XCENA raises $135M to make memory, not compute, the center of AI performance
CEO Jin Kim and a Samsung/SK hynix veteran team rebrand MetisX as XCENA and double down on CXL computational memory with MX1, at a $570M valuation.
By Ryan Merket · Published
Why it matters
If memory is the constraint, not compute, then adding GPUs delivers diminishing returns. XCENA’s CXL computational memory aims to raise AI cluster utilization and cut cost by moving compute to the data.
XCENA has raised $135 million at a $570 million valuation, betting that AI's real choke point is memory bandwidth and locality rather than more flops, according to TechCrunch. CEO Jin Kim is steering the South Korea-based company's rebrand from MetisX to XCENA and pushing a product thesis built by alumni of Samsung and SK hynix: move compute closer to where data lives and make memory a first-class accelerator.
The founder bet: rewrite the memory hierarchy for AI
Kim and team founded XCENA to attack what they argue is an architectural tax in modern AI: shuttling tensors between host memory, CPUs, and GPUs. Their materials frame the issue plainly: if models spend cycles waiting on memory, more GPUs will not fix it. XCENA rebranded from MetisX as part of that strategic reset, the company notes in its newsroom.
What they are building
XCENA's flagship platform, MX1, is a CXL computational memory device that pools high-capacity DDR5 and adds near-data processing (NDP) cores over the open Compute Express Link standard. In plain terms, it lets systems expand memory beyond CPU socket limits and execute certain operations in-memory, reducing data movement and latency. The company says this can cut energy use and total cost of ownership for memory-bound AI workloads.
Hardware is only part of the bet. XCENA ships a full-stack software layer so customers do not have to rewrite everything to see gains. The XCENA SDK offers high-level runtime APIs for drop-in use and low-level device APIs for fine control of data placement and compute execution, plus simulation tools and drivers for major operating systems.
Why memory, not more compute
Large models are increasingly bottlenecked by how quickly and efficiently data can be fetched, transformed, and fed to accelerators. XCENA's approach leans on CXL's memory pooling and coherency to:
- Expand effective memory capacity available to CPUs and accelerators.
- Offload memory-side ops to NDP cores, keeping data local.
- Improve utilization of expensive GPUs by feeding them more consistently.
If it works in practice, the result is not just faster tokens per second but better economics at cluster scale.
Where it fits and who it is for
XCENA says it is targeting hyperscalers, telcos, and research institutions that are scaling AI infrastructure and running into memory ceilings. The company is headquartered in Seongnam-si, Republic of Korea, with a presence in Sunnyvale, California, per its website. The team's background spans Korean memory giants, and the product roadmap is anchored to the CXL ecosystem rather than proprietary interconnects.
Details like investor names, round stage, and shipping timelines were not disclosed in the materials we reviewed. But the contours are clear: Kim is turning XCENA into a memory-first company that aims to complement, not replace, the GPU. If XCENA can deliver real gains without forcing customers to rethink their entire software stack, its MX1 line could become a staple in AI racks that have been starved for capacity and bandwidth.