South Korean startup Xcena has raised $135 million to develop a memory-centric processor called MX1. Instead of shuttling data between separate memory and processing units, the MX1 embeds thousands of RISC-V cores and vector engines directly alongside memory. This near-data processing approach is designed to reduce the costly round trips that slow down AI inference workloads.
The chip supports CXL 3.0 and 3.2 along with DDR5 memory controllers, enabling fast data movement in modern data centers. Xcena claims the MX1 can deliver up to 3.9x improvements in time-to-first-token for AI inference benchmarks-a key metric for large language model performance.
Founded in 2022 under the name MetisX, the company rebranded to Xcena at the end of 2024. CEO Jin Kim leads the startup, which had previously raised about $50 million. Working samples of the MX1 were targeted for October 2025, with a production version planned for 2026.
Memory bandwidth has become a critical bottleneck for AI models. Xcena’s approach aims to place processing power directly at the memory level, a strategy that could prove valuable as demand for efficient inference grows. The company won the FMS 2025 Best of Show award and showcased its technology at the OCP Global Summit 2025.