The Hyperscaler AI Bubble Risks

The risks are in financing expectations, utilization rates, and arms-race logic, not in a mismatch between capex and revenue

There might be a bubble in AI, but not in the way that many people seem to think. The common complaint is this: the hyperscalers are spending hundreds of billions of dollars on building out data centers, and revenue directly attributable to these expenses is only around $50 billion or so. Here’s The Economist making this claim:

By our reckoning, the toal revenue from the tech accruing to the West’s leading AI firms is currently $50bn a year. Although such revenues are growing fast, they are still less than 2% of the $2.9 trillion investment in new data centers globally that Morgan Stanley…forecasts between 2025 and 2028.

The problem with this kind of argument is that it treats AI capex as if it should show up one-for-one in an “AI revenue” line. This is a category error. Let’s review.

1. Revenue labeling does not equal value creation. Most AI value is embedded: higher ARPU, lower churn, better conversion, better ad yield, more engagement, premium features bundled into existing SKUs. That flows into existing revenue lines (Search, Ads, Commerce, Prime, Office, Cloud), not a clean AI line item.

2. Four payoff channels, only one is top-line. AI capex monetizes via:

   1. new revenue,

   2. margin expansion (unit cost declines),

   3. defensive retention (avoided revenue loss), and

   4. real options on future products.

   A 1:1 capex—>revenue heuristic ignores 3 of the 4 payoff channels.

3. Timing mismatch (build now, harvest later). Capacity is built ahead of demand. Training clusters and inference fleets create a J-curve: high depreciation + low utilization initally; utilization ramps as products, models, and distribution catch up. Compare current “AI revenue” to multi-year capex is duration mis-matched.

4. Attribution is hard to impossible. If AI improves ranking, fraud detection, latency, or personalization, the P&L never shows “AI”. It shows higher gross margin or lower losses. Anyone looking for a tidy line item is measuring the wrong thing.

5. Defensive capex can be NPV-positive with zero new dollars. If your rivals add AI copilot features, standing still loses market share. “Revenue not lost” won’t show as incremental “AI revenue,” but it is still an economic return on capex.

6. Cost of capital leverage. At $10-20B per campus, shaving 50-100 bps off WACC via credible forward hedges (power, compute, offtake) can exceed early AI revenue. Financing alpha is not operating alpha, but both are returns to capex.

7. Ecosystem externalities. Early returns often accrue upstream (chips, power, interconnect), with downstream capture lagging until standards, tooling, and distribution harden. Misreading that phasing as “no ROI” is a classic platform-cycle mistake.