The Backlog Is the Story

The cleanest read on AI this week was not a model release. It was Chris Semenuk pointing out that the backlog at GE Vernova is north of almost $90 billion. That is a power-equipment order book, not a chip order book, and it is the shape of the AI trade now.

Semenuk’s framing is the spine: power demand growth in the US, he says, ran at zero to minus one for two and a half decades and is now running 2 to 4. That is the regime change. Everything else this week is downstream of it. NextEra, per Semenuk, just guided to 8 to 9% EPS growth plus a 4% dividend yield out to 2035. Utilities do not usually talk in decade tenors. They are now because someone has to build what the hyperscalers ordered.

Morgan Stanley’s Michael Zezas put a number on how big this has gotten relative to the actual economy: his economics team has roughly a quarter of US GDP growth this year coming from the current AI buildout. That is not a sector call. That is AI capex showing up as the macro.

The equity market is still pricing it as a tech story, and that is where the tension lives. Zezas also flagged that consensus 2027 EPS growth for AI infrastructure stocks is 42%, which he called insane. Josh Brown noted that equal-weight tech (RSP) is up 39%. The exuberance is sitting on the semis and the Mag 7, not on the things actually constrained. Intel’s Lip Bu Tan said the quiet part flatly: “we are supply constrained.”

The capital markets tell on themselves. A senior fixed-income investor noted that Nvidia is raising $20 billion, its first raise since 2021, and that Google did a $10 billion equity offering, its first since around 2001. Companies that print cash do not tap markets unless the build in front of them is bigger than the cash flow behind them. Semenuk’s other tell: 70% of the IIJA has already been committed, and the next leg, he argues, is not roads but kitting out the factories and substations that got built.

The honest way to hold this: the AI trade and the power/industrial trade are the same trade, and only one side of it is priced for 42% earnings growth. Vernova’s backlog and NextEra’s 2035 guide are the same sentence as Tan saying he cannot make enough silicon. The model benchmarks (GLM 5.2 at a tenth the cost of Fable 5, per an AI markets analyst) keep compressing the value of any single model while expanding the value of whoever pours the concrete and wires the substation.

The week’s quietest line came from Semenuk: manufacturing jobs are not coming back, but manufacturing is. That is the spine. The AI story stopped being about who has the best model and became about who has the turbine slot in 2028.