A four-year-old chip startup just raised serious money on a contrarian idea: AI’s biggest bottleneck isn’t compute, it’s memory. According to TechCrunch AI, XCENA, which runs offices in South Korea and Sunnyvale, California, closed a $135 million Series B at a $570 million valuation. That brings its total funding to $185 million, and it tells you where smart money thinks the next infrastructure fight is headed.
What stands out here is the target. While everyone else races to dethrone Nvidia on AI training, XCENA is going after the layer underneath all of it.
What Actually Happened
XCENA designed a chip, the MX1, that puts compute power right next to DRAM, the fast short-term memory a processor actively uses. Today, every word an AI generates triggers a relay race: data leaves memory, runs through a CPU, travels to a GPU, then comes back. That round trip repeats for every token, routing through the most expensive and power-hungry chips in the building each time.
XCENA’s pitch is to kill the trip. The MX1 handles routine data operations near the memory itself, connecting to the CPU through CXL, what the company calls a dedicated express lane between processor and memory. As CEO Jin Kim told TechCrunch AI, “CPUs and GPUs have both gotten smarter over the decades. Memory never did. XCENA wants to change that.”
The founding team isn’t random. Kim co-founded the company in 2022 with CTO Dohun Kim and CPO Harry Juhyun Kim, all veterans of Samsung and SK Hynix, the memory giants that supply the chips inside Nvidia’s GPUs.
Why This Matters
The core thesis is that AI inference, the part where models answer your questions, isn’t just a compute problem anymore. It’s a memory scaling problem. GPUs are great at the heavy matrix math behind training. But a lot of the surrounding work still runs on CPUs:
- Preprocessing incoming requests
- KV cache management, the system that stores prior conversation context so a model doesn’t reprocess it
- General data caching and orchestration
XCENA wants that work done inside the memory module. The company claims what used to need 10 servers could potentially run on one. For hyperscalers spending tens of billions a year, even a small gain in memory efficiency can translate to hundreds of millions in savings. That math explains the investor enthusiasm.
The timing argument is hard to ignore. This month, Samsung, SK Hynix, and Micron, the three companies that dominate global memory, each crossed a trillion-dollar valuation for the first time. Memory prices and memory stocks are climbing together. Kim reads that as proof the industry is shifting toward memory-centric architectures.
The Competitive Picture
XCENA isn’t alone in this space. Its closest public rivals are Astera Labs and Marvell, both working on next-generation memory connectivity. Marvell is the established player, and Kim says the difference comes down to intellectual property.
His claim: “We have thousands of cores.” Based on public specs, Marvell leans on a handful of general-purpose cores. XCENA’s are built on RISC-V, the open-source chip blueprint, each one kept small and tuned specifically for data processing. The company also designs its own internal memory hierarchy, interconnect bus, and DRAM controller, a degree of vertical integration most chip firms outsource.
What To Expect Next
This is still early. The MX1 is a prototype. Here’s the rough timeline from the reporting:
- Mass-production chips roll off Samsung’s foundry lines by the end of 2026
- Revenue is expected to start in 2027
- Talks with major memory vendors are in early stages, and Kim declined to name names
XCENA now has more than 90 staff across Pangyo, outside Seoul, and Sunnyvale, with Seoul VC firms Atinum and IMM Investment co-leading the round. It’s also courting international investors for more.
My take: even if XCENA’s specific numbers prove optimistic, the bet itself is the story. The whole industry has spent two years obsessed with GPU supply. If inference economics really are turning into a memory problem, the companies sitting on that layer become a lot more interesting. Worth watching whether the 2027 revenue promise holds. You can find the full details in the original report at TechCrunch AI.