Samsung's HBM4 Yields Reportedly Reach 80%, but the Number Is Single-Sourced

Six months after Samsung Electronics became the first company in the world to put HBM4 into mass production, Korean media reported this month that the yield on its sixth-generation high-bandwidth memory has climbed to roughly 80% — up from below 60% when the line started in February. If the figure holds, it removes the single biggest reason Nvidia, AMD and the hyperscalers have spent three years buying most of their AI memory from one company.

The number originated with Seoul Economic Daily, which published it Aug. 9 citing an unnamed industry source. Samsung has not confirmed it. TrendForce, DigiTimes, Infostock Daily and ETNews picked it up over the following days, and it reached most English-language readers through those aggregators rather than independent reporting. Treat it accordingly: a single-sourced Korean trade report about a metric no memory maker discloses, arriving in a week when Samsung had every incentive for it to circulate. What can be verified independently is the direction of travel, and that is not in dispute.

What the reporting actually says

Samsung had internally targeted 80% for the end of 2026, according to Seoul Economic Daily's source; process improvements ran ahead of plan. The industry treats 80% as the "golden yield," the threshold at which defect rates fall low enough to make a part reliably profitable. Infostock Daily credits progress on thermal compression non-conductive film (TC-NCF), the bonding process long regarded as Samsung's structural weakness against SK hynix's mass reflow molded underfill (MR-MUF) approach. ETNews points instead to the base layer: Samsung pushed yields on its 1c DRAM — the 10nm-class node underneath HBM4 — above 80% earlier this year.

The corroborating disclosures are firmer. Kim Jae-jun, executive vice president of Samsung's memory business, told analysts on the second-quarter earnings call in late July that third-quarter HBM4 revenue would more than triple sequentially and account for more than 60% of total HBM revenue in the second half. Song Jai-hyuk, Samsung's president and chief technology officer, said at an internal briefing in late June that reliability-test yields on seventh-generation HBM4E had passed 70%. Samsung is targeting roughly 38% of the HBM market by year-end — the level of its DRAM share.

The framing of Samsung "catching up" overstates the gap. Industry insiders quoted by Infostock Daily put SK hynix's own HBM4 yield in the 80% range. Kim Ki-tae, the executive vice president who runs HBM sales and marketing, told SK hynix's second-quarter call that HBM4 competitiveness is complete only when it includes the ability to supply in large volumes with stable yields and quality — and that current mass-production yields are close to those of the mature HBM3E generation. (Both Korean executives' remarks are translated.) UBS models SK hynix holding first place this year with 48% of HBM bit shipments, then losing it narrowly in 2027: Samsung 41%, SK hynix 39%.

Qualification is settled on paper. Nvidia CEO Jensen Huang confirmed on June 5 that Samsung, SK hynix and Micron had all passed certification for HBM4 on the Vera Rubin platform, which entered full production days after its GTC Taipei announcement on June 1.

"Agentic AI is a new kind of workload. One prompt can launch a thousand-step journey of reasoning, retrieval, tool use and response generation," Huang said in the statement tied to Vera Rubin's launch. "Vera Rubin was built for this moment — an AI factory engine that delivers intelligence at scale."

Certification is not allocation. Supply-chain analysts put SK hynix at roughly 60–70% of Vera Rubin HBM4 volume, Samsung at 25–30% and Micron with the remainder. On the AMD side Samsung is further ahead: it signed an HBM4 and DDR5 agreement with AMD in March and is priority supplier for the Instinct MI455X.

Analysis: memory, not logic, is the constraint

A yield number moves markets because HBM — not TSMC wafer starts, not CoWoS — is now the binding constraint on AI accelerator supply. Silicon Analysts prices HBM4 at roughly $550 per 36GB 12-high stack against about $300 for HBM3E, a jump from $8.33/GB to $15.28/GB, and estimates HBM at 30–40% of an accelerator's manufacturing cost, up from under 20% two generations ago. On a B200, eight HBM3E stacks cost about $3,250 against roughly $850 for the logic die.

TrendForce expects HBM bit shipments to grow 50–60% year over year in 2027 and still fall short of demand, leaving suppliers with pricing power throughout the year. HBM wafer input across the top three suppliers is projected to rise from about 18% of total DRAM wafer input at the end of 2025 to 30% by the end of 2027 — capacity taken out of conventional DRAM.

The consequence is visible in product specs. Since the start of the third quarter, Nvidia has been evaluating cheaper HBM configurations for Rubin Ultra, expanding beyond the original 12-high HBM4E baseline to include 8-high HBM4E, 12-high HBM4 and 8-high HBM4, per TrendForce. Whether Rubin Ultra reaches 14–16 Gbps I/O or settles for 11–12 Gbps depends on memory validation, not GPU design.

That is what a credible Samsung second source is worth — and it is not lower prices. Nothing loosens through 2027, and large-scale new fab supply is limited until 2028 given the three-plus years from groundbreaking to wafer output. What it buys is negotiating position: an alternative bidder in annual contract talks, and insurance against one supplier's disruption. SK hynix's leverage is real — it booked roughly 17 trillion won from Nvidia in the first half, about 13% of sales — but not unconditional. Its wage talks collapsed at a fifth round over a proposal to pay more than half of this year's profit-sharing in locked-up treasury shares, and a new union launched Aug. 14 with 2,400 members.

What to watch

Samsung's third-quarter results are the first hard test: they will show whether HBM4 revenue actually tripled. Fourth-quarter shipment volumes from both Korean makers will settle whether capacity or quality decides the generation. And Nvidia's final Rubin Ultra memory specification remains the clearest signal of how tight 2027 supply really is. Foreign investors dumped 5.9 trillion won of Korean chip stocks in the first week of August and were buying again by mid-month. Neither move confirms an unverified number.