geoeconomics | commentary | note on nvda earnings (20251120)

NVDA beat analyst earnings expectations today by \$0.04/share and beat revenue expectations by a little over \$2B. Financial media quoted analysts' opinions that this signaled a lack of market concern regarding an AI bubble. This short note instead argues that NVDA beating estimates carries essentially no information about an AI bubble.

Bayes' theorem tells us that \[ p(\text{bubble} | \text{NVDA beats earnings}) = \frac{p(\text{NVDA beats earnings} | \text{bubble}) p(\text{bubble})}{p(\text{NVDA beats earnings})}, \] and likewise for the no-bubble case. (Yes, we know it's more complicated than "there is a bubble or there is not.") We can compare the hypotheses of bubble versus no-bubble via the Bayes factor: \[ BF(\text{bubble}) \equiv \frac{p(\text{bubble} | \text{NVDA beats earnings})}{p(\text{no bubble}|\text{NVDA beats earnings})} = \frac{p(\text{NVDA beats earnings} | \text{bubble})}{p(\text{NVDA beats earnings} | \text{no bubble})}\frac{p(\text{bubble})}{p(\text{no bubble})}. \] Leaving our priors out of it momentarily, it is worth examining the two respective likelihood terms. For the purpose of this note we will define a bubble as the state where market valuation greatly exceeds the "fair value" \( E[\sum_{t\geq 0}\frac{R_t - C_t}{(1+r_t)^t}] \) computable using all publicly available information.

\( p(\text{NVDA beats earnings} | \text{no bubble}) \)

In the case where there really is no bubble, we should expect NVDA's valuation to fluctuate around its fair market valuation (by definition). In other words, in this case -- and all other variables marginalized out -- it's roughly equally likely that NVDA beats or misses earnings. We can represent our probability estimate by any symmetric interval in this case; We'll use (almost) ICD 203 language and say that there are "roughly even odds" that NVDA beats or misses earnings, so setting \( p(\text{NVDA beats earnings} | \text{no bubble}) = [0.4, 0.6] \).

\( p(\text{NVDA beats earnings} | \text{bubble}) \)

This case is more complicated. We'll enumerate three mutually exclusive possibilities here; you can probably think of more.

1. AI companies (both hyperscalars and frontier developers) know their claims are overhyped and know that they cannot continue to invest in physical capital, so they slow down purchasing of chips. (This seems to be the signal for which many analysts are searching.) In this case, it does indeed seem likely that NVDA misses earnings, so we'll set \( p(\text{NVDA beats earnings} | \text{bubble, case 1}) = [0.2, 0.4] \). (Maybe it should be lower?)
2. AI companies don't know that their claims are overhyped and continue to invest in chips. In this case, though the causal mechanism is completely different, in practice this looks like fair valuation as far as NVDA is concerned -- the demand is actually still there! -- even though there is a bubble because long-run fundamentals (e.g., limits on marginal cost of tokens under a transformer architecture) do not support the beliefs of the AI company executives. In this case we use the "fair valuation" probability again: \( p(\text{NVDA beats earnings} | \text{bubble, case 2}) = [0.4, 0.6] \).
3. AI companies are engaged in oligopolistic pursuit of both market share and capital to claim dominance in the "AI competition", a classic bubble formation story. With this rationale -- and with the incredible ease with which they are able to fundraise -- they can actually increase the demand for chips for both signaling (to investors) and scaling (for market penetration) purposes. In this case, even though there is a bubble, NVDA ia likely to beat earnings: \( p(\text{NVDA beats earnings} | \text{bubble, case 3}) = [0.6, 0.8] \).

You can argue with minutae of the probability judgements but we think the point should be fairly clear. In sum, since we have no a priori reason to favor one hypothesis over the other, we have \[ p(\text{NVDA beats earnings} | \text{bubble}) = N^{-1}\sum_n p(\text{NVDA beats earnings} | \text{bubble, case }n), \] which in this case turns out to be \( p(\text{NVDA beats earnings} | \text{bubble}) = [0.4, 0.6] \). This makes the posterior probabilities exactly equal to the prior probabilities; we gained no information about the bubble question.

House view

We have not been shy about disclosing our view that the transformer-based large-model ("AI") boom is a sizable bubble. Hyperscalars and frontier AI companies will continue to fill NVDA's orderbook for as long as capital is easy and cheap for them to access. As in all coordination games, it will take a very powerful signal to break the equilibrium of providing AI companies with cheap capital. If that signal emerges -- e.g., in the form of an esteemed credit institution publicly rejecting a debt solicitation from a premier AI firm -- the bubble will burst very quickly.