SILICON PENINSULA

Silicon Peninsula #004 — SK hynix ships 16-high HBM4 first; Samsung calls it unnecessary

Friday, September 18, 2026 · 5-minute read · Free edition


The tape

Seoul and New York disagreed about Korean memory on Thursday. KOSPI closed at 6,715.41, −0.04%, with Samsung Electronics at ₩252,500 (−0.39%) and SK hynix at ₩1,745,000 (−0.80%). Morning gains of more than 1% were handed back on afternoon selling in the two large-caps; foreigners sold a net ₩2.28T ($1.6B) for a seventh straight session while institutions bought ₩158.7B. (Seoul Economic Daily)

Wall Street then bid the same trade up. Chips led a rebound from Monday's AI-safety sell-off as Treasury yields eased: SOXX +3.39% to $519.10, Micron +5.50% to $977.50, AMD +6.36% to $545.09, SK hynix's Nasdaq line +4.64% to $182.99, Nvidia +2.54% to $219.34 and Broadcom +2.29% to $347.30, against QQQ +1.73% and SPY +1.13%. (SOXX, MU, AMD, SKHY, NVDA, AVGO, QQQ, SPY)

Three signals

1. SK hynix is mass-shipping 16-high HBM4, and it is the only one of the three doing it. Supply-chain tracking reported this week that SK hynix began volume shipments of a 48GB, 16-layer HBM4 part for Nvidia's Vera Rubin platform during Q3, ahead of any company announcement. Sixteen layers carry 33% more capacity than the 12-high stack. Micron has been in high-volume production of 12-high HBM4 since March and has sent 16-high samples without announcing volume; Samsung began standing up HBM4 lines in February, with its 12-high part in final qualification. Why it matters: Nvidia buys HBM by the stack, and every stack consumes one scarce advanced-packaging slot. Shipping the tallest qualified stack is the cleanest way to turn the same allocation into more revenue — which is how SK hynix converted its HBM3E position into pricing power, and why a repeat matters more than the layer count suggests. (글로벌이코노믹, Benzinga)

2. Samsung's stated position is that nobody needs 16-high yet. Samsung has said customer demand for 16-layer parts is "limited" and commercialising one "unnecessary" for now, while it finishes qualifying 12-high HBM4; its 12-high HBM3E cleared Nvidia qualification in mid-September. Why it matters: Read charitably this is a yield argument, not a capability one — 12-high at high yield can out-earn 16-high at low yield, and Samsung has been burned before by shipping into a qualification it had not stabilised. But it is a real fork, and a falsifiable one: if Vera Rubin SKUs standardise on 48GB stacks, Samsung's allocation is capped at the low-density tier of the most profitable product in memory. (Benzinga)

3. Micron is attacking on density and geography instead of stack height. Its $2.75B Sanand assembly-and-test plant in Gujarat entered commercial production on Sep 17, shipping finished DRAM and NAND worldwide with the first modules going to Dell. Two days earlier it demonstrated the first 512GB DDR5 RDIMM, TSV-stacked at up to 9,200 MT/s, putting 12TB in a 24-slot dual-socket server at over 60% lower power than four 128GB modules; volume production is targeted for 2H27. Why it matters: Korea's back end for commodity DRAM and NAND sits at home and in China, and its answer to US tariff exposure is a 2029 Indiana packaging plant. Micron now has a tariff-flexible back end running today and a module roadmap competing for the same server sockets Samsung sells into. (TechPowerUp, Unite.AI)

Korea-only radar

Items reported in Korean that we did not find in English coverage.

Number of the day

52–78 weeks — the estimated lead time on TSMC's chip-on-wafer-on-substrate packaging, against a roughly 20% supply shortfall for 2026. Capacity is seen at 120,000–130,000 units a month by late 2026 and 141,000–170,000 in 2027 — still short. This is why stack height, not wafer starts, is the memory makers' lever. (글로벌이코노믹)

Calendar


Silicon Peninsula is written from primary Korean sources every weekday morning (KST). Numbers are quoted from the cited source; where Korean outlets use won, we keep won. Founding-member Pro seats ($99/yr, first 50): reserve one here. Reply to this email with a company you want tracked. Forward it to someone who models memory prices for a living.

Get the next issue in your inbox

Korea's chip supply chain, read from the source, every weekday morning.

Weekday mornings, 07:30 KST. No spam, unsubscribe anytime.