Nvidia Asks Samsung and SK Hynix to Boost Supply of 8-Layer HBM4 Configuration
Nashnova编辑部
Nvidia has asked Samsung and SK Hynix to raise the share of 8-high HBM4 shipments in H2 2026 while cutting the 12-high share — not a demand cut, but a rebalancing of heat, yield, and capacity trade-offs.
Why does Nvidia want more 8-high and less 12-high?
The core issue is heat and yield. HBM4 doubles input-output connections to 2,048, sharply raising thermal stress at both chip and rack level.
This means → more layers concentrate more heat, making bonding and packaging yields harder to control — even though individual DRAM die yields are already stable.
In plain terms = Nvidia still wants large memory capacity, but stacking too high drives down the usable ratio and strains cooling. Adjusting the mix keeps the Vera Rubin accelerator platform on firmer ground.
What does the memory market look like right now?
Tight. Samsung said in its Q2 earnings that it expects server DRAM, enterprise SSD, and HBM demand to accelerate in H2 2026, with the market staying undersupplied.
SK Hynix disclosed that additional supply requests from major tech clients keep rising; it has signed long-term agreements with roughly ten core customers to lock in supply.
This reflects an infrastructure investment cycle that is far from peaking — memory makers remain firmly in a seller's market.
How much further can prices rise?
Money Today reported that Samsung's negotiations with clients on 2027 HBM volumes and pricing are nearing completion.
Goldman Sachs estimates 2027 HBM average selling prices will rise 44% year-on-year to roughly $17 per gigabyte, putting total market size at about $116.3 billion.
This means → even with Nvidia reshuffling the layer mix, HBM's volume-and-price thesis is intact — demand is growing, and pricing follows.
What does this mean for Samsung's foundry business?
Samsung's HBM4 pairs advanced DRAM with a 4 nm logic base die — the bottom chip that connects and controls the memory stack — folding memory growth directly into its advanced-logic manufacturing footprint.
Samsung said Q2 foundry profitability improved significantly, driven by HBM base-die demand and U.S. client orders. It plans to expand 4 nm LPU and base-die sales in H2, targeting double-digit revenue growth.
In plain terms = the more HBM Samsung sells, the fuller its foundry order book gets — memory and foundry are now tied together on HBM4.
What does this signal for the broader supply chain?
SK Hynix noted that evolving AI workloads are shifting the competitive focus in memory from single-chip specs to system architecture and packaging.
This means → Nvidia's tilt toward 8-high is not a downgrade; it is supply-chain optimization across the entire compute platform — matching memory capacity precisely to system-level needs.
This reflects a shift in HBM4's growth logic: from pure volume expansion to finer-grained product-mix management. For memory makers, the ability to flexibly adjust the mix for each client is becoming the key to winning orders.
市场有风险，内容仅供研究参考，不构成投资建议。