A fictional exchange between two invented investors who embody two opposing schools of thought. 

Cassandra Wilde runs a disruptive-innovation fund and argues that exponential technology breaks conventional valuation.

Gerald Granthwaite co-founded a Boston value house and has spent four decades calling bubbles. Neither is a real person; the views are archetypes, not quotations from anyone living. Moderated by InvestSense

‍

InvestSense: The question is simple to ask and impossible to answer cleanly. Mid-2026: the five largest US companies are roughly 30 per cent of the S&P 500, the Shiller P/E has been north of 40, Nvidia has touched a five-trillion-dollar market cap, and hyperscaler capital expenditure is heading towards a trillion dollars a year. Bubble, or boom?

Cassandra, you first.

Wilde: It's a boom, and the bubble framing is a category error. Every previous general-purpose technology — the steam engine, electricity, the internal combustion engine — looked "overvalued" at the moment of fastest adoption, because the people pricing it were extrapolating a linear world onto an exponential one. The cost of training and running a given level of machine intelligence has been falling at a rate that has no precedent in industrial history. When your unit cost collapses by an order of magnitude every couple of years, the addressable market isn't a market — it's most of the  economy. You don't value that with a trailing P/E. You value it with a learning curve.

Granthwaite: I've heard "you can't value it the old way" at the top of every bubble I've lived through. In 1989 it was Japanese land. In 1999 it was eyeballs and clicks. In 2007 it was the idea that diversified mortgage pools couldn't fall together. The phrase "this time is different" is the single most expensive sentence in the English language. I don't doubt the technology is real. The railways were real, too, and railway equity holders were largely wiped out while the rest of us got to ride the trains. Profound technology and a profitable investment are two completely different propositions, and bull markets routinely confuse them.

InvestSense: Gerald, what's your actual evidence that we're stretched, rather than just expensive?

Granthwaite: Start with arithmetic, because arithmetic doesn't have a narrative. Long-run equity returns come from a small number of building blocks: the starting dividend yield, real earnings growth, and any change in valuation multiple. The US market is starting from a dividend yield near the floor of its recorded history and a multiple near the ceiling. For the index to deliver its historical real return from here, you need either earnings to grow far faster than the long-run trend forever, or the multiple to stay permanently elevated. Both at once. Our own seven-year real-return forecasts for US large caps have been negative. That's not a mood; it's what mean reversion does to a starting point this extreme. Concentration makes it worse, when a third of the index is six or seven correlated names betting on the same theme, you've removed the diversification you thought you'd bought.

Wilde: But mean reversion to what? Your building-block model assumes the trend growth rate is a constant of nature. It isn't. If AI raises the productivity growth rate of the whole economy, even modestly, even by a fraction of a percent compounded, then the "trend" you're reverting to is itself shifting upward, and the multiple that looks insane against the old trend is reasonable against the new one. You're anchoring on a mean that's moving. That's the trap value investors fell into for the entire 2010s: forever short the thing that was actually compounding, forever "right" about valuation and wrong about wealth.

Granthwaite: And here's the problem with that defence: the productivity hasn't shown up. A National Bureau study earlier this year found that the overwhelming majority of firms reported essentially no measurable impact of AI on output or the workplace, even as executives projected gains. That's the textbook productivity paradox, the gap between what's promised and what's banked. I'm not saying it never arrives. I'm saying you're being asked to pay, today, the full price of a productivity miracle that the data cannot yet find. The spending is certain. The return

InvestSense: Let's put a number on the spending. Cassandra, the bear case isn't only valuation, it's the capex.

Wilde: And I'd argue the capex is the most bullish fact in the whole picture, not the most bearish. The companies doing the spending are the most cash-generative businesses in history. They are funding the build-out largely from operating cash flow, not from junk debt the way the telcos did in 1999 when they laid fibre on borrowed money and went bankrupt before anyone used it. These are real customers, real revenue, real chips shipping out the door. The demand for compute is, as far as anyone can measure, insatiable.

Granthwaite: "Largely" is doing heavy lifting in that sentence, and the margin is where the danger lives. The largest model developer has reportedly committed to something on the order of well over a trillion dollars of future spending against revenue measured in the low tens of billions. That is not a cash-flow story; that is a faith-based capital structure. And the financing is increasingly migrating into private credit and off-balance-sheet, data-centre-backed structures — which is precisely the kind of place where, in 2008, the system discovered it owned far more risk than its own reports admitted. I notice credit spreads on this paper have already started to widen. When the marginal dollar of a boom comes from leverage rather than profit, you've moved from boom to bubble whether anyone rings a bell or not.

Wilde: Spreads widen and narrow. One weak jobs report in June knocked the chipmakers down hard, and they were largely back within a fortnight — because the underlying demand didn't change, only the rate-cut timing did. That's volatility, not a thesis. And there's a deeper point: a great deal of this capex is deflationary. It is being spent to drive the cost of intelligence towards zero. Even the bears concede the technology "actually works" — that's the line that separates this from the dot-com names that had no revenue at all. If the infrastructure is overbuilt, society inherits cheap, abundant compute, the way we inherited cheap, abundant bandwidth after 2000. Even Bezos called this an "industrial bubble" — wasteful in places, transformative in aggregate.

Granthwaite: I'd accept "industrial bubble" as a fair description, and notice what it concedes: that there is a bubble. The internet was the real thing, and the Nasdaq still fell roughly 80 per cent and took the better part of a decade to recover. "The technology wins" and "your equity gets cut in half first" are entirely compatible.  They usually go together. The investor's question isn't "is AI important?" — obviously it is. The question is "what return do I earn buying it at this price?" And the honest answer, from the arithmetic, is: probably a poor one, with a fat tail of something worse.

InvestSense: Is there any common ground here? Because you're not actually disagreeing about the technology.

Wilde: We're disagreeing about time horizon and about what a multiple means. I'd accept that the index is concentrated and that a 20-to-30 per cent drawdown in the AI complex is entirely possible — likely, even, at some point. I just don't think that's a reason to be absent. The cost of being out of the genuine winners over a decade dwarfs the cost of riding a correction. The risk I worry about is selling the compounder to avoid the wobble.

Granthwaite: And I'd accept that calling the timing of a bubble is a fool's errand — bubbles inflate further and longer than any sane model predicts, because late-cycle liquidity and fiscal stimulus keep the music playing. So I'm not telling anyone to short it. I'm telling them that price is the one variable that ultimately always matters, that the starting yield-plus-growth maths is unforgiving from here, and that the antidote to extreme concentration isn't a brave forecast — it's owning the parts of the world the narrative has left cheap. International, value, small caps, the unloved. Diversification is the only free lunch, and it's on special precisely because no one wants it.

InvestSense: A closing line each. Bubble — yes or no?

Wilde: No — a generational platform shift, mispriced on the downside by people using yesterday's ruler. The mistake of the decade will be selling too early.

Granthwaite: Yes — a real technology wrapped in an unreal price. The mistake of the decade will be confusing the importance of AI with the wisdom of paying any price for it.

InvestSense: Which is roughly where the real debate sits: nearly everyone agrees the technology is durable, the fight is entirely about the price tag and the horizon. On that, the tape will eventually settle it, though probably not on a schedule that flatters either side.

‍

Both speakers are fictional. The arguments are deliberately drawn from the two best-known camps in the current debate — disruptive-innovation optimism versus valuation-and-mean-reversion scepticism — and the data points (market concentration, Shiller P/E, hyperscaler capex, model-developer spending commitments, the productivity-paradox study, widening AI-credit spreads) reflect publicly reported conditions as of June 2026.