This chip startup just raised $135M on a bet that AI’s biggest bottleneck isn’t compute — it’s memory - Benzatine Infotech

In the world of artificial intelligence, every query to platforms like ChatGPT initiates a complex data relay process. This involves information transitioning from memory, being processed by a CPU, undergoing heavy computation via a GPU, and then returning — with this cycle repeating for each word generated by the AI. This structure presents a significant bottleneck, as it requires routing through some of the most expensive and power-hungry chips in the industry for every request. XCENA, a startup with bases in South Korea and the U.S., is focused on addressing this inefficiency. Established four years ago, the company has engineered a chip that brings computing capabilities closer to DRAM, the fast memory chips that actively store data for processors. By enabling routine data operations to occur near memory, XCENA aims to eliminate the costly back-and-forth between CPUs, GPUs, and memory. If successful, this innovation could drastically reduce AI infrastructure costs, an appealing prospect that has garnered significant investor interest. Recently, XCENA secured $135 million in a Series B funding round, boosting its valuation to $570 million and raising its total funding to $185 million. The startup was co-founded in 2022 by CEO Jin Kim, CTO Dohun Kim, and CPO Harry Juhyun Kim, all of whom have extensive backgrounds with major memory chip manufacturers like Samsung and SK Hynix. "Over the years, CPUs and GPUs have evolved significantly, but memory technology has lagged behind. XCENA aims to change that," Kim told TechCrunch. The recent surge in memory prices and stock performances highlights a broader shift towards memory-centric architectures in AI infrastructure. Notably, this month, Samsung, SK Hynix, and Micron, the leaders in the global memory chip market, all achieved a trillion-dollar valuation for the first time. XCENA's business model is predicated on the belief that "inference is not just a compute issue; it increasingly hinges on memory scalability," Kim explained. Their MX1 chip employs Compute Express Link (CXL) to create a direct connection between the CPU and memory, allowing data to be processed without needing to leave the memory module. The innovative design means that tasks which traditionally required ten servers could potentially be handled by just one. According to Kim, while GPUs are adept at matrix multiplication for AI model training, several support tasks, such as data orchestration and caching, still rely on CPUs. The MX1 chip will manage these operations directly within the memory module, streamlining efficiency. Since the latter half of last year, demand for memory solutions has surged, setting a favorable stage for XCENA. The company is currently in initial discussions with various global memory vendors, although Kim has chosen not to disclose their names. Their ideal clientele includes hyperscalers investing billions annually in AI infrastructure, where even marginal gains in memory efficiency can lead to substantial cost reductions. Currently, the MX1 is in the prototype phase, with mass production expected to commence at Samsung’s fabrication facilities by the end of 2026, and revenue generation projected for 2027. While competitors in the neural processing unit space, such as Astera Labs and Marvell, vie for dominance in training workloads, XCENA is focusing on optimizing the memory-intensive layer that underlies these processes. Kim highlighted that XCENA's unique value proposition lies in its advanced intellectual property. With thousands of specialized cores designed based on the open-source RISC-V architecture, the company emphasizes efficiency and performance. Their comprehensive approach includes designing their internal memory hierarchy, interconnect bus, and DRAM controller — a level of integration that sets them apart from many larger chip manufacturers. Seoul-based venture capital firms Altinum and IMM Investment led the recent funding round, joined by Corstone Asia and existing investors SBI Investment and Mirae Asset Capital. With a workforce of over 90 across their offices in Pangyo and Sunnyvale, XCENA is also exploring opportunities for additional funding from international investors.