XCENA builds computational memory chips for AI workloads. Its chips move data processing closer to DRAM. This reduces latency, power use, and data transfer costs.
The startup raised $135 million in Series B funding. Its total funding now stands at $185 million, with a $570 million valuation.
XCENA was founded in 2022 by Jin Kim, Dohun Kim, and Harry Juhyun Kim. The founders previously worked at SK hynix and Samsung.
What does XCENA's computational memory actually do?
XCENA’s flagship product, MX1, is a CXL-connected computational memory device that expands memory capacity while also doing work inside or near the memory layer itself. In plain English, that means a server can keep more data close at hand and offload chores like preprocessing and cache handling. It can also handle certain data-processing steps before the information makes a costly trip back to the CPU. XCENA is aiming that shift at AI inference, big data, vector databases, and other workloads where data movement drags on performance.
The hardware story is more ambitious than a plain memory expander. XCENA built MX1 around thousands of custom RISC-V cores and vector engines. It also includes memory compression and its own internal memory hierarchy, interconnect bus, and DRAM controller. The company describes support for CXL 3.2, PCIe 6.0 dual x8 links, and up to 2 TB of pooled DDR5 memory on the platform. That's an aggressive spec sheet.
There’s also a software layer, which matters a lot more than startups like this sometimes admit. XCENA provides an SDK with simulation tools and drivers. It also includes high-level runtime APIs and lower-level device APIs, so customers don't have to rewrite everything from scratch just to test the hardware. In a 2025 preview, the company said it would show MX1 with XFLARE, a library built to accelerate database queries. That hints at how XCENA wants to land inside real enterprise and hyperscale workflows rather than live as a science project.
Before MX1, a lot of this surrounding work stayed on the CPU while the GPU handled the heavy matrix math. After MX1 — at least in XCENA’s ideal setup — that orchestration gets pushed into the memory path itself. Jin Kim’s sales line is that what once needed 10 servers could, in some cases, shrink to 1. It's a huge claim. It needs real production proof.
Who founded XCENA and how far along is the company?
Founding story
XCENA started in 2022 with Jin Kim, Dohun Kim, and Harry Juhyun Kim. The company originally operated as MetisX before rebranding to XCENA, and from the start it aimed at large-scale data processing in AI, big data, vector databases, and even DNA analysis. That focus wasn’t random. It came straight out of the founders’ memory and SoC backgrounds.
Why these founders fit the job
Jin Kim had already been a corporate VP at SK hynix and led next-generation architecture work after earlier roles at Samsung Electronics and SK Telecom. XCENA described him as one of the company’s youngest executives. Dohun Kim brought 18 years of SoC R&D experience from SK hynix and Samsung SDI, while Harry Kim came in with 17 years spanning SoC and related software work at SK hynix and Samsung Electronics. This isn’t a team that woke up one morning and decided AI chips sounded hot. They’ve spent years inside the exact part of the stack they’re now trying to redesign.
Product status and early signals
For all the fundraising buzz, MX1 still isn’t a mass-market product. It’s a prototype, and XCENA is exploring the chip with select partners for validation. Mass production is scheduled on Samsung’s foundry lines by the end of 2026, and the company expects revenue to begin in 2027. XCENA also has more than 90 employees across Pangyo, near Seoul, and Sunnyvale. It's in early conversations with global memory vendors.
Funding and what the money buys
The new round is big by any deeptech standard: $135 million in Series B funding at a $570 million valuation. TechCrunch reported that Atinum and IMM Investment co-led the round, joined by Corstone Asia plus existing backers including SBI Investment and Mirae Asset Capital. XCENA’s own announcement adds a longer roster of financial and strategic investors. The money will go toward global expansion, customer deployments, go-to-market work, and next-generation computational memory products.
