HBM4 Squeeze: SK hynix Sells Out Half Its 2027-28 Capacity

SK hynix has locked in long-term agreements for more than 50% of its 2027-28 HBM4 production capacity, and it announced the figure alongside the first ₩100-trillion quarter in its history. Days later, reports said Nvidia cut Vera Rubin's HBM4 bandwidth from 22 TB/s to 19.2 TB/s per GPU, with no public explanation. The two stories connect: memory makers can sell every stack they build, and Nvidia needs eight of them on every Rubin GPU.

What do the record numbers say?

SK hynix's Q3 2026 results, reported Oct. 28, show revenue of ₩104.6155 trillion, operating profit of ₩83.3338 trillion and net profit of ₩39.7756 trillion. Operating profit runs at roughly 80% of revenue — a margin that would have sounded absurd for a memory maker a few years ago.

The step up from Q2 is just as striking. Second-quarter revenue was ₩79.3187 trillion with ₩60.5426 trillion in operating profit, so Q3 came in about 32% higher on revenue and 38% higher on profit in a single quarter. SK hynix described the contract structure as reflecting market conditions and keeping profitability stable, and said it expects demand to exceed supply through 2030, citing leading-edge process difficulty and long fab construction lead times.

On HBM4 specifically, shipments began in earnest in Q3. The company expects HBM4 to pass HBM3E in shipment volume and revenue in Q4, and it plans to run the M15X facility at maximum capacity by end of 2026. It is also accelerating a ₩40 trillion share buyback and cancellation program.

Is it officially an HBM4 shortage?

Not in those words. SK hynix confirms a broad supply squeeze through 2030 but never uses the phrase "HBM4 shortage." Nvidia's Nov. 5 commentary called HBM contested, said supply was tightening and prices rising, but quantified nothing. Samsung and Micron haven't confirmed a shortage either.

What reporting adds is a possible link between supply and spec. Nov. 2 reports said suppliers including Samsung and SK hynix struggled to hit Nvidia's earlier 22 TB/s target. Nvidia hasn't confirmed that. Other cited constraints include HBM4 validation, advanced packaging throughput and substrate warpage. Reports of volume Rubin shipments slipping into 2027 also remain unconfirmed — Nvidia describes Vera Rubin as ramping at partners. The accurate label today is a supply squeeze with real contract and pricing effects, not a proven shortage. No supplier or Nvidia has published an HBM4 price.

What did the Rubin spec revision change?

Nvidia's Nov. 5 developer blog states Rubin carries 288 GB of HBM4 per GPU — held from GB300 — at 19.2 TB/s. NVLink 6 dropped from 3.6 TB/s to 3.0 TB/s per GPU, roughly a 17% cut. The per-GPU bandwidth trim works out to about 13%.

The Vera Rubin NVL72 page lists 72 Rubin GPUs, 36 Vera CPUs, 20.7 TB of HBM4 and 1,400 TB/s aggregate rack bandwidth. Nvidia's own pages still disagree in places: an HGX Rubin NVL8 page reportedly continues to list 22 TB/s per GPU and 130 TB/s across eight GPUs.

The per-stack arithmetic is telling. SK hynix showed Nvidia's Vera Rubin superchip with 12-layer, 36 GB HBM4 on Nov. 11 at the TSMC OIP conference. Eight stacks at 288 GB means 19.2 TB/s divided by eight, or 2.4 TB/s per stack versus 2.75 TB/s under the old figure. Rumored per-pin rates fell from roughly 10.75 Gb/s to 9.375 Gb/s. Samsung advertises 11.7 Gbps scalable to 13, and Micron claims over 2.8 TB/s per stack at above 11 Gb/s per pin. Both sit above Rubin's current spec, suggesting the revision reflects system-level validation, power or packaging margins rather than a supplier ceiling.

Where do the three suppliers stand?

• SK hynix: Q3 shipments underway, Q4 crossover guided, 12-layer 48 GB HBM4E shown Nov. 11, plus 96 GB and 192 GB SOCAMM2 parts.
• Samsung: HBM4 in mass production since Mar. 17, sales "scaled up significantly" per Oct. 29 report; Chosun Biz pegs HBM4 at roughly 60% of its HBM sales (unconfirmed). It shipped 12-layer HBM4E samples June 1, with HBM4E specced at 16 Gbps and 4.0 TB/s.
• Micron: High-volume HBM4 production per its Nov. 12 page. Fiscal Q3 (ended May 28) revenue hit $41.46 billion versus $23.86 billion the prior quarter and $9.30 billion a year earlier, with $28.24 billion GAAP net income. HBM4E volume production is targeted for calendar 2027.

None of the three has published an HBM4 price, unit count, or Rubin share.

Who holds the leverage now?

Nvidia announced Jensen Huang's resignation as president, CEO and board member on Nov. 12, effective immediately. CFO Colette Kress became interim CEO, and the search for a permanent chief began that day, following an independent investigation announced days earlier.

For practitioners, the leverage math is straightforward. A buyer facing a supplier with over half its 2027-28 capacity contracted — and demand outrunning supply to 2030 — negotiates from weakness. Interim CEOs rarely rewrite large supply contracts, so suppliers likely keep their terms near term. A permanent CEO could change that.

Buyers should ask for exact SKUs and spec-sheet dates, since Nvidia's NVL72 and HGX NVL8 pages show different bandwidth figures. Memory-bound inference operators should model the 13% cut directly rather than assume compute-bound behavior. And anyone planning consumer or edge hardware should budget for firm memory pricing through 2027 — the forecasts disagree on severity, not direction. Apple's Mac margins have already shown memory costs biting, and the same DRAM supply chain feeds everything from data centers to laptops.