The Seoul and U.S.-based chip startup closed a $135 million Series B at a $570 million valuation, arguing that the industry's fixation on GPU supply is obscuring a deeper infrastructure problem that no amount of raw compute can solve.
Every time a large language model generates a single word, data leaves memory, passes through preprocessing on a CPU, travels to a GPU for heavy computation, then makes the return trip back. That round-trip happens repeatedly, thousands of times per inference request, through some of the most expensive and power-hungry chips in the data center. It is a structural inefficiency that the AI industry has largely tolerated because GPUs kept getting faster. Xcena, formerly known as MetisX, is betting that tolerance is about to run out, and investors just handed the company $135 million to prove the point.
The Series B, reported by TechCrunch today, values the startup at $570 million and brings total capital raised to $185 million. Atinum Investment led the round, with existing backers IMM Investment and LB Investment participating alongside new entrants. The raise is notable not just for its size but for the argument behind it: that memory bandwidth, not GPU supply, is the defining constraint on AI infrastructure as training and inference workloads continue to scale.
The company's core thesis is a compute-near-memory architecture. Rather than shipping data back and forth across the conventional chip hierarchy, Xcena places processing capabilities directly adjacent to DRAM, handling routine data operations at the memory layer before the expensive GPU-bound journey ever begins. The design uses RISC-V cores kept deliberately small and efficient, optimized specifically for data movement rather than general computation. Xcena also designs its own internal memory hierarchy, interconnect bus, and DRAM controller, giving it control over the full data path rather than relying on off-the-shelf components that were never designed with AI throughput in mind.
The founders know this stack from the inside. CEO Jin Kim, CTO Dohun Kim, and CPO Harry Juhyun Kim all came out of Samsung and SK Hynix, the memory giants whose high-bandwidth memory chips sit inside Nvidia's flagship GPUs. That lineage matters. Understanding how HBM behaves under AI workloads is not something you learn from a textbook, and the Xcena team spent years designing the components that every major hyperscaler now depends on. They founded the company in 2022 under the MetisX name before rebranding to Xcena to signal a broader vision for how computing architecture needs to evolve.
The bottleneck debate heating up across the industry
The memory wall is not a new concept in computer science, but it has become newly urgent as AI model sizes have grown and inference has moved from research to production at enormous scale. Microsoft recently attributed $25 billion of its record AI infrastructure budget to increased memory and chip costs, a figure that underscores just how central the memory layer has become to the economics of running AI systems. When a hyperscaler is spending at that scale, a meaningful improvement in memory efficiency does not represent a marginal optimization. It represents hundreds of millions of dollars in potential savings per year.
That is precisely the customer Xcena is targeting. The company's ideal buyer is a hyperscaler already deploying billions in AI infrastructure, where the cost of moving data inefficiently compounds at a massive scale. Traditional memory vendors like Micron and SK Hynix supply the raw hardware and have their own roadmaps for improving bandwidth, while Marvell and similar players have pushed aggressively into custom interconnect silicon. Xcena's positioning sits between those camps, arguing that neither pure memory density gains nor better interconnect routing fully addresses the data movement problem at the point where it actually occurs.
That framing puts Xcena in an interesting competitive position. It is not directly challenging Nvidia for training workloads, which is the more crowded and well-funded fight. It is targeting the layer that sits underneath all of it, a place where purpose-built silicon could reduce the computational overhead that the GPU-centric stack currently absorbs without question.
The road to production
The practical test of the thesis comes soon. Mass production chips are scheduled to roll off Samsung's foundry lines by the end of 2026, with revenue expected to begin in 2027. That timeline is aggressive but not unrealistic for a company whose founders built institutional knowledge at Samsung before striking out independently. Using Samsung as the manufacturing partner also provides a degree of supply chain credibility that fabless chip startups often struggle to establish, particularly given the geopolitical sensitivity around advanced semiconductor production.
The broader AI hardware investment cycle has seen a wave of contrarian bets over the past eighteen months, from inference-optimized silicon startups like Fractile to memory architecture plays like VERTICAL COMPUTE in Europe. What distinguishes Xcena's round is the combination of size, valuation, and the specificity of the claim being made. This is not a general AI chip play hedging across multiple workloads. It is a focused argument that data movement is the problem, that moving compute to memory is the solution, and that the hyperscalers who figure this out first will have a structural advantage over those still throwing GPUs at every bottleneck they encounter.
Whether the market agrees will become clear when production chips ship. If Xcena can demonstrate measurable efficiency gains at hyperscaler scale, the $570 million valuation will look modest in retrospect. If the memory bandwidth thesis turns out to be one bottleneck among many rather than the central constraint, the company will face the harder challenge of repositioning against better-funded competitors in adjacent segments. For now, the bet is placed and the timeline is set.
Also read: Try Verity Builds AI Interview Practice Around the Candidate's Own CV and Target Role • ElevenLabs Dubbing v2 bets that AI can finally crack the global localization market at scale • AskFred Gives Small Businesses a 24/7 AI Assistant Built From Their Own Data
