A year ago, AI was a story about models. Today it's a line item in the stock market's earnings. Roughly half of the S&P 500's earnings growth now traces back to AI-related activity (Goldman Sachs). Read that again: the broad market's growth story is the AI story. Everything else is a rounding error.
But here's what I think most commentary gets wrong: the market isn't pricing AI's intelligence. It's pricing its construction program.
Look at where the money is actually going. Data-centre capital expenditure was up 92% year-on-year in the second quarter (Dell'Oro Group). Hyperscalers keep raising capex guidance. Utilities are dusting off load-growth forecasts they haven't used in twenty years. When you buy the AI trade today, you're not buying today's cash flows — you're underwriting a buildout.
The pricing happens in two places, and they tell different stories. The builders — chips, cloud — are priced as if earnings compound indefinitely. The enablers — power producers, grid equipment, construction — are priced as if the buildout never stops. Both can be right, but both depend on the same hinge: somebody has to pay for the output. Every dollar of capex is a bet that enterprises convert AI spend into margin, and that consumers keep paying for it indirectly.
That's the "who pays" question, and it's the only valuation question that matters now. If enterprise ROI arrives — real productivity, real revenue — today's multiples will look cheap in hindsight. If it doesn't, we've built a very expensive bridge to nowhere, and the repricing will be violent.
My view: the next big repricing won't be triggered by a model release. Nobody's multiple is going to halve because a benchmark moved two points. It'll be triggered by something boring — power prices in a key market, interconnection queues stretching into the 2030s, or an earnings call where a CIO says they're cutting AI software spend because the pilots didn't convert.
So I watch electricity, not GPUs. Grid constraints are the physical speed limit on the AI multiple. You can print money; you can't print megawatts.
Here's my question back to you: what would make you sell the AI trade? Not a smarter model from a competitor — but a utility telling a hyperscaler there's no power until 2031? That's the scenario nobody has priced.