中文 EN
Login Register
N
NVIDIA 2026-06-12
Technology Integration Impact: Major Conf: 92%

NVIDIA and SK Hynix Lock Down HBM4/5 Roadmap, Cementing Vera Rubin Supply Chain

Summary

NVIDIA and SK Hynix sign a multi-year agreement to co-define HBM4 production and HBM5 pre-research for Vera Rubin GPUs. Samsung also enters HBM4 supply as a second source. The deal elevates SK Hynix from vendor to co-developer, potentially creating a de facto memory standard barrier that marginalizes Micron and others.

Key Takeaways

In June 2026, NVIDIA and SK Hynix signed a multi-year technology partnership to co-develop next-gen HBM. The deal covers HBM4 production for Vera Rubin (single-stack bandwidth >1.6TB/s, 33% uplift from HBM3e's ~1.2TB/s) and HBM5 pre-research (expected post-2028 with Hybrid Bonding and 3D logic die stacking). SK Hynix holds ~50-55% HBM market share as the primary HBM3e supplier for Blackwell/Ultra. Samsung has started shipping HBM4 samples as a second source for Vera Rubin, with its 2nm GAA foundry orders up 130% YoY. Micron, though in Blackwell's HBM3e supply chain, risks marginalization in HBM4-5 generations. This agreement elevates SK Hynix from vendor to co-developer, effectively defining the memory standard roadmaps for next-gen AI chips, creating a de facto industry barrier.

Why It Matters

This deal is a strategic encirclement by NVIDIA:

  • Defense against AMD/Intel: By co-defining HBM4/5 interfaces exclusively with SK Hynix, NVIDIA ensures optimal memory bandwidth for Vera Rubin while raising compatibility costs for rivals (who must redesign memory controllers if they adopt different HBM standards).
  • User asset lock-in: Enterprises deploying Vera Rubin clusters become dependent on SK Hynix's HBM4/5 supply for batch size and time-to-first-token performance, locking them into the NVIDIA+SK Hynix roadmap and hindering migration to AMD MI400 or other architectures.
  • Hidden physical limits and cost traps: The press release omits HBM4's power/thermal challenges—8-12 stacks per GPU could push total GPU power beyond 1500W, escalating liquid cooling costs. HBM5's Hybrid Bonding yield is unproven, risking delays. SK Hynix's enhanced pricing power as co-developer will likely inflate GPU TCO.

PRO Decision

Vendors (AMD, Intel, Micron): Immediately push for open HBM standards via JEDEC to avoid NVIDIA+SK Hynix lock-in. Accelerate development of multi-vendor-compatible memory controllers. Micron must lead in Hybrid Bonding HBM4 and form alliances with AMD/Intel.
Enterprises: CIOs/architects must audit Vera Rubin procurement for multi-supplier risk—demand NVIDIA guarantee at least two HBM4 sources (Samsung is in, but ensure interchangeability). Assess liquid cooling costs and budget for HBM5 maturity risks. Adopt cross-GPU AI frameworks (e.g., PyTorch/XLA) to preserve migration options to AMD/Intel.
Investors: Short-term bullish for SK Hynix (pricing power) but Samsung's entry dilutes monopoly. Long-term, watch Micron's HBM4 Hybrid Bonding breakthrough. NVIDIA's moat strengthens, but supplier concentration risk rises—any SK Hynix production hiccup could cripple Vera Rubin deliveries.

Source: SK hynix Newsroom / Silicon Semiconductor
View Original →

Get 3-5 key AI infrastructure signals weekly →

💬 Comments (0)