NVIDIA books $96.2bn at 75% margins and guides 74%: compute is not getting cheaper

Start with the margin line, because that is what prices capacity next year. Seventy-five points on $96.2 billion is about $72.2 billion of gross profit in a single quarter ^[1][10][20]. Hold 74.0% on the guided $108.0 billion and gross profit is roughly $79.9 billion ^[5][21]: a hundred basis points of compression ^[22] arriving with about $7.8 billion more gross profit than the quarter just closed ^[23]. A supplier trading price for volume does not print that. This one is not being asked to.

The plan assumption to delete, then, is the one that had each accelerator generation landing with softer unit pricing. Cost-per-token improvements are still available, but they have to be found in perf-per-watt, utilization and model-side work, because the vendor's take rate has moved by one point across two quarters ^[10][5].

The shape of growth needs re-basing too. Sequential revenue growth was 18% and the guide implies about 12.3% ^[1][24], while the sequential dollar add goes from roughly $14.7 billion to $11.8 billion ^[26][25]. For scale: the quarter cleared the consensus runtimewire cites, $92.1 billion, by $4.1 billion ^[9][14], which is around a third of the increment already guided for the current quarter ^[16]. Anyone reading earnings surprises as a demand signal is calibrating on the smaller number.

Concentration is the third input to change. Data Center at $89.0 billion is 92.5% of the company ^[2][17], and its sequential add of $13.8 billion ^[11][18] is nearly twice the whole Edge Computing line of $7.2 billion, which grew 27% year over year against Data Center's 117% ^[7][2]. There is no diversified second engine to smooth a Data Center air pocket, and none to compete for fab and rack capacity either.

Where the binding constraint actually sits shows up in the announcements that have nothing to do with silicon. NVIDIA secured land, power and shell capacity with SB Energy at the PORTS-Pike campus in Ohio ^[15], and lined up six financial firms to mobilize outside capital for the buildout, subject to definitive agreements ^[6]. The queue forms around megawatts and balance sheets as much as around dies. Worth noting that in the same quarter, the company sent approximately $26.0 billion back to its own shareholders ^[13], about 44% of GAAP net income ^[3][19], while organizing third-party money for its customers' infrastructure.

Jensen Huang's framing is that compute is revenue ^[8]. From the buy side, that reads as compute being somebody else's revenue at 74 points of margin, for as far ahead as the company is willing to guide ^[5].