What a Base Die Requal Did to NVIDIA's Rubin Forecast
A KeyBanc Capital Markets preview note ahead of Nvidia's August 26 print raised its annual supply estimate for the Rubin (R200) GPU from 1.75 million units to roughly 1.9 million. The reason given: SK hynix has requalified its HBM4 base die and started shipping. That is a narrower, more specific claim than "SK hynix won HBM4" headlines usually carry, and it is worth taking apart on its own terms rather than folding into the broader Samsung-versus-SK hynix scoreboard.
This post does one thing: it converts that single word, requalification, into hardware — stacks, gigabytes, and terabytes per second — and checks whether the resulting numbers are consistent with what both companies have already disclosed.
KEY TAKEAWAYS
1. The revised Rubin forecast (about 1.9 million units) implies roughly 15.2 million HBM4 stacks and 547 million GB of HBM4 capacity a year for this one GPU family alone — a figure that sits inside SK hynix's own, already-trimmed 2026 HBM4 output plan.
2. Nvidia's published Rubin spec of "288GB, up to 22TB/s" lines up with a 10.6 Gbps pin speed, not the 11.7 Gbps top tier — meaning a base-die requalification is not the same claim as securing the fastest available grade.
3. The KeyBanc note itself is internally inconsistent on the exact numbers (25 vs. 20 million incremental R200 units in different sections; 1.9M vs. "1.9-2.0M" for the full-year figure) — this post uses the more conservative reading throughout.
Update, Aug 29: Nvidia reported fiscal Q2 2027 results on August 26 — $96.2 billion in revenue and a Q3 guide of $108 billion. The requalification story below predates that print; this post is about the volume math behind one word in a preview note, not a forecast of the print itself.
What "requalification" actually means here
Base die requalification is a specific step in the HBM qualification sequence: development complete, sampling, qualification, requalification (after a design or process change), mass-production approval, and shipping are all different stages. KeyBanc's note credits SK hynix with clearing requalification and beginning shipments — not a new spec win, and not an official company announcement. The underlying story has been messier than a single headline suggests.
In late 2025, Korean trade press reported SK hynix might be redesigning its HBM4 base die after an I/O speed shortfall and very low viable-sample yields; the company called it unfounded rumor and said it was finishing optimization. By February 2026, Samsung had started HBM4 mass-production shipments at 11.7 Gbps. By March, trade press (TrendForce, KED Global) reported both Samsung and SK hynix had been tapped as Vera Rubin HBM4 suppliers. In mid-April, Digitimes and others reported SK hynix was trimming its 2026 HBM4 volume plan 20-30%, citing friction in the Rubin ramp itself. In June, Jensen Huang said all three memory makers had qualified and were shipping. On its Q2 call in late July, SK hynix said HBM4 mass-production shipments had started that quarter, yield was nearing mature HBM3E levels, and second-half volume would expand well beyond the first half.
The KeyBanc requalification note in August sits at the end of that sequence. A July call describing quarter-old mass shipments and an August note describing a requalification "success" are consistent only if they describe different things — most likely different speed grades, with the top-tier grade still open. That reading matches Korean reporting that Nvidia is evaluating two pin-speed tiers (10.6 Gbps and 11.7 Gbps) for Rubin, with the top-tier award reportedly not decided until year-end.
Seven months of HBM4 headlines end with a base-die requal, not a new spec win.
From GPU count to gigabytes: sizing the revised forecast
Take KeyBanc's revised figure — roughly 1.9 million R200 units a year, the more conservative end of the range the note itself gives — and convert it using Rubin's own published memory configuration: 288GB of HBM4 per GPU, delivered as eight stacks of 12-hi, 36GB HBM4.
At 1.9 million units, that is about 15.2 million HBM4 stacks and 547 million GB — or 43.8 billion Gb, in the unit Korean HBM coverage typically uses — of HBM4 capacity a year, for this one GPU family alone. The incremental 150,000 units on top of the prior estimate work out to roughly 1.2 million additional stacks and 43.2 million additional GB (34.6 billion Gb).
A 1.90 million-unit Rubin forecast converts to about 547 million GB of HBM4.
ZDNet Korea reported on April 14 that SK hynix's original 2026 HBM4 output plan of about 60 billion Gb (750 million GB) had been cut 20-30%, to a range of roughly 42-48 billion Gb (525-600 million GB) — while HBM3E output was expected to exceed 80 billion Gb for the year.
