Samsung starts mass production, ships HBM4 memory chips

Samsung has announced the start of mass production and shipment of its new HBM4 memory modules, claiming industry-first status.

The HBM4 chips, built on a 4nm logic process, deliver a consistent transfer speed of 11.7 Gbps, with potential to reach 13 Gbps.

The modules offer a maximum bandwidth of 3.3 TB/s per stack, a 2.7x increase over HBM3E, and are available in capacities from 24 GB to 36 GB, with plans to expand to 48 GB.

Samsung reports a 40% improvement in power efficiency and enhanced thermal management compared to previous generations.

The company plans to expand HBM4 production capacity significantly in 2026, with sample shipments of HBM4E starting later this year and custom HBM samples in 2027.

The development aims to support AI, data centers, and high-performance computing markets amid increasing demand for advanced memory solutions.

🔗 Source: Samsung

🧠 Food for thought

Implications, context, and why it matters.

Samsung’s HBM4 launch aims to win back Nvidia supply

Samsung’s announcement marks a return to top-tier memory, with customers reportedly saying “Samsung is back” 1.
Reports say Samsung has cleared, or is close to clearing, qualification and verification for its HBM4 (High Bandwidth Memory 4) for Nvidia’s upcoming Rubin AI platform (Nvidia’s next-generation AI accelerator platform), which could put Samsung back on the supplier list 2.
The timing matters since SK hynix’s operating profit beat Samsung’s for the first time in 2025, driven by SK hynix’s lead in HBM for AI chips 3.
The supplier pool could tighten further since SemiAnalysis (a semiconductor research and analysis firm), quoted by Wccftech, says Micron’s share of Nvidia Rubin HBM drops to zero and it sees no sign of Nvidia ordering Micron HBM amid reported validation issues 4.

More memory supply may not ease the AI hardware crunch

Samsung’s production scale-up gives AI chipmakers like Nvidia another HBM source, in a market where SK hynix has led 3.
Extra supply alone may not lower prices since one industry estimate puts HBM4 pricing more than 50% above HBM3E due to higher production costs, including about a 30% jump tied to outsourced base-die manufacturing, meaning a supplier makes the foundational logic layer that the memory stack is built on 5.
Output also runs into limits from specialized “TSV capacity” (through-silicon via capacity), a choke point in vertical chip stacking that takes more steps than standard memory production 6.
Those constraints feed into next-generation accelerator pricing since SemiAnalysis writes that Nvidia’s roadmap includes Rubin, with HBM content rising over time 6.