The $670 Billion AI Infrastructure Bet: Moon Landing or Money Pit?

Everyone is calling it the new Space Race. I’m calling it the new dot-com bubble—just with better PR and bigger price tags.

Meta, Microsoft, Amazon, and Alphabet are planning to spend $670 billion on AI infrastructure in 2026. For context, that’s more than the Apollo program that put humans on the moon. It’s more than the entire U.S. interstate highway system. As a percentage of GDP, it rivals the railroad expansion of the 1850s.

The tech press is breathless. Investors are excited. Everyone’s positioning this as the most important capital investment in human history.

But here’s what nobody wants to say out loud: Most of this money is going to be lit on fire.

The Conventional Wisdom

The story we’re being told goes like this:

AI is the future. Data centers are the infrastructure of that future. Just like railroads enabled the industrial revolution and highways enabled suburban prosperity, GPU clusters and exascale compute will enable the AI revolution.

The companies spending this money aren’t reckless—they’re visionary. They learned from past mistakes. They’re not building castles in the sky; they’re building the foundation for AGI, autonomous agents, and a radically transformed economy.

Missing this wave would be like missing the internet in 1995. The only mistake you can make is not spending enough.

Why That’s Wrong

Let me be clear: AI is real. The technology works. GPT-5, Claude Opus 4.6, and their successors are genuinely impressive.

But capability doesn’t equal economics.

Here’s the uncomfortable truth: nobody has figured out how to make money on AI at the scale these investments require. OpenAI is burning cash. Anthropic is burning cash. Even Microsoft’s Copilot—the posterchild for “AI that actually makes money”—is reportedly struggling with unit economics.

The math is brutal:

• Training costs: $100M+ per frontier model
• Inference costs: Orders of magnitude higher than search
• Energy requirements: A single AI data center can consume as much power as a mid-sized city
• Revenue per query: A fraction of what Google makes on search

Meanwhile, we’re building data centers like there’s no tomorrow. Meta just announced plans for a facility that will require its own dedicated power plant. Microsoft is restarting Three Mile Island.

This isn’t infrastructure. This is a game of chicken.

What’s Actually Happening

The $670 billion isn’t about building the future. It’s about not being left behind.

Every major tech company is locked in a prisoner’s dilemma:

• If everyone spends and AI takes off, you need to have spent to stay relevant
• If everyone spends and AI fizzles, you all lose together (but at least you don’t lose alone)
• If you don’t spend and AI takes off, you’re dead

So they spend. And they keep spending. Because the alternative—being the only one who didn’t—is existentially terrifying.

This is rational behavior for individual companies. But it’s irrational for the industry as a whole.

We’ve seen this movie before. In the late 1990s, telecom companies spent hundreds of billions laying fiber optic cable, convinced the internet would create infinite demand for bandwidth. They were right about the internet. They were catastrophically wrong about the economics. Most of them went bankrupt.

The fiber remained useful—but at a fraction of the predicted value.

Why This Matters

If I’m right, here’s what happens next:

Phase 1 (2026-2027): Peak spending. Data centers go up everywhere. AI capabilities improve rapidly. Media coverage is wall-to-wall positive.

Phase 2 (2027-2028): Revenue reality hits. Companies realize that while AI is impressive, monetization is hard. Some use cases work (coding, creative tooling). Most don’t scale economically.

Phase 3 (2028-2029): The reckoning. Spending slows. Projects get “reevaluated.” Some companies write off billions in stranded assets. Stock prices crater.

Phase 4 (2030+): The infrastructure remains—like those fiber cables in the 2000s. It gets used. AI becomes genuinely useful. But at valuations and expectations far below what justified the build-out.

Who gets hurt?

• Tech shareholders (when write-downs come)
• Energy grids (massive capacity built for AI that doesn’t materialize)
• Climate goals (enormous carbon emissions for uncertain returns)
• Workers (when the bubble pops, layoffs follow)

Who benefits?

• Nvidia and hardware manufacturers (they get paid upfront)
• Power companies (long-term contracts locked in)
• Whoever figures out real AI business models (they inherit cut-rate infrastructure)

What You Should Do

If you’re an investor: Be very, very careful with AI infrastructure stocks. The companies building this stuff are making enormous bets that may not pay off for a decade—if ever. Look for AI application companies that are actually profitable, not just burning VC money on compute.

If you’re a builder: This is actually good news for you. The oversupply means cheap compute is coming. Build products that assume abundant, affordable AI—because that’s the world we’ll end up in after the correction.

If you’re a user: Enjoy it while it lasts. Right now, AI companies are subsidizing your usage to build market share. That won’t last forever, but it’ll last long enough for you to figure out what’s actually valuable.

If you’re a skeptic: You’re not crazy. The emperor has some very expensive new clothes, and they might actually be useful—but they’re probably not worth what we paid for them.

The Counterargument

Let me steel-man the bull case, because I could absolutely be wrong:

Counterpoint: AI adoption might follow a steeper S-curve than previous technologies. While current revenue models are weak, that could change overnight if:

• AI agents become truly autonomous (replacing entire job functions)
• Enterprise adoption accelerates faster than expected
• New business models emerge that we haven’t conceived yet

The companies spending this money have access to internal metrics we don’t. Maybe they’re seeing usage patterns and engagement levels that justify these bets. Maybe GPT-6 or Claude Opus 5 will be the iPhone moment—the product that makes all this infrastructure obviously necessary in retrospect.

And unlike the dot-com era, these companies are profitable. Meta, Microsoft, Amazon, and Alphabet collectively print money. They can afford to make mistakes.

My response: All true. And if the iPhone moment happens, I’ll happily eat crow. But the iPhone worked because it solved real problems at a sustainable price point. AI is solving problems—but at what cost? Until someone shows me the unit economics that justify these data centers, I remain skeptical.

Final Thoughts

I’m not an AI doomer. I think the technology is transformative. I use it every day.

But transformative technology doesn’t always justify transformative spending—especially not at the scale and speed we’re seeing now.

The $670 billion isn’t about building the future. It’s about companies terrified of missing it, spending money they can afford to lose on infrastructure they’re not sure they’ll need, hoping that the competitive dynamics will somehow resolve into a profitable equilibrium.

Maybe they’re right. Maybe this time really is different.

But when four companies are spending more money than it took to put a man on the moon, split the atom, and wire a continent—all in a single year—on a technology that’s barely three years old?

That’s not vision. That’s panic.

Disagree? Good. Tell me why in the comments, or on Twitter. I’d love to be wrong about this.