Is AI capex a bubble?

The honest answer is that the bull and bear cases both rest on real evidence. Demand signals are genuine and revenue is growing, and it is equally true that a large share of the spending is debt-funded and depreciated on assumptions nobody outside the companies can check.

How long does an AI GPU actually last?

Nobody outside the hyperscalers knows precisely. Companies depreciate them over roughly five to six years, while utilization patterns and failure data suggest heavily used training fleets may wear out faster. The gap between those two matters directly to reported earnings.

What is the real constraint on AI data centers now?

Electricity and grid interconnection. Chip supply eased and memory became the pinch point, but the binding constraint on new large sites is now how fast a utility can deliver power, which is measured in years rather than quarters.

How do I get AI exposure without buying chip stocks?

The adjacent layers are power generation, grid equipment, data center REITs and industrial real estate. They offer less upside than a leading accelerator maker and considerably less customer-concentration risk, because electricity has more than four buyers.

InvestingThe big one

The AI Capex Trade

Hundreds of billions are being spent on chips, buildings and power on the bet that demand shows up. Here is where the money goes, who books it, and what breaks if the bet is wrong.

15 articles · about 164 min in total

Start with Is This an AI Bubble? A Computer Engineer Reads the Financials

This is the largest capital deployment in the history of the technology industry, and it rests on an assumption nobody can verify yet: that the revenue shows up before the depreciation does.

The way I think about it, there are three separate bets stacked on each other. That AI demand keeps growing. That the companies spending the money capture that demand rather than passing it to customers. And that the assets they are buying last as long as the accounting says they do.

That third one is the quiet part. Hyperscalers depreciate AI accelerators over roughly five or six years. If the real useful life of a hard-run training GPU is meaningfully shorter, then a material slice of reported big-tech profit is an accounting assumption. Nobody outside those companies can settle it, which is precisely why it is worth tracking.

The path is ordered to follow the money physically. Chips first, then the customers buying them, then the balance sheets funding it, then the memory and power constraints, and finally the buildings and land it all sits on. The bottleneck moved from silicon to electricity, and the trade moved with it.

Key takeaways

GPU depreciation schedules are an accounting assumption rather than a measured fact, and a shorter real useful life would mean current hyperscaler earnings overstate profitability.
The AI capex bottleneck has moved from chip supply to high-bandwidth memory and then to electrical generation and grid interconnection.
Nvidia derives a large share of revenue from a small number of hyperscaler customers, which makes it a concentrated business rather than a diversified one.
Debt-funded AI capex transfers risk from the equity holder to the bondholder and makes the spending harder to cut quickly if demand disappoints.

Frequently asked questions

Is AI capex a bubble?: The honest answer is that the bull and bear cases both rest on real evidence. Demand signals are genuine and revenue is growing, and it is equally true that a large share of the spending is debt-funded and depreciated on assumptions nobody outside the companies can check.
How long does an AI GPU actually last?: Nobody outside the hyperscalers knows precisely. Companies depreciate them over roughly five to six years, while utilization patterns and failure data suggest heavily used training fleets may wear out faster. The gap between those two matters directly to reported earnings.
What is the real constraint on AI data centers now?: Electricity and grid interconnection. Chip supply eased and memory became the pinch point, but the binding constraint on new large sites is now how fast a utility can deliver power, which is measured in years rather than quarters.
How do I get AI exposure without buying chip stocks?: The adjacent layers are power generation, grid equipment, data center REITs and industrial real estate. They offer less upside than a leading accelerator maker and considerably less customer-concentration risk, because electricity has more than four buyers.

All reading paths

The AI Capex Trade

Key takeaways

Step 1: Is This an AI Bubble? A Computer Engineer Reads the Financials

Step 2: AI Chip Stocks in 2026: The Bull Case and the Bear Case Are Both Right

Step 3: Nvidia Customer Concentration: The Chart That Should Keep You Up at Night

Step 4: How Long Does an AI GPU Actually Last? The Quiet Number Holding Up Big Tech Earnings

Step 5: The Earnings Call Tells You Which Hyperscaler Is Faking the CapEx ROI

Step 6: Amazon Is Borrowing to Build AI. That Changes the Whole Trade.

Step 7: SK Hynix Just Raised $26.5 Billion. The Thing It Sells Is the Real AI Bottleneck.

Step 8: The AI CapEx Trade Is Changing: From Picks-and-Shovels to Power

Step 9: Nuclear and SMR Stocks in an AI-Powered Grid

Step 10: Data Center REITs: The Real Estate Play Hiding Inside the AI Boom

Step 11: Data Centers Are the New Industrial Real Estate Asset Class

Step 12: Semiconductor Stocks and the AI Wave: Separating Signal from Hype

Step 13: Apple Is the Most Valuable Company Again, and It Spends 6% of What Alphabet Spends on AI

Step 14: Meta's Iris Chip Isn't an Nvidia Killer. It's a Cost Cut.

Step 15: The AI Infrastructure Roadmap for 2025

Frequently asked questions