Samsung Has Yet to Secure Mass Production Orders for NVIDIA's HBM4
nashnova research
Industry sources say Samsung has received no formal mass-production orders from Nvidia for HBM4 — its revenue is limited to paid evaluation samples. This directly contradicts both Samsung's and Jensen Huang's public claims of ongoing production, suggesting Samsung's actual position in the HBM supply chain may be far weaker than official statements imply.
"Mass-production sales" vs. "mass-production orders" — what's the gap?
Samsung's Q1 earnings in April said it had begun HBM4 mass-production sales for Nvidia's Vera Rubin platform. Jensen Huang said in June that Samsung was "in production." But industry sources say Samsung has not received a formal mass-production order — its revenue comes only from paid evaluation samples.
This means → what Samsung calls "mass-production sales" may just be getting paid for samples, not shipping at scale. "In production" and "receiving bulk purchase orders" may be separated by a threshold Samsung has not yet crossed.
In plain terms = Samsung says "we're selling," Nvidia says "they're producing," but insiders say "no order has been placed." The three narratives don't align — actual progress likely trails the public statements.
How is the share split? How much can Samsung get?
Nvidia reportedly discussed allocating Vera Rubin HBM4 supply roughly 60% to SK Hynix, 30% to Samsung, and 10% to Micron. But these were verbal discussions before any formal purchase order — the final split can still change.
A more critical variable: Nvidia has cut its overall HBM4 procurement plan, shifting part of its memory budget toward server products including low-power DRAM and SOCAMM.
This means → the pie itself is shrinking. Even if Samsung holds the 30% verbal share, actual order volume could fall below expectations. And if it hasn't even received a formal order yet, that 30% is just a number on paper.
Is Samsung's speed advantage enough?
Samsung's HBM4 runs at a stable pin speed of 11.7 Gbps, peaking at 13 Gbps — a 22% improvement over the top pin speed of the previous HBM3E generation. It uses Samsung's 6th-gen 1c DRAM process, with a base die built on a 4nm logic node.
Pin speed was a key comparison point when Nvidia discussed specs with HBM4 suppliers. Initial supplier proposals were around 9 Gbps; Samsung pushed a faster design — seen as a signal it was closing the gap after falling behind in the HBM3/HBM3E generations.
In plain terms = the speed is genuinely faster, but this is catching up, not leading. Customers don't pick suppliers on peak speed alone — they also weigh thermal performance, power consumption, yield, and system-level performance after integrating memory with AI accelerators.
How deep is SK Hynix's moat?
SK Hynix started earlier on HBM4 development and customer qualification. It delivered 12-layer HBM4 samples to key customers in March 2025 and has an established track record supplying earlier HBM generations to Nvidia.
Nvidia and SK Hynix announced a multi-year technology partnership in June to co-develop memory products for Nvidia's future AI infrastructure roadmap.
This reflects an advantage that goes beyond a technical head start — it is a deeply embedded customer relationship. Nvidia reportedly used Samsung's faster design as leverage in price negotiations with SK Hynix, but that did not change the supply relationship. Samsung's technical catch-up is, for now, more of a bargaining chip than an order guarantee.
市场有风险，内容仅供研究参考，不构成投资建议。