HyperAccel Struggles to Secure Samsung Foundry Capacity, SK hynix Memory

What happened

South Korean AI semiconductor startup HyperAccel is struggling to secure foundry capacity from Samsung Electronics and memory supplies from SK hynix ahead of planned mass production of its first-generation Bertha LPU in the second quarter of next year. The company requested production capacity for 20,000 Bertha accelerators from Samsung Electronics; with Samsung's orders already at full capacity, production was delayed beyond its preferred schedule, and after discussions it secured an initial allocation of 10,000 units. Bertha, a language processing unit (LPU) for data centers, is manufactured using Samsung Foundry's 4-nanometer SF4X process. Each Bertha accelerator card carries eight channels of LPDDR5X memory totaling 192GB; accounting for manufacturing yields and spare inventory, reliably manufacturing 10,000 accelerator cards requires approximately 100,000 LPDDR5X chips. HyperAccel requested these 100,000 memory chips from SK hynix but was told it would be difficult and secured only a partial allocation. CEO Kim Joo-young disclosed the difficulties during a Q&A after a breakfast seminar hosted by the AI Semiconductor Forum at the Grand InterContinental Seoul Parnas hotel on Nov. 12, arguing that fabless startups are disadvantaged in securing foundry capacity and procuring tens of thousands of memory chips, and called for government-level support for startup supply chain access. The company plans mass production of the first-generation Bertha LPU in the second quarter of next year, followed by the second-generation Carmen in 2029. It recently raised approximately 150 billion won in a Series B round with participation from Krafton and LG Electronics, with cumulative funding exceeding approximately 210 billion won.

Impact and signal

Affected components and suppliers

Source and time

Source: The Elec · 11/12/2026 · View original source