AI inference chip startup d-Matrix, backed early by Samsung and SK, formally joins Nvidia's ecosystem.

d-Matrix, a U.S.-based AI inference chip startup valued at $2 billion (approximately 2.7 trillion won), is joining Nvidia’s AI infrastructure ecosystem. The company will integrate its proprietary AI accelerator into Nvidia’s data center deployment platform. This development has drawn renewed attention given that SK Hynix, Samsung, and Mirae Asset have backed the firm since its earliest funding rounds in 2022.

Last week, d-Matrix announced a partnership to integrate its inference accelerator, the “Raptor XPU,” into Nvidia’s AI infrastructure platform. The integration links the Raptor XPU with data center GPUs, CPUs, and network infrastructure through Nvidia’s NVLink Fusion architecture. Nvidia is actively shifting inference workloads—core data center operations—from general-purpose GPUs to dedicated XPUs to improve efficiency and reduce costs. The d-Matrix partnership exemplifies Nvidia’s broader strategy of absorbing custom AI chip designers into its ecosystem.

Founded in 2019 as a fabless AI inference chip designer, d-Matrix raised $275 million (approximately 370 billion won) in a Series C round last November, achieving a $2 billion valuation. SK Hynix led the $44 million (approximately 60 billion won) Series A round in 2022, recognizing the company’s potential from its earliest stages. Samsung Venture Investment and Mirae Asset joined the $110 million (approximately 150 billion won) Series B in 2023, with Mirae Asset returning for the Series C round in November 2025. Samsung Venture Investment is jointly owned by six Samsung affiliates, including Samsung Electronics.

The semiconductor giants’ early backing stems from d-Matrix’s core technologies: Digital In-Memory Computing (DIMC) and 3DIMC. DIMC embeds compute functionality directly into memory, while 3DIMC stacks memory vertically atop a compute die. In the AI semiconductor market, reducing the physical distance between memory and compute units is becoming critical as performance demands rise. As AI models expand, the power and latency required to shuttle data between memory and processors increase, turning the so-called “Memory Wall” into a major bottleneck. While High Bandwidth Memory (HBM) transfers data horizontally with GPUs via an interposer, 3DIMC attaches memory vertically and directly to the logic or compute die, effectively eliminating physical data movement distances. d-Matrix claims this architectural shift reduces power consumption to one-tenth that of HBM4.

Analysts note that 3DIMC will not replace HBM but will instead drive overall memory demand alongside it. Beyond traditional GPU-centric HBM, new AI workloads are expected to spur demand for inference-optimized 3D DRAM, Processing-in-Memory (PIM), and similar architectures. Major memory manufacturers are already diversifying beyond HBM, actively developing AI-specific products such as AI-DRAM, 3D DRAM, and PIM.

Samsung’s strategic investments extend beyond South Korea, notably to Euclide, a Dutch AI semiconductor startup. On the 14th, Samsung Electronics led a €200 million (approximately $230.7 million) Series A round for Euclide. The financing also featured participation from Somerset Capital Partners and the Scale-up Europe Fund, managed by European investment firm EQT. Founded in 2024, Euclide designs AI inference chips using an architecture fundamentally distinct from the GPUs that cemented Nvidia’s dominance. The company is rearchitecting both processing and memory/data movement specifically for inference tasks. Euclide aims to release a physical chip by 2028 and onboard thousands of enterprise customers by 2030, though commercial validation remains pending.

Euclide CEO Bernardo Kastrup, a former executive at Dutch equipment manufacturer ASML, highlighted Samsung’s strategic value in a CNBC interview. “Samsung is one of the world's largest memory manufacturers,” he noted, adding that the company “understands the [semiconductor] supply chain and systems well, and has a massive network.” The statement underscores Samsung’s ability to provide operational and supply-chain advantages beyond mere capital. In a separate interview, Kastrup confirmed production details: “The chips are currently being manufactured in Korea. We are leveraging Asian manufacturing lines, just like Nvidia.” He further explained, “We are building a multi-chiplet system.”

Samsung has fast-tracked foundry collaboration with the two-year-old startup, citing high confidence in its potential. Both firms are currently fabricating a first test chip ahead of a 2028 commercialization target. The investment positions Samsung to capture volume production orders for Euclide’s next-generation chips at its foundries, with custom memory supply for these inference accelerators also highly likely. Industry observers view Samsung’s investment as a showcase of its end-to-end “turnkey” capabilities across foundry, memory, and packaging. Analysts interpret the move as a strategic effort to lock in a next-generation European AI semiconductor client across both Samsung’s foundry and memory divisions.

This follows Samsung’s recent investment in Mistral AI, France’s leading AI startup, with which it formed a strategic partnership to co-develop AI models tailored for semiconductor design and manufacturing. From Mistral AI to Euclide, Samsung is executing an aggressive expansion into European AI ventures. From Samsung’s standpoint, the investment also serves as preparation for the evolving AI accelerator landscape. Major tech firms including Google, Meta, and Microsoft are each developing proprietary AI chips to reduce reliance on Nvidia GPUs. By partnering with Euclide, which is engineering inference chips architecturally distinct from GPUs, Samsung gains critical visibility into emerging hardware demand.

Nvidia’s partnership with d-Matrix is widely interpreted as a deliberate strategy to absorb custom AI chip designers into its ecosystem. While GPUs remain dominant in AI training, the inference segment is increasingly viewed as a space where dedicated accelerators can deliver superior efficiency and cost advantages. South Korean semiconductor firms’ early investments are regarded as more than financial wagers; they are strategic maneuvers to secure next-generation memory clients. This outlook is reinforced by projections that as the AI inference market expands, demand for advanced memory architectures like 3D DRAM and PIM will grow in tandem.