Every figure in this article comes from our HBM Qualification Tracker — a sourced, dated, confidence-scored timeline of who passed whose qualification, maintained daily and reviewed by a human before publication. Click any relationship there for the underlying source on each milestone.
Why qualification — not capacity — decides HBM share
High-bandwidth memory is not a commodity you win with fab capacity announcements. Before a single stack ships for revenue, it must pass the customer's qualification: months of thermal, power, signal-integrity, and reliability validation against that specific accelerator platform. Fail, and your roadmap slides a full product cycle. Pass first, and you ship at premium prices into a sold-out market while competitors re-spin.
That is why the qualification event — not the press-release roadmap — is the single highest-signal datapoint in the HBM market. The 2022–2026 record makes the point brutally well.
Chapter 1 (2022–2023): SK Hynix runs alone
SK Hynix supplied the industry's first HBM3 to NVIDIA, with mass production beginning June 2022 — the memory inside the H100, the chip that defined the AI buildout [1]. For roughly two years, no competitor passed an equivalent NVIDIA qual at the leading edge. Samsung's HBM3 did pass NVIDIA qualification in July 2024, but reportedly only for the China-market H20. First-mover qualification, more than any other factor, is how SK Hynix built its ~50% revenue share — a dynamic you can see quantified in our HBM Market Analysis.
Chapter 2 (2024–2025): the HBM3E split — Samsung's saga, Micron's sprint
HBM3E is where qualification timing visibly redistributed share.
SK Hynix extended its lead: industry-first HBM3E volume production in March 2024, supplying the H200 [2].
Micron, the late entrant, qualified fast and quietly: HBM3E 24GB 8-Hi in volume production for the H200 by February 2024 [5], production-capable 12-Hi stacks shipped September 2024, NVIDIA certification for the GB300's 12-layer HBM3E by April 2025, and design-ins on AMD's MI350 with mass production from Q3 2025.
Samsung hit the wall. In April–May 2024, its 8- and 12-layer HBM3E reportedly failed NVIDIA's heat and power requirements [3]. A third validation attempt reportedly stumbled again in June 2025, with a retest set for September. Samsung finally passed NVIDIA's 12-layer HBM3E qualification in September 2025 — third supplier, after SK Hynix and Micron [4] — and reached volume shipments in October 2025, with 2026 supply reportedly sold out. The ~18 months between first failure and pass spanned essentially the entire Blackwell cycle.
Two patterns worth noting from the timeline. First, qualification is customer-specific: while Samsung was failing at NVIDIA, it was qualifying 12-Hi HBM3E at AMD for the MI350 (June 2025) and clearing Broadcom's 8-Hi evaluation — a reminder that "Samsung HBM3E status" is not one fact but a matrix. Second, failures rarely come with press releases; they surface through Korean trade press and supply-chain reporting, get denied, and are later confirmed by who actually ships. That is exactly why we log every milestone with its source and a confidence score.
Chapter 3 (2025–2026): HBM4 — the first three-supplier generation
The HBM4 race for NVIDIA's Vera Rubin compressed the qualification cycle dramatically:
- June 2025 — Micron ships HBM4 36GB 12-Hi samples to key customers, first to sample publicly [6]
- October–December 2025 — SK Hynix and Samsung deliver paid HBM4 samples to NVIDIA; certification underway
- March 2026 — Samsung and SK Hynix win Vera Rubin HBM4 slots [7]; Micron enters high-volume HBM4 production
- June 2026 — Jensen Huang confirms all three vendors qualified and in production for Vera Rubin, with Q3 shipments [8]
For the first time in HBM's history, a generation starts its ramp with all three suppliers qualified. That changes the competitive question. Through HBM3/3E, the question was binary: can you ship at all? From HBM4 onward it becomes economic: who ships at the best yield, price, and allocation? — which is precisely where qualification data stops being the whole story and pricing/share data picks up.
What four years of qualification events teach
- Sampling-to-volume runs roughly 9–15 months when it goes well (Micron HBM4: samples June 2025 → high-volume March 2026), and 2+ years when it doesn't (Samsung HBM3E).
- One vendor's failure is another's pricing power. Samsung's 2024 failure left SK Hynix and Micron splitting Blackwell-era demand in a shortage.
- The matrix beats the headline. Vendor × customer × generation × stack-height is the correct granularity; anything less collapses contradictory facts into noise.
- Provenance is the product. Every status above links to a primary source, a named executive, or attributed trade reporting — and is dated to the month. When statuses are contested (as Samsung's were for 18 months), the source trail is the only way to know what was actually true, when.
Track it live
The full matrix and per-relationship timelines — 16 relationships, 25+ sourced events and growing daily — are free on the HBM Qualification Tracker. For programmatic access, hit /api/v1/hbm-qualification or call the get_hbm_qualification MCP tool from any agent. When the next qualification breaks, it lands there first — sourced, dated, and human-reviewed.