Micron Aims to Nearly Double HBM4 Wafer Capacity by the End of 2026

What Micron actually confirmed, on March 16

At Nvidia's GTC event, booth 1407, Micron said in a March 16, 2026 press release that its HBM4 36GB 12H stack, built for Nvidia's Vera Rubin platform, had reached high-volume production alongside a PCIe Gen6 SSD and a 192GB SOCAMM2 module. The company listed HBM4 bandwidth above 2.8 terabytes per second and pin speeds over 11 gigabits per second, a 2.3-times bandwidth jump over the prior HBM3E generation alongside a power-efficiency gain above 20%. Sumit Sadana, Micron's executive vice president and chief business officer, said the next era of AI "will be defined by tightly integrated platforms developed through joint engineering innovations across the ecosystem." The same release disclosed early samples of a denser HBM4 48GB 16H stack, a 33% capacity increase over the shipping part.

The capacity number is a report, not a confirmation

What Micron has not confirmed is how much of that memory it can actually make. TrendForce reported on September 4, 2026, citing Electronic Times, that Micron plans to add up to 60,000 HBM wafers a month, taking total capacity from roughly 40,000 to 50,000 wafers today to around 100,000 by year-end. A separate write-up from German outlet igorslab.de, published two days later on September 6, relayed similar figures and added a detail TrendForce's piece left out: the 100,000-wafer figure has not yet been officially confirmed by Micron itself.

Why wafer count, not chip design, is the constraint

HBM4 bandwidth numbers get most of the attention in vendor announcements, but the number that actually rations supply to customers is wafers per month, not gigabits per second. A single accelerator package can carry multiple HBM4 stacks, and Nvidia's Vera Rubin platform is expected to use more stacks per GPU than its Blackwell predecessor, which means the same generational jump that improves bandwidth per chip also raises the total wafer demand per accelerator shipped. Micron confirmed the bandwidth improvement in March; it has not confirmed how many Vera Rubin units its current wafer output can actually supply, which is the gap the September capacity reports are trying to fill from the outside.

HBM4 is also set to take over Micron's product mix

Both reports put 12-Hi HBM4 at 20% to 33% of Micron's HBM output in early 2026, rising to as much as 50% by year-end as the company shifts capacity away from HBM3E. That shift matters because Samsung and SK Hynix's own memory inventory has already fallen below 10 days of supply, leaving Micron room to sell whatever HBM4 it can build without having to discount it against a glut.

Company	Monthly HBM wafer starts	Q2 2026 HBM revenue share
SK Hynix	150,000-200,000	50%
Samsung	150,000-200,000	32%
Micron	40,000-50,000 (targeting 100,000)	18%

1. Current, low estimate40K wafers/month
2. Current, high estimate50K wafers/month
3. Year-end 2026 target100K wafers/month

Source: TrendForce, citing Electronic Times, Sept. 4, 2026; igorslab.de, Sept. 6, 2026

Samsung and SK Hynix still hold most of the market

Even at 100,000 wafers, Micron would sit at roughly half of Samsung's or SK Hynix's individual output. igorslab.de put each of those two companies' HBM wafer starts at 150,000 to 200,000 a month, and their second-quarter 2026 HBM revenue share at 50% for SK Hynix, 32% for Samsung, and 18% for Micron. Micron's push to double capacity narrows that gap; it does not close it. Samsung and SK Hynix have separately rejected an $18 billion prepayment request from Korea's state utility, KEPCO, a sign that even the two leaders are managing their own expansion spending carefully rather than accepting outside capital on offer.

The gap also shows up in what each company can promise a customer today. Micron's own confirmed number, greater than 2.8 terabytes per second of bandwidth per HBM4 stack, applies however many stacks a buyer can get; the constraint Nvidia's Vera Rubin customers actually face is Micron's allocation, not Micron's chip design. A 12-Hi HBM4 stack from Samsung or SK Hynix carries comparable bandwidth, which means the wafer-count gap, not the bandwidth figures, is what determines how much of the Vera Rubin supply chain Micron can claim.

What a doubled Micron would mean for pricing

TrendForce said in a June 2, 2026 press release that suppliers would have "strong justification for raising HBM prices" heading into 2027, without naming a specific figure for the increase. A Micron that reaches 100,000 wafers a month by year-end would not flip that dynamic on its own; even doubled, its output stays below either Samsung's or SK Hynix's current run rate, and all three companies are selling into demand from Nvidia, AMD and the hyperscalers building their own accelerators at the same time. The more likely effect of a successful Micron ramp is a third qualified supplier for buyers who depend on two today, which matters for a buyer's negotiating position even where it does not change the headline price.

What would confirm the number

Micron's own public statements to date cover product specifications and its Vera Rubin partnership, not manufacturing capacity targets. Until Micron itself discloses a wafer count on an earnings call or in a filing, the 100,000 figure rests on Electronic Times' industry sourcing, repeated independently by TrendForce and igorslab.de but not yet matched by an on-the-record number from the company that would have to build it.