Line those two figures up and the structural point becomes visible: one GPU family's calculated HBM4 draw (43.8 billion Gb) falls squarely inside SK hynix's own, already-reduced annual HBM4 plan. Put differently, a single memory supplier's full-year HBM4 output plan is now roughly the same order of magnitude as what one customer's GPU line alone would consume. That is the structural read this post is built around — not who "won," but how thin the margin for error has become once volume, not just qualification status, enters the picture.
Why "22TB/s" is not proof of the top-tier win
HBM4 widens the interface to 2,048 I/O per stack (versus 1,024 for HBM3E), so bandwidth per stack is I/O count times pin speed, divided by eight (bits to bytes). At the JEDEC floor of 8.0 Gbps, that is about 16.4TB/s across eight stacks. At a mid-tier 10.6 Gbps, it is 21.7TB/s. At the top-tier 11.7 Gbps, it is 24.0TB/s.
Nvidia's own published Rubin (R100) spec sheet lists "288GB, up to 22TB/s." That number sits almost exactly on the mid-tier 10.6 Gbps calculation (21.7TB/s), not the top-tier 24.0TB/s figure. If the public spec itself was written around the mid-tier pin speed, then a base-die requalification — even a genuine one — does not by itself confirm that the fastest grade has been secured. It is consistent with a supplier clearing the bar the published spec actually requires, which is a real result, just a different one than "fastest available memory."
Nvidia's published 22TB/s spec lines up with the mid-tier pin speed, not the fastest one.
KeyBanc raised its annual Rubin (R200) supply estimate by roughly 150,000 units.
What I actually watch
| Checkpoint | Why it matters |
|---|---|
| SK hynix's next quarterly HBM4 mix disclosure | Whether second-half bit growth clears the mid-tier pace implied here, or stays near the trimmed low end |
| Confirmation of a top-tier (11.7 Gbps) award for Rubin | The gap between "qualified and shipping" and "shipping the fastest grade" — reportedly not settled until year-end |
| Whether R200's HBM4 configuration matches R100's published spec | All the bandwidth math above uses R100's public spec; R200 could differ |
| HBM4E sampling and 2027 ramp commentary | Samples due to major customers in H1 2026 with a 2027 mass-production target — the next qualification cycle starts before this one resolves |
Value chain read-through
| Segment | Read-through |
|---|---|
| HBM suppliers (Samsung, SK hynix, Micron) | Even a "conservative" GPU forecast now implies HBM4 volumes on the same scale as a supplier's full annual plan — tight capacity, thin room for execution slips |
| Foundry / base-die process (TSMC 12nm for SK hynix base dies, Samsung's own 4nm) | Base-die requalification cycles are now a direct lever on GPU-side supply estimates, not just a memory-side detail |
| Advanced packaging / CoWoS-linked assembly | Stack count (15.2 million implied) flows straight through to interposer and substrate demand |
Risks to this view
• The KeyBanc figures are a single sell-side preview note, not a company disclosure, and are internally inconsistent on the exact incremental unit count and full-year range — this post used the more conservative numbers throughout.
• No primary source (filing or press release) confirming an SK hynix HBM4 base-die requalification was located; the claim traces to the KeyBanc note and industry reporting, not a company statement.
• The bandwidth math assumes Rubin's HBM4 configuration matches Nvidia's published R100 spec; R200's actual configuration has not been independently confirmed.
• The 10.6/11.7 Gbps tier framing comes from Korean trade press describing Nvidia's evaluation process, not from an official spec disclosure.
The next post in this line will look at how yield actually degrades as HBM layer count rises, using public conference disclosures (IEDM, ISSCC) rather than qualification-stage headlines — the harder, less visible half of the story this one only touches at the margins.
Sources: KeyBanc Capital Markets Nvidia earnings preview (Aug 2026, Korean-language summary); ZDNet Korea (Apr 14, 2026); TrendForce; KED Global; Digitimes; SK hynix Q2 2026 earnings call; Nvidia published Rubin (R100) specifications. Links where available in the original reporting.
Disclaimer: This post is for informational and educational purposes only. It does not constitute investment advice or a recommendation to buy or sell any security. All investment decisions are your own responsibility.
Comments
Post a Comment