Or, frankly, are we in a bubble?
That's what Bernstein analyst Mark Moerdler asked Microsoft executives on their earnings call this week. Not hedged with qualifiers. Not softened with corporate speak. Just the direct question that apparently needed asking.
When analysts start saying the quiet part out loud on earnings calls, something has shifted.
Two months ago, I wrote about how AI infrastructure spending was following the exact playbook of the fiber optic bubble—revolutionary technology paired with catastrophic overbuilding. I suggested watching earnings calls for the first cracks, the moment when someone blinks and the dominos start falling.
Well, this week delivered something even more telling than a company blinking. It delivered Wall Street finally asking whether any of this makes sense.
The Spending Accelerates Into Madness
The numbers from this week's earnings are genuinely staggering. Alphabet, Meta, and Microsoft together spent $78 billion on capital expenditures in a single quarter. That's up 89% from last year. Most of it destined for data centers and the GPUs to fill them.
Let that sink in for a moment. Three companies. One quarter. $78 billion.
Microsoft alone hit $34.9 billion in the September quarter—a company record. Meta warned that 2026 spending would be "notably larger" than 2025. Google raised its full-year guidance from $85 billion to $93 billion, with a "significant increase" expected next year.
This isn't spending stabilizing. This is spending accelerating.
And here's Microsoft's CFO Amy Hood explaining why:
I thought we were going to catch up. We are not. Demand is increasing.
Let's be clear about what "demand is increasing" actually means. Yes, it means more people are using AI services. It doesn't mean those services are profitable. It doesn't mean the revenue from that demand justifies the capital being deployed.
Azure cloud revenue continues growing—but at roughly the same rate as the prior quarter, even as capital spending hit record levels. Google Cloud revenue beat estimates, sure. But these companies are spending four to five times faster than revenue is growing.
Strip away the corporate optimism and what do you have? Companies spending record amounts to meet demand that isn't generating returns. They're not even building for future growth anymore—they're desperately trying to keep up with today's Ferrari-level computational requirements, losing money on every query while hoping to make it up on volume.
Let's Do Some Kitchen Table Math
Here's the arithmetic that apparently nobody on Wall Street wants to work through.
These companies are on track to spend somewhere north of $400 billion on AI infrastructure in 2025. Now, here's what most people don't understand about AI data centers: they're nothing like Amazon's fulfillment centers or traditional corporate infrastructure.
An Amazon warehouse gets built and used for decades. The racks hold boxes. The conveyor belts move packages. The technology inside might get upgraded, but the fundamental infrastructure has a long useful life.
AI data centers? They're obsolete almost immediately.
Think about what's actually happening inside these facilities. The GPU chips that power AI—which represent about 35% of the total cost—are being superseded by new iterations every year or two. When Nvidia releases a new generation of chips that's dramatically more powerful and efficient, the previous generation doesn't just become less desirable. It becomes uneconomical to operate.
But it's not just the chips. The entire infrastructure is evolving at breakneck speed. Cooling systems designed for current chips can't handle the next generation's heat output. Power systems need to be upgraded. Racking designs change. Even the physical layout of the buildings is being rethought as the technology advances.
This isn't like the transcontinental railroads, where you laid track once and used it for a century. This isn't even like fiber optic cable, which at least still carries data today. AI data centers are depreciating assets from the moment they're completed.
Based on how rapidly this technology is advancing, you're looking at depreciation schedules in the 3-5 year range for much of this infrastructure. Some industry observers think even that's generous—the practical obsolescence might be faster.
Let's be conservative and call it a 5-year depreciation schedule. That's $80 billion in annual depreciation from 2025's spending alone.
Current AI revenue across the industry? Somewhere in the $15-20 billion range.
The depreciation is four times the revenue. And that's before we even talk about the operating costs—the electricity bills that could power medium-sized cities, the armies of engineers required to keep everything running, the ongoing maintenance and upgrades.
To just break even on depreciation with, say, 25% gross margins (optimistic given power costs), you'd need $320 billion in annual revenue. To actually earn a return that justifies this capital—let's say a modest 20% ROIC—you'd need closer to $480 billion in revenue.
From a starting point of $15-20 billion.
And remember, that's just 2025's spending. They're planning to spend even more in 2026. Each year's buildout adds to the revenue requirement, while the previous year's infrastructure is already depreciating into obsolescence.
The math doesn't work. It's really that simple.
When the Hedging Language Starts to Appear
During Meta's earnings call, Mark Zuckerberg floated an interesting hypothetical. If they end up spending too much on infrastructure, they could always sell the computing power to other companies.
"We haven't done that yet," he said. "But obviously, if you got to a point where you overbuilt, you could have that as an option."
Read that again. The CEO of Meta is openly discussing their Plan B for overbuilding—before the building is even finished.
That's not confidence. That's hedging.
