Xcena raises $135M to develop memory-centric AI chip MX1

South Korean chip startup Xcena has closed a $135 million Series B round to commercialize its MX1 computational memory chip. The round was led by Atinum Ventures, with IMM Investment and LB Investment joining.

The thesis is straightforward: while most of the AI hardware arms race has centered on cramming more compute into data centers, Xcena thinks the actual chokepoint is data movement. Moving bits back and forth between processors and memory is slow, power-hungry, and increasingly the thing that keeps AI inference from scaling efficiently. The MX1 chip is designed to process data where it already lives, rather than shuttling it across the bus.

What the MX1 actually does

The MX1 integrates over 1,000 RISC-V cores for near-data processing, meaning computation happens right next to the stored data instead of requiring constant back-and-forth transfers. The chip is built on CXL 3.x architecture, supports DDR5 memory, and runs on PCIe Gen6.

Xcena’s particular focus areas include vector search, analytics, and KV cache management. KV (key-value) caches are what large language models use to store context during inference, and they’re becoming enormous as context windows grow.

The MX1 won the “Most Innovative Memory Technology” award at FMS 2025. Samples are expected by October 2025, with mass production slated for 2026.

The funding trajectory

Xcena was originally founded as MetisX in early 2022 before rebranding in December 2024. The seed round in 2022 brought in roughly $6 million. A $44 million Series A followed in 2024, valuing the company at approximately $167 million post-money. Now the $135 million Series B, denominated at roughly KRW 150 billion, represents a significant jump. Total funding before this latest round sat around $50 million. The funds are earmarked for MX1 commercialization, customer deployments, and development of next-generation product lines.

Why this matters beyond traditional AI

The competitive landscape includes established players like Samsung and SK Hynix, both of which are investing heavily in CXL-based memory solutions from the DRAM side. Xcena’s differentiator is the computational element, embedding processing directly into the memory device rather than simply making faster memory sticks.

The risk, naturally, is execution. Xcena needs to deliver working samples by October 2025, convert those into customer design wins, and then ramp manufacturing in 2026.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.