Nvidia qualified HBM4 from Samsung, SK Hynix, and Micron for Vera Rubin, moving those parts into full production. SK Hynix secured 60–70% of the initial allocation, Samsung 25–30%, and Micron the remainder.
Key facts
- Nvidia qualified HBM4 from Samsung, SK Hynix, and Micron.
- SK Hynix secured 60-70% of Vera Rubin HBM4 allocation.
- Samsung received 25-30% of the allocation.
- Micron took the remainder.
- HBM4 qualification is a gating factor for Vera Rubin's 2026 ramp.
Nvidia has cleared a critical supply-chain bottleneck for its next-generation Vera Rubin architecture by qualifying HBM4 memory from all three major DRAM suppliers. According to @rohanpaul_ai, Samsung, SK Hynix, and Micron have all passed qualification and those parts have been moved into full production.
HBM4 qualification is the single largest gating factor for Vera Rubin's 2026 ramp. Each Vera Rubin rack will demand roughly 1.5–2 TB of high-bandwidth memory, and HBM4's per-stack bandwidth is expected to exceed 2 TB/s — roughly double HBM3e. Without all three suppliers producing qualified parts, Nvidia risked a repeat of the HBM3 supply crunch that constrained Hopper shipments through early 2024.
The allocation split reveals Nvidia's supply-chain strategy. SK Hynix, which led HBM3 and HBM3e production, took the largest share at 60–70%. Samsung received 25–30%, a smaller slice than its broader DRAM market share would suggest, likely reflecting its later entry into HBM3e qualification. Micron took the remainder, maintaining its role as a third-source hedge.
This is not a simple procurement update — it is a signal that Vera Rubin's production timeline is on track. Nvidia typically qualifies memory 12–18 months before volume shipments. With qualification now closed and production released, the earliest Vera Rubin systems could ship in late 2026, consistent with Nvidia's stated 2026 roadmap.
What the allocation doesn't tell us is pricing or yield specifics. None of the three suppliers disclosed HBM4 bit-cost or qualification yield rates. Samsung's smaller share suggests it may still be ramping HBM4 yields, while SK Hynix's dominant position reflects its head start in 12-stack HBM4 prototypes shown at ISSCC 2025.
For inference operators, the qualification is good news: three qualified suppliers means better pricing leverage and less risk of a single-vendor production outage. For AI chip competitors (AMD, Intel, startups), it means Nvidia's memory supply for Vera Rubin is effectively derisked — a competitive disadvantage that narrows the window for catching up on the memory front.
What to watch
Watch for Vera Rubin's first system-level benchmarks at Nvidia GTC 2027, expected in March 2027. Also track Samsung's HBM4 yield disclosures in its Q1 2027 earnings call — if yields approach SK Hynix's, the allocation split may shift for the follow-on Rubin Next architecture.
