Nvidia guides fiscal 2028 to $673B to give its memory suppliers a planning horizon

Back the base out of the guide and the arithmetic gets calmer than the growth rate does: $673B at roughly 70% growth implies fiscal 2027 finishing near $396B ^[10], and 44% growth on that same base is about $570B ^[11], so the guide sits about 18% above where the Street's dollars were rather than at twice them ^[12]. The doubling lives in the percentage, which is a different statement about a different quantity. The $103B of daylight between the two figures is within about 5% of the single quarter Nvidia has already guided, $108B give or take 2% ^[7][13] - one extra current quarter, spread across four.

That base is itself a forecast. Reported Q2 of $96.2B ^[2] plus guided Q3 of $108B ^[7] is $204.2B, which leaves Q1 and Q4 to sum to roughly $192B for the $396B base to hold ^[17], and a soft closing quarter shrinks the denominator and takes the headline number with it, because 70% of less is less. (Five years ago the company billed about $27B a year ^[9], so the guide is roughly 25 times that ^[16], which is the figure that will get quoted.)

The purpose is stated plainly in CryptoBriefing's account of the August 26 report: Huang tied the longer horizon to supply chain partners needing longer planning horizons, with the components for next-generation parts, high-bandwidth memory above all, already constrained ^[8], and he was candid that supply caps what can ship, implying customer demand runs above the 70% ^[4]. So the number is a build plan more than a demand disclosure. It asks for an average quarter of $168B against the $108B guided now, a 56% step up in run rate ^[14], and applying the reported data-center share puts about $599B of it in data centre ^[19], which is a line item on somebody else's capital budget.

This reads a few different ways. One, supply really is the binding limit, the guide is a floor, and any loosening in memory is upside. Two, the guide is an anchoring device that gets memory and foundry capacity built on suppliers' balance sheets, and the material describes no purchase commitment or prepayment sitting behind it, only the forecast ^[3] - Nvidia has given up the quarterly reset option and bought two years of supplier planning with a sentence. Three, demand is real but concentrated in a handful of buyers whom this material does not name ^[3], in which case $673B is a bet on a small number of capex committees.

On the evidence available, the second reading carries the most weight, because the only reason given for extending the horizon is supplier planning and the only ceiling named is memory ^[8][4], and TSMC, Samsung and SK Hynix are the parties who must now size capex against a public number while managing their own capacity limits ^[5].

That read would break under either of two conditions: fiscal 2027 exiting below the implied $396B base ^[10], which would mean the growth rate was quoted off a base that never arrived; or gross margin settling below the 71-72% band already flagged for the quarters after Q3 ^[1], which would say the supply is costing more than the guide assumed and the memory constraint is being cleared with cash rather than with time.