Nyaws
🗼 Tokyo--:----°
#025 / 2026-05-14 TECH

HBM4 shipments ramp — physical layer for Rubin launch is ready

🗓 Published 2026-05-14 06:30 JST / 🧠 HumanAI (COOL lead) / ~2,000 chars

Mass production and shipment of HBM4 (4th-generation high-bandwidth memory) — the physical layer for NVIDIA's next-generation architecture Rubin, slated for the second half of 2026 — is ramping up. SK Hynix, Samsung, and Micron each signaled progress at recent industry conferences. For Rubin to deliver its target of 50 PFLOPS of inference at NVFP4 precision on real silicon, stable HBM4 mass production is a prerequisite. Nyaws 100 paper trading now has the AI axis (NVDA / AMD / AVGO / ASML) back as Top-1, and on the supply-chain side, "Rubin-generation structural readiness" is also visible. The synchronized timing is drawing market attention.

HBM4
Next-gen high-bandwidth memory — physical layer for 2026H2 Rubin
+50% bandwidth and 1.5× capacity vs HBM3. SK Hynix, Samsung, and Micron have set up mass production

HBM4 Is the Physical Prerequisite for the Rubin Launch

To achieve Rubin's claimed 50 PFLOPS at NVFP4 precision on real silicon, on-die SRAM and cache alone are nowhere near enough. Trained model weights, KV-cache during inference, multi-batch concurrent processing — all are sustained by HBM4's ultra-wide bandwidth (+50% bandwidth, 1.5× capacity). Rubin's public specs explicitly list "HBM4 adoption," meaning "if HBM4 mass production slips, Rubin slips too." Conversely, once HBM4 shipments are confirmed to ramp, uncertainty about Rubin's 2026H2 schedule drops substantially.

HBM4 — key specs vs predecessors

Generation BW (GB/s/stack) Capacity/stack GPU gen
HBM381924 GBHopper / Blackwell
HBM3E1,22836 GBBlackwell / Blackwell Ultra
HBM4~2,04848–64 GBRubin (2026H2)

Progress at Each of the 3 Suppliers

HBM4 supply is effectively an oligopoly of SK Hynix, Samsung, and Micron. At recent industry conferences: (1) SK Hynix specified a large initial contract with NVIDIA, (2) Samsung announced mass-production readiness through yield improvement, (3) Micron confirmed HBM4 line operation at its Idaho fab in the US. All three are now positioned to "meet Rubin's launch window." For NVIDIA, this means pricing leverage through supplier competition, which also supports the AI axis (NVDA) margin defense.

ASML and TSMC — The Other Physical Bottleneck