South Korean AI chip startup Rebellions is directing its technology roadmap toward custom memory architectures to secure its supply chain and position for a potential initial public offering. CEO Sunghyun Park said the strategy relies on partnerships with SK Hynix and Samsung Foundry to meet the bandwidth demands of large-scale language model inference.
Shifting the memory architecture
Rebellions originally planned a 3D SRAM stack for its next-generation chips but shifted to 3D-stacked DRAM. Park said the industry is moving away from commodity memories, requiring specialized architectures and memory pooling for scale-up and scale-out solutions. The company is currently co-designing HBM memory and logic dies with its domestic partners.
The company's second-generation accelerator, Rebel, delivers 1 POPS of FP16 compute and 144 GB of HBM4e within a 300-watt power envelope. Technical leaders tracking AI for CTOs will note that this memory-centric approach relies heavily on co-designing logic and memory dies. Park believes this will define the next wave of hardware winners. "A year ago the emphasis was on chiplets, but now the emphasis is on memory and memory-centric architectures," he said.
Market expansion in the Middle East and telecom
Rebellions is drawing considerable interest from Saudi Arabia, where entities like Humain and Aramco view AI infrastructure as central to their 2030 goals. Park said recent regional instability has not slowed these ambitions. He described sovereign AI deployments in the region as heterogeneous platforms where Nvidia and non-Nvidia hardware coexist.
Domestically, the largest deployment of Rebellions hardware is at SK Telecom. A multi-rack cluster of first-generation chips partially powers Adot, the carrier's proprietary AI assistant. Adot handles up to 50 million API calls daily for services like phone call summarization, making it the largest token user in South Korea.
Capital strategy and disaggregation
Rebellions raised $400 million in a pre-IPO round in March, bringing total funding to $850 million. The company is evaluating Nasdaq and domestic listing options, though Park said no definitive plans are established. The recent substantial IPO of Cerebras has sharpened investor focus on low-latency inference, a tailwind for Rebellions' strategy.
The startup is also testing disaggregated inference architectures with Arm and SK Telecom. In this setup, Rebellions hardware accelerates the decode stage. Park said he remains uncertain whether disaggregation is the correct overall direction, noting that the current Rebel chip already includes HBM to handle prefill workloads.
Why this matters for executives and strategy
Hardware startups are increasingly betting on memory partnerships rather than pure compute density to differentiate their products. As companies evaluate infrastructure investments, the strategic principles of AI for Executives & Strategy highlight that supply chain control over custom memory logic will dictate market share as much as raw compute. Organizations planning large-scale deployments should anticipate heterogeneous hardware environments and prioritize vendors with flexible, open-source software stacks.
Your membership also unlocks:
