The numbers keep climbing. Every quarter brings another announcement of a data centre campus that costs more than the GDP of a small nation. The hyperscalers are spending like there is no tomorrow, and the market is cheering them on.
But the maths does not work. The energy alone for these facilities is staggering. Cooling, networking, staffing, land. The operating costs hit before a single model is trained. And the revenue side is thin. Most AI products are sold at a loss or given away free to build usage. Someone has to pay the electricity bill eventually.
The comparison to past bubbles is uncomfortable but valid. The dot-com era had real internet companies, just overvalued. The housing bubble had real houses, just overpriced. AI infrastructure has real chips and real buildings, but the demand curve is speculative. Enterprises are piloting, not deploying at scale. Consumers use chatbots for fun, not for revenue.
I am not saying the technology is worthless. Far from it. But the spending is priced for perfection. Every company is building as if they will capture the entire market, and that is mathematically impossible.
The correction will come. It always does. The question is whether it is a gentle pullback or a violent crash. My guess is violent, because the leverage is enormous. Debt-funded construction, supply contracts signed years in advance, and a stock market that has priced in infinite growth.
When the music stops, the companies with real cash flow will survive. The ones borrowing to build monuments will not. Watch the balance sheets, not the press releases.
Paul