The 08-26 column asked whether Huang would quantify the demand picture on the call, and flagged that this column keeps prodding a demand wall. He answered, and then some: NVIDIA issued a rare long-range forecast — roughly 70% revenue growth in fiscal 2028 — and CFO Colette Kress said customers’ forecasts point to their growth doubling next year, “however … we are supply-constrained.” That is the first time NVIDIA has publicly committed to a demand curve two years out, and it confirms with a dated horizon exactly what the earnings print implied: capacity, not customers, is the constraint.
The rare forecast is the news; the memory bill is the fine print
What happened. Beyond the record $96.2 billion quarter ($89.0B data center, +117%, GAAP and non-GAAP margin 75.0%) and the $108 billion Q3 guide, the call carried the long-term outlook that NVIDIA typically never gives: ~70% fiscal-2028 revenue growth. AWS said the pair will deploy an additional 2 million NVIDIA GPUs across its global infrastructure in 2027 and 2028. The fine print is cost: soaring memory prices push margins to roughly 74% in Q3 and a bottom near 71-72% in Q4. The market read it as the boom being intact: shares rose 6.7% to $223.71 before the bell, roughly $340 billion of market value, with CoreWeave +5.8% and Nebius +7.2%; at least ten brokerages raised targets, Morgan Stanley calling “70% growth supply constrained … a remarkable figure.”
Why it matters. The demand wall is not the thing to watch anymore — NVIDIA just committed to two years of ~70% growth in writing and named the real constraint: memory, not orders. For anyone provisioning, the unwelcome corollary is that 2027-28 compute does not get cheaper on this evidence; it gets supply-capped and memory-cost-inflated, which pushes against the cheaper-token thesis this column has been pricing. The legible, bookable signal is the AWS 2M-GPU ramp — that is hyperscaler capacity you can actually plan against, and the rented-neocloud supply chain just got a multi-year demand backstop from the chipmaker itself.
Source: nvidianews.nvidia.com, reuters.com, reuters.com, seekingalpha.com
What I’m watching
Sounds like the sell side re-pricing to a ~70% fiscal-2028 model — at least ten houses already moved today — and whether Vera Rubin revenue shows up in the Q3 data-center split. The scorecard number is the Q4 gross-margin bottom: if memory prices keep climbing, the $108B guide is a ceiling, not a floor.