AI Revenue Is Growing. The Math Still Doesn't Work.

The AI industry is generating real revenue now. That’s the problem.

For three years, bulls could dismiss the bears by pointing at growth curves and saying “give it time.” That move is getting harder. When The Economist publishes a piece titled “AI revenues are growing fast, but not fast enough,” the argument has shifted terrain. It’s no longer about whether the revenue is real. It’s about whether the revenue can ever be real enough to justify what’s already been spent — and what’s being committed right now.

The numbers are brutal in their simplicity. Microsoft, Google, Amazon, and Meta have collectively authorized hundreds of billions in AI capex over the next 18 months. OpenAI is burning cash at a rate that would make a 2021 SoftBank portfolio company blush. Anthropic raised its latest round at a valuation that bakes in years of aggressive growth before any plausible path to cash-flow breakeven. The revenue exists, and it’s growing fast. But it’s growing fast off a base that was close to zero, and the denominator — total capital deployed — is compounding at a different, scarier rate.

This is the classic overhang problem. The market doesn’t care that revenue doubled year-over-year if the infrastructure spend tripled. What you get is a widening gap dressed up as a momentum story. The AI sell-off in chipmakers this week — South Korean markets tumbling, Wall Street tech shares set for further declines in their wake — isn’t panic. It’s arithmetic. Investors are doing the same calculation that anyone with a spreadsheet can do: at current GPU unit economics, current utilization rates, and current average contract values, the cash-on-cash return profile looks like enterprise software circa 2001. Not a zero. Just priced wrong.

The steelman for the bulls is real and worth taking seriously: we are in the infrastructure build phase, and infrastructure always looks like waste before it looks like value. The railroads of the 1870s destroyed their investors and built the American economy. AWS lost money for years. The capex is building a platform on which applications we can’t yet imagine will generate returns we can’t yet model. Fine. Acknowledge all of that. The counterpoint isn’t that AI is a fraud. The counterpoint is that the application layer is not developing as fast as the infrastructure layer, and the gap between them is where the valuation risk lives.

Look at what’s actually getting bought. Enterprises are signing pilots, not transformational contracts. The $50-per-seat Copilot licenses are real but they’re running into the oldest problem in enterprise software: adoption. You can sell a seat; you can’t sell behavior change. The stories of genuine productivity transformation are still anecdotal, still confined to specific functions — coding, legal document review, customer support deflection. They’re not showing up yet in aggregate labor cost data at the companies buying the most AI. If they were, you’d see it in margins. You mostly don’t.

None of this means the technology fails. It means the timeline is being repriced. There’s a difference between “AI transforms everything in 5 years” and “AI transforms everything in 12 years,” and the current capex commitments were priced for the former. Repricing toward the latter doesn’t require a narrative collapse. It just requires a long, grinding period where revenue growth is real but not that fast — and where the companies carrying the most infrastructure debt either get acquired, consolidate, or find ways to charge customers a lot more than customers currently want to pay.

The companies most at risk aren’t the hyperscalers. Microsoft and Google can absorb a slow decade; they have cash engines that don’t depend on AI working on any particular timeline. The risk sits with the pure-play frontier labs and the mid-tier infrastructure players who bet their entire existence on the fast scenario. OpenAI needs the application layer to explode. Anthropic needs enterprise AI spend to grow faster than enterprise AI skepticism. If neither happens in the next 18 months, the funding math gets existential rather than merely uncomfortable.

The AI sell-off will get called a buying opportunity by someone on CNBC by end of week. It might even be right. But the underlying tension doesn’t resolve with a rally. The gap between infrastructure spend and application revenue has to close by one of two mechanisms: revenue accelerates dramatically, or capex commitments get pulled back. The first is what everyone is betting on. The second is what a market repricing forces. Watch which one moves first.

Infrastructure built too fast isn’t wasted — but someone always pays for the timing mismatch, and it’s rarely the people who built it.