Flying Too Close to the Sun
Big Tech is spending $900 billion on AI infrastructure this year. That’s 3.6% of U.S. GDP, a buildout bigger than the railroads, the interstate highways, or early telecom.
And the cracks are showing:
Alphabet lost roughly $500 billion in market value in a single session after its CapEx and negative free cash flow spooked investors.
Amazon’s trailing free cash flow sits at -$42.4 billion.
Meta posted its first negative quarterly free cash flow in over a decade.
CoreWeave is paying 15-16% to borrow around $50 billion.
The industry isn’t just running hot. It’s running on borrowed money.
Intel: The Slow-Motion Unraveling
Once the king of silicon, Intel is now a case study in how fast an empire can shrink.
Rolling layoffs: from about 132,000 employees in 2022 toward a core target of 75,000, with more cuts hitting its data center and AI group in 2026
Vanishing perks: the cuts came with shrinking benefits and a steady drip of voluntary separations
Foundry bleeding: a $2.09 billion operating loss in one quarter, with outside customers contributing just $293 million in revenue
Stock down 54% from its highs, with single-day drops as deep as 28%
The bailout: about $36 billion in fresh capital, including a 10% stake handed to the U.S. government
When a company that once defined the industry needs the federal government as a shareholder, the warning is hard to miss.
Nvidia: The High-Wire Act
Nvidia is the opposite story, and arguably the scarier one. It posted $96.22 billion in quarterly revenue, with margins near 75%.
But its customers are buying those chips with borrowed money. Hyperscalers and GPU rental startups are leveraged to the hilt. If they trim spending or default, Nvidia’s orders drop with them. Even a 3-5% pullback erases tens of billions in value in a day.
And the damage spills over. AI chips are starving the supply of memory for everything else, so 2026 PC shipments are projected to fall 10.4% while prices rise 17%.
The Money Merry-Go-Round
Here’s how the machine works:
Hyperscalers invest in AI labs, often through compute credits.
The labs spend those credits right back on the hyperscalers’ clouds.
That shows up as cloud revenue growth.
The growth justifies more debt to build more data centers.
To keep the debt off their books, companies use special purpose vehicles. Meta’s $30 billion Hyperion project carries only 20% on Meta’s balance sheet. The other $24 billion sits in an outside entity funded by private credit.
The Power Wall
Money isn’t the only limit. The grid is.
Data centers could go from 3-4% of U.S. electricity to 9-17% by 2030
AI racks that once drew 10-15 kW now push 100+ kW
Local protests and moratoriums are spreading as household power bills climb
Utility rates are becoming an election issue, and voters want data center developers to foot the bill.
The Human Cost
Behind every balance sheet is a workforce bracing for the next email.
Endless reductions: layoffs, buyouts, and hiring freezes have become routine
Squeezed budgets: non-AI teams are being cut to fund chip purchases
Shaky pay: with RSUs tied to volatile stock prices, compensation has become a moving target
Constant dread: the question has shifted from “if” to “when”
The 2027 Trap
Here’s the part executives don’t want to say out loud:
The debt behind these buildouts runs 7 to 15 years.
The chips inside them go economically stale in 12 to 18 months.
The cost of producing AI output is falling about 70% a year, so demand has to more than triple annually just to keep revenue flat.
The leases and loans from the 2023-24 rush start coming due in 2027. Operators will need to replace hundreds of billions of dollars in aging GPUs that haven’t paid for themselves, just as lenders are pulling back. Weaker players face write-downs and defaults, and the chipmakers selling to them feel it next.
