeepSeek (and the others following in its footsteps) showing early glimpses of how to achieve dramatically better AI performance with far less computational infrastructure.
These efficiency improvements aren't theoretical—they're happening now. AI systems are getting dramatically better at matching computational power to task complexity. The Ferrari-to-Toyota transition isn't a question of if, but when. And when it happens, the current overbuilt infrastructure will reveal itself as the expensive monument to miscalculation that it is.
We're building computational capacity based on current inefficiencies, then acting surprised when those inefficiencies get solved. It's like building highways assuming cars will forever get two miles per gallon.
When Smart Companies Get Blindsided by Tomorrow
Let's acknowledge something crucial: I could be wrong about the timing, magnitude, or even the fundamental premise of this analysis. But that's precisely why erring on the side of skepticism makes investment sense.
History is littered with brilliant companies blindsided by innovations they never saw coming. BlackBerry had logical reasons why physical keyboards were superior—better typing accuracy, professional appeal, proven experience. They never saw the iPhone revolution coming. More recently, Meta poured billions into the metaverse with serious strategic analysis and top talent. Where are those metaverse discussions now?
Warren Buffett's principle applies here:
It's better to be roughly right than precisely wrong.
Even smart companies with good data regularly get caught off guard by technological shifts they didn't anticipate.
The $40 Billion Trigger
Unsustainable AI spending represents what market watchers call a "trigger point"—the moment when financial reality meets technological fantasy. Meta spending $40 billion annually on AI infrastructure while borrowing money at 7% interest rates is the kind of math that works right up until it doesn't.
When Meta (or another major player) blinks first and announces reduced AI infrastructure spending, the domino effect will be swift. Negative comparisons will ripple through earnings calls. Investors will suddenly ask uncomfortable questions about return on investment. The phrase efficiency improvements will go from engineering curiosity to existential threat.
The Economic House of Cards
Here's where the AI infrastructure bubble becomes particularly concerning: unlike the isolated fiber overbuilding of the dot-com era, today's AI spending boom is propping up entire economic sectors. Everything connected with AI infrastructure is experiencing explosive growth, but the rest of the economy appears increasingly soft underneath the surface.
The AI infrastructure buildout isn't just creating a technology bubble—it's become the pillar supporting broader economic growth. Electricity demand, data center construction, chip manufacturing, cooling systems—entire supply chains depend on the assumption that current AI infrastructure requirements will continue growing exponentially.
When efficiency improvements make current infrastructure investments obsolete, the correction won't be contained to a few overextended tech companies. It will ripple through every sector that has built business models around the AI infrastructure boom. The result could be a significant economic downturn that exceeds the scope of the dot-com crash.
Preparing for the Inevitable
Smart money doesn't ask whether an AI infrastructure correction will happen—it considers when and how severe. The fiber bubble offers some uncomfortable guidance.
Timing: Bubbles can continue longer than rational analysis suggests, especially when easy money policies fuel continued investment. The current expectation of Federal Reserve rate cuts could extend the AI infrastructure boom beyond its natural lifespan.
Magnitude: When efficiency improvements become widely apparent, the correction will be notable. Companies that have built entire business models around current AI inefficiencies will face challenging questions about their value propositions.
Sector Rotation: Healthcare, energy, and other sectors that have underperformed during the AI boom may offer attractive opportunities when investors flee AI infrastructure plays for more traditional value propositions.
The Infrastructure vs. Application Divide: Just as internet applications proved transformative despite fiber overbuilding, AI applications will likely create enormous value. The question is whether current infrastructure investments match actual long-term requirements.
How I Am Playing This
Given everything I have written above, why would I want to own any AI-related stocks? Simple. As long as the money is being spent, money is also being made. So I might as well participate while the sun is shining. Or in other words:
When all the tech companies are getting high on AI, it's good to own shares in the dealer.
My investment approach is straightforward: instead of betting on the companies that burnt through $155 billion this year on AI infrastructure, I'm betting on the companies receiving that money. It's the classic picks and shovels strategy.
I've allocated a starter 2.5% position in my portfolio to the Tortoise AI Infrastructure ETF (Ticker: TCAI), with plans to grow this to 5%. While TCAI just launched in August 2025, it captures exactly what I'm looking for: a diversified play across all three categories of AI infrastructure spending—48% technology components, 36% energy infrastructure, and 16% data centers. One ETF, one trade, comprehensive exposure to the entire supply chain.
The positioning makes sense even if I'm wrong about bubble timing. These infrastructure providers have diversified customer bases across multiple tech giants, contracted cash flows, and recurring revenue models. Even if Meta cuts spending, Amazon might increase it. Even if Google optimizes its data center buildout, Microsoft still needs power infrastructure. The dealers win regardless of which specific companies are buying.
But I'm not planning to hold indefinitely. I'll be monitoring every major tech company's earnings calls, watching for the first signs of AI infrastructure spending moderation. Code words like "optimizing capital allocation" or "rightsizing our infrastructure investments" will be my signal to cut the position. The beauty is that spending cuts get telegraphed quarters in advance during earnings calls, giving me time to exit before the infrastructure providers feel the impact.
Again, this isn't a buy-and-hold forever position—it's a tactical bet on the sustainability of current AI infrastructure spending levels, with a clear exit strategy when the music stops.
25 years later, are the tech companies smart enough to learn from $40 billion mistakes before they become $400 billion disasters? Based on the current trajectory, the answer appears to be probably not.
