The biggest complaint I have with the A.I. “bubble” narrative is also the simplest: The run-up in tech shares this year hasn’t succeeded in getting the S&P back to record highs, and the aggregate index valuation isn’t anywhere near dot-com bubble levels.

It’s hard for me to imagine that “peak A.I.” will occur without new highs on tech-dominated, cap-weighted benchmarks. If you look at the Nasdaq Comp, a sizable share of constituents still trade miles below their two-year highs. And the A.I. frenzy is all of six months old, at most. The Bored Ape Yacht Club mania lasted longer than that.

The simple figures below are helpful, I think. One notable is that Meta, Alphabet, Apple and Microsoft aren’t trading all that expensive compared to the S&P. Tesla is, naturally. And so is Nvidia. But you’re never going to get a “fair value” price for Tesla and if you believe the “picks and shovels” Nvidia narrative, I suppose you could justify paying ~50x. If the A.I. gold rush proceeds according to optimistic projections, revenues and (hopefully) profits can “grow into” the share price, bringing down the multiple — assuming, of course, the shares don’t keep rising in perpetuity.

Another notable from the figure on the right above is the extent to which the median multiple has tracked the aggregate multiple very closely during this tech “bubble” compared to the almost complete decoupling in the dot-com era.

In the latest installment of the bank’s “Top Of Mind” series (essentially compendiums of interviews and commentary around whatever the market/macro topic du jour happens to be), Goldman’s Dominic Wilson and Vickie Chang offered four points of caution on the burgeoning A.I. mania.

“Bubbles are complicated phenomena, often driven by momentum and self-fulfilling price dynamics, but several reasons explain why productivity booms can lead markets to overpay,” they wrote. Below are those four reasons, abridged and truncated for brevity.

1. Investors may fall prey to a fallacy of extrapolation. With genuine innovation, productivity gains will be real. To the extent that markets price initial increases in profit growth as persistent, the long-term potential shift in the earnings trajectory may be overestimated.
2. Investors can fall prey to a fallacy of aggregation. During periods of innovation, some individual companies may be capable of stretches of stunning earnings growth driven by a new technology. But it is a mistake to assume that what can be true for an individual company can be true on aggregate. With potential “winners” sometimes more obvious than losers, investors may price a chance of increased profitability across a broad range of potential winners. The result may imply a rate of economy-wide profit growth that is unlikely to be feasible.
3. Activity fueled by the bubble itself can appear to justify the optimism. As asset prices rise, they may encourage a boom in investment and consumer spending. This in itself may provide a boost to the profitability of companies supplying those areas. But if increased revenues and profits are ultimately based on unsustainable demand that is generating economic imbalances, then those gains too will eventually unwind. In other words, a domestic boom created by overvalued asset prices can fuel the perception that higher profit growth can be maintained.
4. To the extent that an acceleration in productivity growth leads to monetary policy that is easier than it “should” be, it can help fuel asset price overvaluation.

That latter point is especially interesting right now. Goldman noted that it’s “particularly a risk when a boom overlaps with other disinflationary forces, as it did for the US in the late 1990s.” Currently, disinflationary forces are in short supply, and the Fed is on its toes, but it’s likely the neutral rate is underestimated.

If any A.I. boom does act to suppress inflation, that’d be welcome, but it could also prolong the state of denial at the Fed vis-à-vis conceding that r-star is higher, thereby raising the risk that rates never make to “sufficiently restrictive” at the risk of unanchored longer-term expectations.

The figures above are self-explanatory. The chart on the left gives you a sense of how quickly ChatGPT became ubiquitous and the figure on the right is eye-opening to the extent it shows that as recently as a year ago, tech firms were barely discussing A.I. on analyst calls outside of Nvidia and Meta.

In the same compendium, Goldman’s Ryan Hammond and David Kostin chimed in on “the perils of euphoric expectations.”

“At the index level, long-term EPS growth expectations are roughly in line with historical averages, suggesting investor optimism on A.I. adoption is not at extreme levels [but] at the stock level, the current valuation of the largest A.I. beneficiaries, like NVDA, is similar to the valuation accorded in the 2000s to some of the largest dot-com boom beneficiaries, though not as high as the most extreme example (CSCO),” they wrote, before reminding investors that “even though most TMT companies were still able to generate strong sales growth between 2000 and 2002, the failure to meet lofty investor forecasts led to a sharp 50%+ contraction in P/E multiples and a plunge in share prices.”