[E-Daily Marketin Song Seung-Hyeon Reporter] The biggest bottleneck for AI servers is memory. As the KV cache—where large language models (LLMs) store past computations—accumulates, the required memory capacity increases exponentially. This problem is becoming even more acute as AI agent services become more widespread. XCENA, a CXL-based memory fabless startup, has set its sights squarely on this bottleneck. The investment market has bet 200 billion won on this solution.

According to the venture capital (VC) industry on the 4th, XCENA has finalized a Series B funding round of approximately 200 billion won, co-led by Atinum Investment and others. This represents 1.4 times the initial target of 145 billion won. Existing investors—Mirae Asset Capital, Mirae Asset Venture Investment, LB Investment, Stick Ventures, Wonik Investment Partners, and SBI Investment—participated in the follow-on round, while a large number of new investors, including the Industrial Bank of Korea, Shinhan Venture Investment, Kyobo Securities, and Kyobo Life Insurance, also joined. The company’s valuation was set at 800 billion won. This represents approximately a threefold increase from the 250 billion won valuation during its Series A round just two years ago. It is reported that competition over the allocation of investment shares continued until the latter stages of the round.

Kim Jin-young, founder and CEO of Exina, is a memory semiconductor expert who started his career in Samsung Electronics’ Memory Business Division and rose to become the youngest technical executive at SK Hynix. Drawing on his experience overseeing the development of next-generation solution products at SK Hynix, he founded Exina in 2022. The company’s CTO and CPO are also former SK Hynix employees who were responsible for next-generation memory development. This “manpower” is cited as the primary reason Exina was able to secure an investment in the 200 billion won range just three years after its founding.

What is particularly noteworthy is the structure of the market Exina is targeting. Existing AI servers rely on GPUs to handle computations, with data constantly being exchanged with memory such as HBM. The more frequent the data transfers, the more bottlenecks occur, and this problem intensifies as the AI inference market grows. KV caches are a prime example. KV caches, which large language models (LLMs) use to store past computations for reuse, require more memory as usage increases. As AI coding tools and agent services become more widespread, the KV cache problem is becoming increasingly significant. This is why the shortage of memory semiconductors for big tech AI companies is worsening.

Exina’s flagship product, the MX1, takes a fundamentally different approach. Instead of moving data out of memory for processing, it processes it right next to the memory. This method reduces unnecessary data movement by integrating computing units with memory components such as DDR5 DRAM and SSDs. The underlying technology is CXL (Compute Express Link). It is a next-generation interface technology that connects various computing resources—such as CPUs, GPUs, memory, and accelerators—at high speeds to enhance server memory scalability and resource utilization. While existing CXL products focus on memory expansion capabilities, the MX1 distinguishes itself by combining data processing capabilities with expansion, thereby taking system efficiency to the next level. In Silicon Valley, companies such as Astera Labs and Marvell Technologies are developing similar products, and adoption is beginning, particularly among hyperscalers.

Exina is currently working with North American hyperscalers and CXL ecosystem partners to validate the MX1 product and optimize systems. The company plans to manufacture the product using Samsung Electronics’ 4-nanometer foundry process, with the goal of entering mass production following the tape-out of the next-generation product in the second half of this year. With CXL 3.0-based CPUs scheduled for release this year, the CXL market is expected to expand in earnest starting next year.

CEO Kim Jin-young stated, “AI workloads are exposing the fundamental limitations of existing computing architectures, and the MX1 is rapidly moving beyond proof of concept into the practical commercialization phase.” He added, “Through this funding, we will expand our MX1-based customer collaborations and lead the era of memory-centric computing.”