Korean trade daily The Electronic Times reported on September 4, 2026, that Micron Technology is expanding High Bandwidth Memory capacity toward roughly 100,000 wafers per month by year-end — about double last year's industry-estimated run rate — with a sharper mix of 12-high HBM4 aimed at NVIDIA's Vera Rubin accelerators. DIGITIMES carried a similar capacity-doubling readout the same day. Neither piece is a fresh Micron press release; both lean on industry sources and earlier company commentary.
The timing matters because HBM remains the scarce layer in AI racks. Micron is still the clear third supplier behind SK hynix and Samsung, and capacity — not another thought-leadership essay about “memory bottlenecks” — is what changes who can actually ship Rubin-class systems.
Confirmed
- ETNews (Sep 4) cites industry sources saying Micron plans to add up to about 60,000 wafers/month of HBM capacity by end-2026, taking total HBM wafer input to roughly 100,000 wpm, versus roughly 40,000–50,000 wpm last year.
- The same reporting says 12-high HBM4 was only about 20–30% of Micron HBM output early this year and could approach ~50% by year-end, with mass production of 12-high HBM4 described as starting in Q2 2026 for Vera Rubin.
- On Micron's fiscal Q3 2026 earnings call in June, CEO Sanjay Mehrotra said the 12-high HBM4 ramp was progressing about twice as fast as 12-high HBM3E and that cumulative HBM4 shipment revenue had surpassed $1 billion — company-stated figures, not third-party audits.
- ETNews attributes Counterpoint Research's Q2 2026 HBM share snapshot as SK hynix ~50%, Samsung ~32%, Micron ~18%, and estimates Samsung/SK hynix HBM capacity in the 150,000–200,000 wpm range each.
- Micron's March 2026 GTC announcement already placed HBM4 36GB 12H in high-volume production for Vera Rubin; the September stories are about capacity scale, not a first product reveal.
Unknown
- Micron has not, in these September reports, published a matching IR note confirming the exact 100,000 wpm target or the HBM4 mix percentages — treat wafer and share figures as industry estimates until the company restates them.
- How much of the added capacity is already on the floor versus still in equipment POs for Taiwan and Singapore packaging sites is not independently verified in the coverage.
- Whether NVIDIA's Rubin HBM allocation shifts share away from SK hynix/Samsung in lockstep with Micron's ramp remains a forecast, not a disclosed contract schedule.
Our take
This is the useful version of the “memory is the bottleneck” meme: a third-place supplier trying to cut the capacity gap in half before year-end. If the ETNews numbers hold, Micron stops being a rounding error on Rubin boards. If they are aspirational PO chatter, the story is still demand — equipment vendors getting HBM tool orders — not another Micron-sponsored essay about storage strategy.
Series: 1. Nvidia Commits Up to $3 Billion to Lancium, Powering the Stargate AI Campus in Texas · 2. Micron aims to roughly double HBM capacity by end-2026 as HBM4 mix rises for Vera Rubin · NVIDIA AI Factory
