NVIDIA's 70% FY2028 Call Is a Margin Story Disguised as a Growth Story

NVIDIA has never before published a year-ahead revenue range. On the Q2 FY2027 call, Jensen Huang broke that habit and pegged fiscal 2028 growth at \"approximately 70%\" — a figure that, applied to fiscal 2027's $396 billion Street consensus, implies roughly $673 billion of FY2028 revenue, well above the $570 billion the Street was modeling (44% growth, per LSEG). The after-hours stock reaction — a roughly 4–5% pop on Aug 26, 2026 — was the headline takeaway. It is also the wrong one.

The 70% line is anchored to what NVIDIA can physically ship, not what hyperscalers want to buy. Read together with the parallel reset on memory pricing — CFO Colette Kress used the words \"extreme pricing conditions in memory\" on the call — the forecast functions as a preemptive defense against two narratives at once: AI-capex deceleration and a memory-driven gross margin squeeze.

Q2 FY2027 GAAP and non-GAAP gross margins were both 75% — a high-water mark for the AI cycle. Q3 guidance is 74%, plus or minus 50 basis points, and Kress guided margins to bottom at 71–72% in Q4 FY2027 before settling at 72–73% in fiscal 2028. That is a deliberate ~300-basis-point step-down, blamed squarely on memory pricing rather than competition or product mix. Management says executed price increases take effect in Q1 FY2028 and lift the curve back to 72–73%.

For investors trained to price NVIDIA on revenue growth, the gross margin glide path is the more actionable signal. It tells you the supply chain is now the bottleneck and that dollars of incremental growth come with rising input costs — which is precisely the precondition for memory and packaging suppliers to compound.

Huang's language on the call was unusually direct: \"as you are already aware, we are experiencing extreme pricing conditions in memory... we are resetting expectations today,\" and \"we expect supply to remain a bottleneck at least through the end of fiscal year 2028.\" The AI buildout itself is driving the scarcity, with memory content per Blackwell Ultra and Rubin systems rising sharply versus Hopper. That structural shift hands pricing power to the three HBM vendors — SK Hynix, Samsung Electronics, and Micron.

SK Hynix leads on incumbent HBM3E share and is forecast by UBS to capture ~70% of the HBM4 market for NVIDIA's Rubin platform in 2026 — the platform NVIDIA is shipping into the AWS buildout. Q2 2026 revenue at SK Hynix reached ₩79.32 trillion (~$54.55 billion), up 257% year over year, with operating profit of ₩60.54 trillion — record absolute levels even after the company missed the ₩84 trillion consensus.

Micron is the cleanest pure-play on the bottleneck: fiscal Q3 FY2026 revenue of $41.46 billion (+346% YoY) with gross margin of ~84.6% — the highest in the memory group — and management has stated HBM is sold out through calendar 2026. Cloud Memory revenue alone hit $13.77 billion in the quarter; Core Data Center revenue $11.52 billion. With Rubin HBM4 ramping, the FY2027 fiscal year at Micron is structurally backstopped regardless of PC and smartphone weakness.

Memory is not the only chokepoint. NVIDIA's Blackwell and Rubin generations are bounded by TSMC's CoWoS advanced packaging capacity, where NVIDIA is reported to hold roughly 60% of allocated wafer volume through 2026–2027. TSMC's Q2 calendar 2026 revenue hit TWD 1,270 billion (~$40 billion), up 36% YoY, and Q1 calendar 2026 was TWD 1,134 billion — a sequential ramp that confirms CoWoS supply, not demand, is the swing variable. Morgan Stanley and other trackers put NVIDIA's 2026 CoWoS demand at ~780,000 wafers, rising to ~1.2 million in 2027.

Broadcom sits on the other side of the same buildout. Its custom AI ASIC business — designs for Google, Meta, and now increasingly others — is the structural beneficiary of hyperscaler appetite for non-NVIDIA silicon, and the company just delivered Q3 FY2026 revenue of $29.59 billion (+85% YoY), with networking and AI-related segments driving the mix. The 70% guide for NVIDIA does not contradict Broadcom's trajectory; it complements it, because every hyperscaler that wants optionality against a single-supplier compute stack is funding its own ASIC program in parallel.

The same day Huang issued the 70% guide, Amazon.com announced an expansion of its NVIDIA partnership: 2 million additional GPUs across AWS's global infrastructure in 2027–2028, covering Blackwell Ultra, Rubin, and Rubin Ultra silicon, plus a separate 100,000-GPU AI factory program for U.S. government workloads. This is on top of the 1 million GPUs AWS had already committed to at GTC in March 2026 — effectively a tripling of the AWS commitment in five months.

That order book is what makes the 70% ceiling mathematically defensible. Aggregate 2026 AI capex disclosures from the four largest hyperscalers now sit near $725 billion — Amazon ~$200B, Microsoft ~$190B, Alphabet $175–205B, Meta $125–145B — and Goldman Sachs projects cumulative 2025–2027 capex at $1.15 trillion, more than double the $477 billion spent in 2022–2024. Each new Blackwell or Rubin rack carries materially more HBM and more CoWoS than the prior generation, so hyperscaler dollar growth is mechanically amplified into memory and packaging units.

If NVIDIA grows 70% in FY2028 — and that growth is supply-bounded — the most direct loser is the assumption that AI silicon demand is up for grabs. AMD, the only credible merchant GPU alternative, posted Q1 calendar 2026 revenue of $10.25 billion, up from $7.4 billion a year earlier but a rounding error against NVIDIA's $96.22 billion. The MI400 ramp in late 2026 will add share but cannot dent the trajectory: Huang is not fighting for incremental sockets — he is fighting for memory allocation.

Short-term (days–quarters): The market will likely re-underwrite memory and packaging multiples higher as HBM ASPs reset upward. Watch the Q3 calendar 2026 prints from Micron (late September 2026) and SK Hynix (October 2026) for confirmation of the \"extreme pricing\" language. The Q4 FY2027 NVIDIA margin trough (71–72%) is the next binary setup — a beat there, driven by faster-than-expected price increases, would compress the gap between the 70% guide and a higher growth rate.

Long-term (1–3 years): The structural call is that NVIDIA grows at 60–70% through FY2028, gross margins stabilize at 72–73%, and the dollar contribution of memory and packaging per system trends upward. That re-rates the supply chain — not NVIDIA, whose 70% is largely priced in. Risks: any signal that the AWS-style deal flow is front-loaded (i.e., the 2M GPU deal reflects a 2027 pull-in, not a true increment); a faster CoWoS expansion by TSMC or a breakthrough in advanced packaging by Samsung Electronics that breaks the bottleneck; or any hyperscaler CFO publicly softening 2027 capex commentary.