Can Nvidia Defend 75% Gross Margins Into the Capex Cliff?

A gross margin in the mid-seventies is not a number. It is a claim about the durability of a monopoly. The entire AI trade is, in effect, a leveraged bet that this one claim holds.

Start with the structure of the rent. Nvidia's margin is not primarily a silicon story; it is a software-and-switching-cost story wearing a silicon costume. CUDA, the surrounding library stack, and a decade of accumulated developer muscle memory constitute a moat that is expensive to cross even when a competitor's raw FLOPS-per-dollar is attractive. That moat is what lets the company price accelerators at a premium that would be competed away in any commoditized hardware market within two product cycles.

The question is whether the moat survives contact with its own customers' incentives. The hyperscalers — the four or five buyers who constitute the overwhelming majority of data-center demand — are simultaneously Nvidia's largest customers and its most motivated would-be replacements. Each is funding a custom-silicon program precisely because a 75% gross margin paid to a supplier is, from their side of the table, a 75% gross-margin tax they are desperate to route around. Custom accelerators do not need to beat Nvidia on every axis. They need only be good enough for inference at scale, where the workloads are more predictable and the software lock-in is weaker.

Then there is the capex cliff. The current margin is supported by demand that is, in part, a one-time build-out: a land grab for training capacity funded by balance sheets that can afford to prioritize capability over return-on-invested-capital. That phase does not last forever. When the marginal data center has to justify itself on unit economics rather than strategic necessity — when boards begin asking what the AI capex actually earned — the order book becomes more price-sensitive, and price sensitivity is the natural enemy of a 75% margin.

We built a bottom-up model: accelerator ASPs against a custom-silicon adoption curve, gross margin against a normalizing demand mix, and the whole thing against the book. The conclusion of the equity-specific analysis is genuinely two-sided. The bull case — sustained training demand, an inference market that stays GPU-shaped, CUDA lock-in holding — supports the margin for longer than skeptics think. The bear case — inference migrating to custom silicon, a capex digestion year, a single disappointing guide — compresses the margin and, because the position sits near the center of the book, drags the broad market with it. This is the part that matters for an asset allocator: concentration has made the single-stock question into a market question.

An investor with a strong view on the moat could express it directly, sizing the single name to their conviction and their stomach. We hold a strong view on the moat. We simply decline to express it as a single name, because our framework has opinions about idiosyncratic risk that are more durable than our opinions about gross margin.

Running the full distribution of margin paths through our proprietary 14-factor regime engine — weighting the bull and bear silicon scenarios, the capex-digestion timing, and the concentration feedback — the allocation that survives is the one that owns the entire outcome rather than betting on a slice of it.

We did not need to pick the winner. We already own it.

Note the elegant property of this conclusion: if Nvidia defends the margin, our position owns the win, because Nvidia is roughly seven percent of the five hundred and rising. If Nvidia loses the margin, our position owns whichever company captured the rent instead, because that company is also among the five hundred. We are, in the most literal sense, long the outcome of the debate without having to win it. We consider this the highest form of being right.

← Back to Research