At the risk of sounding a bit Cassandra-esque, I return this week with my cautionary tale of the potential for poor returns on capital in the coming years from the current AI cycle. You may remember Cassandra, a figure from Greek mythology who was cursed to speak the truth and be believed by no one. The capital-cycle warning works the same way. It is accurate, it is old, and it is dull, so it gets waved off right up until it’s too late.
This week’s Chart shows that on only rare occasions over the last couple of decades did GE (pre-2024 breakup), General Electric, a company J.P. Morgan and Thomas Edison built in 1892, earn its cost of capital.
GE historically had one edge: GE could borrow cheaper than anyone, so its customers borrowed cheaply to buy GE machines. The AAA rating held for over 50 years. Then GE Capital grew into the fifth-largest lender in the country, and the strength quietly became the fragility. The empire cheap money built, cheap money eventually broke.
Here’s the takeaway for me. The technology was real at every step. Electricity, jet engines, medical imaging, all genuine, all world-changing. What sank the returns was the money invested. Too much capital, invested and then mis-invested, competing away the very profits that drew it in.
We have seen this movie before. Railroads in the 1870s laid tens of thousands of miles of track and handed shareholders years of losses. Telecom in the late 1990s poured more than $500 billion into fiber, roughly 80 million miles of it, and left something like 85% of it dark and unlit. Bandwidth got cheap, customers won, and the firms that funded the buildout went bankrupt. The internet was real. The payoff to the people who wired it was not.
Which brings us to the current investment cycle. Big tech spent close to $400 billion on AI infrastructure in 2025. Bain estimates data centers will need something like $2 trillion in annual revenue by 2030 to justify the spend. Current AI revenue sits near $20 billion. That gap is no rounding error, to say the least. Meta, Microsoft and Alphabet are each spending more of their revenue on capex than AT&T did at the peak of the telecom bubble.
The trap is that none of them can stop. Ease off while a rival keeps building, and you risk obsolescence. So everyone builds, capacity floods in, and the group competes away the returns it is racing to capture.
The more recent and concerning wrinkle is the addition of significant amounts of debt to fuel the AI investment boom. It was one thing for the big spenders to crash their free cash flow margins. But now, they’ve run way past that with the addition of debt, some direct, some indirect, and they’re doing so by increasingly creative means. This makes me want to bring out my worry beads.
AI may change the world. Whether this much capital, moving this fast, earns a return in excess of its cost is a very different question.
Sources: The Collapse of GE’s House of Debt Was 130 Years Coming (Bloomberg, David Fickling, 2021); Surviving the AI Capex Boom (Sparkline Capital); Parallels Between the Hyperscalers and the Telecom Firms of the 1990s (MOI Global).