How XCENA computational memory stacks up against Astera Labs and Marvell
This part matters, because XCENA isn’t alone in seeing memory as the next AI bottleneck. Astera Labs already sells its Leo CXL smart memory controllers for memory expansion and pooling, with hardware that supports up to 2 TB. It has also published demo results showing faster LLM response workflows and higher throughput in inference-style workloads. Marvell’s Structera line goes after the same general problem with near-memory accelerators and memory-expansion controllers, using 16 Arm Neoverse cores, up to 200 GB/s of bandwidth, and support for more than 6 TB of DDR5 memory capacity on some configurations.
XCENA’s angle is doing more data orchestration inside the memory module itself, with thousands of small custom RISC-V cores instead of a handful of general-purpose cores. The incumbent alternative is still the old server pattern: let CPUs babysit preprocessing, caching, and context management while GPUs do the math. XCENA is trying to cut that handoff overhead out of the loop.
Why does this computational memory round matter?
Because XCENA is still pre-revenue hardware, this round isn’t just a victory lap. It has to carry the company from an interesting prototype to something hyperscalers might actually deploy. XCENA says the funding will support customer validation, global commercial expansion, and development of follow-on products. It’s also growing its Northern California presence to work more closely with customers and partners.
There’s a broader investor read-through here. XCENA isn’t trying to out-Nvidia Nvidia on training chips. It’s targeting the memory-heavy layer underneath inference, database work, and context management — the stuff that gets uglier as models grow, context windows stretch, and AI services become more interactive. If that thesis is right, memory-centric computing becomes less like a niche optimization and more like a budget line item every hyperscaler has to care about.
How big is the computational memory market for AI?
The easiest way to understand the timing is to zoom out. WSTS said global semiconductor sales hit $795.6 billion in 2025, up 26.2% year over year, and said the industry is approaching the $1 trillion mark in 2026. Even more telling, the computer segment grew by more than 60% in 2025, driven largely by data center and AI systems, while memory was one of the categories leading the rebound. This isn't a tiny corner of hardware anymore. AI infrastructure is dragging the whole semiconductor market with it.
Memory is getting pulled into that center of gravity. SK hynix said system-level optimization across CPU, GPU, and memory is becoming decisive in AI inference, not just the performance of a single chip. In a separate 2026 market outlook, it summarized outside estimates that the memory market could exceed $440 billion in 2026. That helps explain why CXL products are showing up across the stack, and why cloud vendors are starting to test CXL-attached memory in real environments rather than just conference demos.
That’s why XCENA is interesting even before revenue shows up. The company is lining up with a structural shift: AI workloads are becoming more memory-hungry and more latency-sensitive. They're also getting a lot more expensive to move around than the industry used to assume. If computational memory becomes a standard design choice instead of an exotic one, XCENA’s current prototype phase could look a lot more important in hindsight. What to watch next is simple: partner wins, silicon validation, and whether end-2026 mass production actually happens on schedule.
Read how Simple Energy raised ₹250 crore to scale its high-performance EV scooter business ahead of a planned FY28 IPO push.
FAQ
– What funding did XCENA raise?
XCENA raised $135 million in a Series B round announced on May 29, 2026. The round valued the company at $570 million and brought its total funding to $185 million, with Atinum Investment and IMM Investment leading and a wider group of Asian financial investors joining in.
– How does XCENA’s MX1 chip work?
MX1 is a CXL-connected computational memory chip that adds memory capacity and performs certain data-handling tasks closer to where the data already sits. XCENA built it to take work like preprocessing and KV cache management out of the usual CPU-GPU-memory shuffle. It also handles caching and some query acceleration, using thousands of custom RISC-V cores plus its own software stack.
– Who founded XCENA?
XCENA was founded in 2022 by Jin Kim, Dohun Kim, and Harry Juhyun Kim. Jin previously held senior architecture roles at SK hynix after earlier work at Samsung Electronics and SK Telecom, while Dohun and Harry both spent years in SoC development at major Korean chipmakers, giving the company unusually strong memory-system credibility for such a young startup.
– Is XCENA an AI chip company or a memory company?
It’s best described as a memory-centric AI infrastructure company. XCENA isn’t mainly selling training accelerators; it sits in the layer between compute and memory, using computational memory and CXL-based architecture to improve how AI inference systems handle data-heavy workloads.
