Competition for AI infrastructure is expanding from GPUs to memory. XCENA, a chip startup operating in both South Korea and the United States, has completed a $135 million Series B funding round, valuing the company at $570 million post-money. The company believes that the cost pressures of generative AI in the inference stage come not only from computing chips but also from the repeated data transfer between CPUs, GPUs, and memory.
The company's valuation rose to $570 million after the funding round.
Following this round of financing, XCENA has raised a total of $185 million. The company was co-founded by CEO Jin Kim, CTO Dohun Kim, and CPO Harry Juhyun Kim, all three of whom previously worked at Samsung Electronics and SK Hynix. These two companies are major global memory chip suppliers and crucial components of Nvidia's GPU supply chain.
TechCrunch reports that XCENA believes AI inference is shifting from a "computing power problem" to a "memory expansion problem." As the model calls upon context, caching, and preprocessing processes to generate each lexical unit, data is transferred back and forth between different chips, resulting in additional latency, power consumption, and hardware costs.
The chip moves some computations closer to memory.
XCENA's product name is MX1, and it is currently still in the prototype stage. The chip connects to the CPU via CXL, moving some data processing tasks closer to the DRAM, reducing the amount of data traveling between the CPU, GPU, and memory.
The company states that these tasks include preprocessing, key-value cache management, and data caching. According to them, work that would normally require 10 servers can potentially be completed on a single server in certain scenarios. However, this effect still needs to be verified through large-scale deployment.
Mass production is planned for the end of 2026.
XCENA anticipates that mass-produced chips will roll off the production line at Samsung's foundry by the end of 2026 and begin generating revenue in 2027. The company's current target customers are hyperscale cloud service providers that invest hundreds of billions of dollars annually in AI infrastructure, as even a small improvement in memory efficiency could result in cost savings of hundreds of millions of dollars.
The company stated that since the second half of last year, market demand for memory-related solutions has increased significantly, and it has already made early contact with several global storage manufacturers, but did not disclose the specific list.
Competitors include Astera Labs and Marvell
In terms of the competitive landscape, XCENA considers Astera Labs and Marvell to be close competitors. Both companies are advancing next-generation memory interconnect solutions, with Marvell already a mature, publicly traded company.
XCENA believes its main difference lies in the chip's internal architecture. The company states that its solution is based on RISC-V, employs a large number of small data processing cores, and develops its own memory hierarchy, interconnect bus, and DRAM controller, rather than outsourcing these components to third parties.
This funding round was co-led by South Korean venture capital firms Altinum and IMM Investment, with participation from Corstone Asia, existing shareholders SBI Investment, and Mirae Asset Capital. XCENA currently has offices in Pangyo, South Korea, and Sunnyvale, USA, with over 90 employees, and is in discussions with international investors regarding follow-up funding.
