XCENA Raises $135M to Put AI Compute Inside the Memory Module
XCENA closed a $135 million Series B at a $570 million valuation. Total funding is now $185 million. The Seoul and Sunnyvale startup makes a chip called MX1 that processes AI inference workloads from inside the memory module, not the GPU.
The Bet
XCENA's thesis is simple: AI inference is a memory problem, not a compute problem. As models grow, the bottleneck moves from raw compute to how fast you can move data between memory and processor. MX1 attempts to sidestep that bottleneck by doing the work where the data already lives.
The chip connects to a server's CPU via CXL (Compute Express Link), a relatively new interconnect standard designed for exactly this kind of memory-attached compute. MX1 handles preprocessing, KV cache management, and data caching directly within the memory module.
XCENA claims this cuts server requirements by 10x. A workload requiring 10 servers could theoretically run on one.
That's a large claim for a prototype.
The Hardware
MX1 runs thousands of cores built on RISC-V, the open-source chip architecture that's become common in custom silicon. XCENA designs its own memory hierarchy, interconnect bus, and DRAM controller internally. That level of vertical integration is unusual for a 3-year-old company with 90 staff.
The team has the background for it. CEO Jin Kim, CTO Dohun Kim, and CPO Harry Juhyun Kim all came from Samsung and SK Hynix. They founded XCENA in 2022. Memory architecture is not a new domain for them.
Mass production is scheduled on Samsung foundry lines by end of 2026. Revenue is expected to start in 2027.
The Market
XCENA is not alone in the memory connectivity space. Astera Labs and Marvell are the established players in CXL-adjacent infrastructure. Both are larger, better funded, and already shipping. XCENA is entering a market with real competition and doing so with a chip that hasn't left prototype status yet.
The Series B was co-led by Atinum and IMM Investment, with participation from Corstone Asia, SBI Investment, and Mirae Asset Capital. The investor list is primarily Korean, which tracks with the company's roots in Pangyo.
What It Means
The near-memory compute idea has been circulating for years. CXL gives it a real interconnect standard to build on. If MX1 ships on schedule and the 10x server reduction holds under production conditions, XCENA's valuation will look reasonable. If either slips, $570 million is a lot of weight on a prototype.
The 2027 revenue window gives them time. Not a lot of it.
Source: Techcrunch