In terms of concentration risk, the AI trade’s now on par with the largest bubbles of the past 150 years.

I mentioned this a few days ago, but I’ll reiterate it here: The “Big 10” AI names, defined as the Mag7 plus AMD, Broadcom and Micron, now comprise more than 40% of overall US equity market cap.

That’s comparable to, or in some instances slightly larger than, the share of market cap commanded by the leadership during four previous bubbles. The exception’s the railroad boom in the latter half of the 19th century, when related stocks at one point comprised nearly two-thirds of market cap in America.

The (mordant) joke writes itself: That’s a lot of upside! For the AI names, I mean. Ostensibly, they can claim another ~20ppt worth of market cap share between them before the train flies off the tracks.

Snark aside, I wouldn’t rule such a scenario out. If Micron’s sales guide is any indication, demand for crucial AI components remains insatiable. Bloomberg reported Friday that Broadcom’s lining up $60 billion in new financing to help Anthropic access compute as the company gears up for a historic public offering before Thanksgiving. And on and on.

Assuming there’s more gas in the tank (coal in the tender) for the AI trade and that the associated concentration risk rises further towards the railroad boom extremes, it’s worth exploring the analogue.

The railroad bubble, BofA’s Michael Hartnett reminded investors in his latest weekly, came in two waves. One lasted from 1861 to 1872, the next from 1877 to 1881. At the peak during the first wave, spending on the buildout reached 5% of US GDP, before European financing dried up ahead of the 1873 bust.

If you assume hyper-scaler capex of between $1.2 trillion and $1.4 trillion in 2027, that’d be less than 5% of contemporaneous US GDP, which is to say short of the railroad boom peak (and, Hartnett noted, lower than the same metric for the fiber-optic network buildout in late-1990s).

So that’s — I’m not sure what the right word is — “comforting,” I guess? More importantly, prices for “bubble inputs,” if you like, are still rising, and rather dramatically at that. Just ask HBM margins.

Soaring prices are in stark contrast to the conditions which prevailed at the end of the second phase of the railroad boom when, as Hartnett recounted, excess capacity and freight-rate deflation meant “revenues and profits were unable to sustain the capex boom.”

The bad news, which is to say the argument for why the AI bubble may be living on borrowed time and thereby won’t likely match the railroad boom on sundry metrics of “revolution”ary excess, is straightforward: Rates are rising.

“The railroad booms were backstopped by an era of falling UST yields,” Hartnett wrote. “That’s evidently not the case today.” No, “evidently” not, and the irony is that the tsunami of IG debt supply to fund AI capex is a factor in explaining why long-end Treasury yields are perched at the highest levels in decades.

The figure above, from the same Hartnett note, shows you 10-year US Treasury yields alongside railroad bond yields for the period encompassing the two booms.

To reiterate what’s quickly becoming an ad nauseam refrain in these pages amid the staggering run-up in G7 bond yields: Rates will matter eventually.

For now, though, that seems a quaint notion in the face of a technology with the potential to save our species (as long as it doesn’t “decide” to kill us first in what, if you ask me, would be a welcome coup de grâce). As Hartnett quipped, describing the zeitgeist, “Rates don’t matter when you’re curing cancer.”