There is an outlook that Micron, a U.S. memory semiconductor corporations, may be unable to supply sixth-generation high-bandwidth memory (HBM4) for the initial production volume of Nvidia's next-generation artificial intelligence (AI) chip "Vera Rubin." As a result, analysts say an impact on HANMI Semiconductor, which supplies Micron with TC Bonders (thermo-compression bonders) for HBM manufacturing, is inevitable. Micron is effectively expanding its production facilities aggressively on the premise of "HBM4 supply for Nvidia."
However, some in the semiconductor industry say it is not yet certain that Nvidia has fully finalized HBM4 supply volumes with the three memory companies (Samsung Electronics, SK hynix and Micron), so it is too early to conclude a "Micron early exit." There is also a view that even if Micron fails to supply HBM4 to Nvidia, the impact on HANMI Semiconductor could be limited.
According to the industry on the 11th, semiconductor analysis firm SemiAnalysis recently reported that there are no signs Nvidia will place HBM4 orders with Micron. It assessed there is a high chance Micron will be unable to supply HBM4 for about 12 months after the Vera Rubin launch. SemiAnalysis accordingly cut Micron's HBM4 share in Vera Rubin to 0%.
SemiAnalysis put SK hynix and Samsung Electronics' Vera Rubin HBM4 shares at 70% and 30%, respectively. Previously, industry analysis suggested Micron could secure about a 5%–10% share of Vera Rubin, but the landscape appears to have shifted as Nvidia recently raised its required HBM4 performance.
Nvidia recently asked the three memory companies to set HBM4 pin (data path) speeds at 11 Gbps (gigabits per second) or higher. Micron said it met this condition, but SemiAnalysis viewed it as "lagging the target specifications." HBM stacks DRAM, and the bottom layer is the "base die," which controls data flow and distributes power. Starting with HBM4, it also performs direct logic functions, making it even more important. Industry analysis says Micron struggled with base die design and failed to meet Nvidia's required performance.
Targeting an official launch in the second half of this year, Vera Rubin boasts higher performance than the AI chip now on sale, "Grace Blackwell" (GB). Its hallmark is a single configuration combining 36 "Vera" central processing units (CPUs) and 72 "Rubin" graphics processing units (GPUs). HBM4 is mounted on Rubin. Nvidia CEO Jensen Huang said at CES 2026 in Jan. that "Vera Rubin is currently in full-scale mass production."
◇ If Micron's "early exit from Nvidia supply" is true, HANMI Semiconductor would also take a hit
For HBM3E (fifth-generation HBM), Micron cleared Nvidia's quality tests about a year earlier than Samsung Electronics. However, it is assessed that performance gains were limited versus SK hynix due to production capacity constraints. Accordingly, Micron is aggressively ramping up production capacity to align with the Vera Rubin launch.
In Aug. 2024, Micron acquired AUO's Taiwan plant and is converting it into a memory production facility. The Singapore HBM packaging plant that broke ground last Jan. will begin mass production as early as the end of this year. In Onondaga County, New York, the company is pursuing a "mega fab" project with a total investment of $100 billion (145.62 trillion won), and it also announced an expansion of its Hiroshima, Japan, plant. Recently, to increase DRAM output, it agreed to acquire the "P5 plant" (P5 fab) of Taiwanese semiconductor company PSMC in Miaoli County for $1.8 billion (about 2.65 trillion won).
Micron cited "responding to HBM demand" as the rationale for its large-scale capacity expansion. As a result, analysts said HANMI Semiconductor's TC Bonder sales could rise. HANMI Semiconductor established a Singapore subsidiary last Oct. in line with Micron's Singapore plant expansion.
Of HANMI Semiconductor's cumulative sales in the third quarter last year (493.671 billion won), 80.94% (399.625 billion won) came from overseas, with Micron reportedly accounting for a substantial portion. In 2024, due to expanded TC Bonder orders from SK hynix, 58.73% (328.278 billion won) of its annual revenue (558.917 billion won) came from the domestic market.
The market sees HANMI Semiconductor's results this year as hinging on Micron's capacity expansion. In the case of SK hynix, it is diversifying its TC Bonder supply chain, which is gradually increasing HANMI Semiconductor's dependence on Micron.
A semiconductor industry official said, "After receiving full-scale production orders (HBM4) from customers, related production equipment (TC Bonders) is ordered accordingly," adding, "While equipment orders also come in during mass production optimization or prototype production stages, full-scale orders are placed after customers finalize production volumes." In fact, SK hynix resumed TC Bonder orders, which had been halted since the second half of last year, with a 9.65 billion won order last month as HBM4 supply became visible.
A researcher at a market research firm who requested anonymity said, "If SemiAnalysis' outlook is accurate, it would be natural for Micron to reconsider its capital expenditure execution plans." At the end of last year, Micron raised this year's capital expenditures from the previous $18 billion (about 26.4 trillion won) to $20 billion (about 29.3 trillion won), but if Nvidia orders are delayed, there is a possibility of a major revision.
Meanwhile, the semiconductor industry is divided over the SemiAnalysis report. Mizuho Securities said, "The view that Micron will drop out of Nvidia's HBM4 supply chain is foolish." Since Micron refuted the performance issues raised in the market during its year-end earnings and said "including HBM4, the 2026 HBM supply is sold out," it would not be an early exit.
On the other hand, some interpret Micron's HBM4 sellout mentioned in its earnings as reflecting orders from companies other than Nvidia. A semiconductor industry official said, "It has been consistently noted that Micron's HBM4 is inferior in performance to Samsung Electronics and SK hynix, so its supply volume to Nvidia could be smaller."