Nvidia Crossed the $5 Trillion Mark, and the Sell Side Is Still Catching Up

Nvidia's revenue went from $26.9 billion in FY22 to $130.5 billion in FY25 to a TTM run-rate near $190 billion as of the most recent quarter. The compound growth is roughly 90% annualised over four years. Free cash flow expanded from $8.1 billion to over $90 billion. Operating margin moved from the 27% range to north of 60%.

This is the most extraordinary financial trajectory in the history of large-cap technology. AWS, in its 2015-2018 inflection, saw revenue compound at roughly 50% annualised. Nvidia has done nearly twice that pace at twice the absolute revenue base.

The consensus error has been to anchor on prior cycle frameworks. The 2018 crypto-driven Nvidia cycle peaked at roughly $30 billion in TTM revenue and reset hard. The current cycle has structural drivers that did not exist in 2018: hyperscaler capex commitments through 2027, a sovereign AI buyer cohort, and an enterprise infrastructure replacement cycle that has not yet started in earnest.

Sell-side estimate revision has been positive for eight consecutive quarters. Buy-side positioning, as inferred from the institutional ownership data we track, has been net additive throughout the period. Short interest sits near the lowest level for the stock in five years.

This is not what a top looks like. Tops are characterised by negative revision dispersion, decelerating beat magnitude, and short interest building. Nvidia exhibits none of those features. The beat magnitude has held at 5-10% for the trailing four quarters, and the company's own forward guidance has been progressively raised in each subsequent quarter.

What this signals is that the buy-side consensus has not yet caught up to the underlying demand curve. We have seen this pattern before in semiconductor cycles, most notably during the 2009-2014 mobile silicon cycle when Qualcomm and ARM-related plays consistently surprised to the upside for roughly five years before the cycle peaked.

The Nvidia pattern looks similar in shape but the demand cycle behind it is structurally different. Mobile silicon was a unit-volume story; AI silicon is a revenue-per-system story. The latter has higher margin durability.

The bear case has shifted three times in two years. The first bear case was that AI training was a one-time build. That argument was disproved by 2024, when inference workloads accelerated faster than expected. The second was that custom silicon (TPU, Trainium, MTIA) would erode Nvidia's share. That has been a slow burn; custom silicon has expanded share at the largest hyperscalers but the absolute Nvidia revenue from those same customers has grown faster than the share loss implies.

The third and current bear case is the most credible. It argues that the AI capex cycle has front-loaded, that 2026-2027 will see absorption rather than expansion, and that the consensus revenue estimates do not yet reflect the digestion phase. We give this argument 30-35% probability over an 18-month window.

What would confirm it: a single hyperscaler quarterly print where capex guidance is held flat or reduced. We have not seen that yet. Microsoft, Google, Meta, and Amazon all guided capex up sequentially in their most recent prints. That is the binding constraint on the bear thesis.

Forward P/E of 26.5x is below the technology peer average for hypergrowth franchises. Salesforce traded at 60x at maturity. Adobe traded at 50x. Microsoft traded at 40x at the peak of the post-pandemic cycle.

The Nvidia multiple compression to 26x is in part a function of the denominator growing fast enough that even in absolute price terms, the stock can move higher while the multiple compresses. That is a healthy multiple dynamic, not a cautionary one.

The relevant valuation question is not 'is 26x forward expensive' but 'is the FY27 EPS estimate that produces the 26x multiple correct'. Consensus FY27 EPS is roughly $5.85. The implied trajectory requires 35% earnings growth from the trailing run-rate. That is below the 73% revenue growth pace and assumes operating margin compression. We think it is achievable; the bears think it is heroic.

The historical parallel for this multiple position is Apple in 2016-2017, when the stock traded at roughly 12-14x forward and the bears argued the iPhone refresh cycle was peaking. That cycle ran for another seven years. The pattern of underestimating durable platform franchises is consistent.