Memory startup scores $135M to fix AI’s hidden traffic jam

When you type a question into ChatGPT, your request sets off a frantic data relay. Information shuttles from memory to a CPU for preprocessing, over to a GPU for heavy math, then back again — and that loop repeats for every single word the model generates. The problem isn't a lack of computing power; it's a structural inefficiency that forces every query through some of the most expensive, power-hungry chips in existence. XCENA, a four-year-old startup with offices in South Korea and the U.S., has designed a chip that moves processing closer to DRAM — the fast memory chips that hold active data. By handling routine operations near memory, it avoids costly round trips between CPUs, GPUs, and memory modules. Investors are betting big that this approach can slash AI infrastructure costs. XCENA just closed a $135 million Series B at a $570 million valuation, bringing total funding to $185 million. CEO Jin Kim, a Samsung and SK Hynix veteran, argues that while CPUs and GPUs have evolved, memory has stagnated. “Inference isn’t just a compute problem; it’s increasingly a memory scaling problem,” he says. The company’s MX1 chip uses CXL — a dedicated express lane between processor and memory — to process data before it ever leaves the memory module. XCENA claims this could reduce server requirements from ten machines down to one. The chip is still a prototype, with mass production slated for Samsung’s foundries by late 2026 and revenue expected in 2027. Rivals like Astera Labs and Marvell are also targeting memory connectivity, but Kim says XCENA’s thousands of small, RISC-V-based cores and in-house design of memory hierarchy and DRAM controllers give it an edge. The Series B was co-led by Seoul-based Altinum and IMM Investment, with participation from Corstone Asia and existing backers. XCENA, which employs over 90 people, is also in early talks with global memory vendors and hyperscalers — the kind of customers where even a modest efficiency gain could save hundreds of millions annually.