This chip startup just raised $135M on a bet that AI’s biggest bottleneck isn’t compute — it’s memory
XCENA raised $135 million to push AI data processing closer to memory instead of routing everything through CPUs and GPUs. The startup says that shift could cut infrastructure costs for large AI operators.
Intelligence analysis by GPT-5.4 Mini
XCENA, a four-year-old chip startup with roots in Samsung and SK Hynix, is pitching a memory-centric alternative to today’s AI infrastructure. Its MX1 chip is designed to handle data tasks inside the memory module itself, which the company says could reduce expensive server traffic and improve efficiency.
XCENA is trying to fix a traffic problem inside AI computers. Right now, information has to travel back and forth between different parts, like a package going through several delivery stops before it reaches its destination.
The company says it can put more of that work right next to the memory, where the data already lives. That could make AI machines use less energy and fewer servers.
Investors liked the idea enough to give XCENA $135 million. The company still has to prove the chip works at a large scale, but it thinks the future of AI is not just about bigger brains, it is also about faster shelves where the information is stored.
Analysis
The bet
XCENA is built around a simple claim: AI systems spend too much time moving data back and forth between memory, CPUs, and GPUs. According to the company, that routing is expensive, power-hungry, and unnecessary for many routine operations.
What the chip does
The startup’s MX1 chip is meant to bring computation closer to DRAM, the memory used for active data. It connects to the CPU through CXL, so data can be processed near the memory module instead of being shuttled to other chips first. XCENA says that this can cover tasks like preprocessing, KV cache management, and data caching.
Funding and timeline
XCENA raised $135 million in a Series B at a $570 million valuation, bringing total funding to $185 million. The round was led by Seoul-based Altinum and IMM Investment, with Corstone Asia and existing backers SBI Investment and Mirae Asset Capital also participating. The company says mass production is planned for Samsung foundry lines by the end of 2026, with revenue expected in 2027.
Competition and market context
The company’s rivals include Astera Labs and Marvell, both working on next-generation memory connectivity. XCENA argues its advantage comes from heavier internal integration: it designs its own memory hierarchy, interconnect bus, DRAM controller, and uses many small RISC-V cores for data processing. The startup is targeting hyperscalers, where small efficiency gains can translate into large savings.
Key points
- XCENA raised $135 million in a Series B at a $570 million valuation.
- The company argues AI inference is becoming a memory-scaling problem, not only a compute problem.
- Its MX1 chip is designed to process data closer to DRAM using CXL.
- XCENA says the chip could replace work that would otherwise require multiple servers.
- Mass production is planned for late 2026, with revenue expected in 2027.