Microsoft and Google at least have existing cloud businesses where they can theoretically absorb excess capacity. Meta doesn't. Their entire AI spending spree is predicated on better ad targeting at Instagram and Facebook. If the return on investment doesn't materialize there, they're stuck with billions in depreciating assets and no obvious way to monetize them.
Unless, apparently, they pivot to becoming a cloud provider—a business they've never been in, competing against entrenched players with established customer relationships and pricing power.
This is the language you hear in late-cycle bubbles. The "well, if this doesn't work, we could always do that other thing" reasoning. The backup plans being articulated before the primary plan has had time to either succeed or fail.
The Systemic Risk Nobody's Discussing
Here's what you should also know: this isn't contained to tech anymore.
That $400 billion in infrastructure spending? That's roughly 1.5% of US GDP. Add in the multiplier effects—all the downstream industries supporting this buildout, from electricity generation to construction to chip manufacturing—and you're easily looking at 2% or more of GDP directly tied to the AI infrastructure boom.
Now consider the wealth effect. Roughly half of all consumer spending comes from the top 10% of earners—the people who own appreciating assets. Much of this year's equity appreciation is coming from a handful of AI-related stocks, while the rest of the market treads water.
When those AI stocks were soaring, that wealth effect drove additional consumption. Shopping, travel, luxury goods, real estate—all propped up by investors feeling wealthy from their Nvidia and Microsoft holdings.
What happens when that reverses?
Strip away the AI infrastructure spending and its downstream effects, and what's left of economic growth? Not much. Remove the wealth effect from AI stock appreciation, and consumer spending takes a hit. This isn't just a tech bubble anymore—it's become a meaningful pillar of the entire economy.
The historical parallel isn't really the dot-com crash. It's the railroad panics of the 1800s.
The Railroads That Went Bust (But At Least the Tracks Lasted)
The transcontinental railroads were genuinely transformative. They also went bankrupt repeatedly, often taking the broader economy down with them.
The problem was overbuilding—multiple companies laying parallel tracks, competing on price below operating costs, all funded by investors who assumed demand would grow into the capacity. When it didn't, the failures triggered multi-year global financial panics.
But here's the critical difference: even after the railroad companies went bankrupt, the tracks remained useful. The steel rails laid in the 1880s carried trains well into the 20th century.
AI data centers don't have that luxury. When these facilities become obsolete—and they will, rapidly—there's no equivalent of "well, at least the tracks are still good." The expensive GPUs? The specialized cooling systems? The power infrastructure designed for specific chip architectures? This is stranded capital on a timeline measured in years, not decades.
If AI infrastructure represents 2%+ of GDP and companies are planning to significantly increase spending in 2026, what happens when efficiency improvements make current infrastructure obsolete? When shareholders finally ask why they're earning negative returns on capital deployed into assets that depreciate faster than they can generate returns?
The capital destruction won't be spread over generations like the railroad busts. Be ready for it to be swift and concentrated.
The Music Hasn't Stopped Yet
Let me be clear about timing: I'm not predicting this implodes tomorrow. Long-term power contracts, equipment orders with multi-year lead times, and committed spending create momentum that carries forward regardless of economics.
As long as there's capital willing to fund AI infrastructure, the buildout continues. Take-or-pay contracts with utilities are too expensive to cancel. Equipment orders can't be unwound. The machine keeps running on inertia.
But finance has a way of allowing problems to fester before they explode into the open. The fiber bubble didn't collapse the moment the first efficiency breakthrough happened—it took time for the implications to sink in, for funding to dry up, for companies to exhaust their capital.
What I'm watching for is the change in tone. The shift from "we need to spend more to keep up with demand" to "we're optimizing our capital allocation." From "the returns are coming" to "we have options if we overbuilt."
This week's earnings calls moved the needle. An analyst asked if we're in a bubble. A CEO discussed hypothetical overbuilding scenarios. Microsoft admitted they're not catching up despite record spending.
These are the early warning signs. The cracks forming before the dam breaks.
How I'm Still Playing This (For Now)
My position in the Tortoise AI Infrastructure ETF (TCAI) remains unchanged. The thesis is simple: as long as the money is being spent, someone is receiving it. Better to own the dealers than the addicts.
But I'm watching every earnings call closely now. The language matters. Meta discussing overbuilding scenarios matters. Analysts asking about bubbles matters. Microsoft admitting they're not catching up despite record spending matters.
When the music stops—and it will stop—I want to be near the exit. The beauty of infrastructure providers is that spending cuts get telegraphed quarters in advance. Management teams don't wake up one morning and decide to slash capex by 50%. They signal it, justify it, prepare the market.
I'll be listening for those signals. And when I hear them, I'm out.
Twenty-five years ago, fiber optic companies couldn't keep up with projected demand either. The infrastructure buildout was epic. The vision was intoxicating. Then efficiency improvements made most of it obsolete overnight.
The technology was revolutionary. The applications were transformative. The infrastructure requirements were catastrophically overestimated.
We've seen this story before. The only question is timing.