d-Matrix, a U.S. fabless AI inference-chip startup valued at $2 billion, is integrating its Raptor XPU into Nvidia’s data-center AI infrastructure through NVLink Fusion, which links the accelerator with GPUs, CPUs and networking equipment. Nvidia is using the partnership to accommodate dedicated inference chips alongside its general-purpose GPUs, targeting better efficiency and lower costs as inference workloads expand. South Korean investors SK Hynix, Samsung Venture Investment and Mirae Asset backed d-Matrix from its early financing rounds, reflecting strategic interest in technologies that bring memory closer to computation. d-Matrix’s Digital In-Memory Computing and 3DIMC architectures are designed to reduce data-movement distances and address the Memory Wall, a bottleneck created by the growing cost of moving data between memory and processors. The company says 3DIMC can reduce power consumption to one-tenth that of HBM4, while analysts expect it to add to, rather than replace, demand for HBM and other AI-memory products. Samsung is also investing in Dutch inference-chip startup Euclide, which aims to produce a physical chip by 2028 and reach thousands of enterprise customers by 2030. Samsung and Euclide are already fabricating a first test chip, and the investment could support future foundry, memory and packaging business. The moves come as Nvidia, Samsung and other major technology companies position themselves for an AI-inference market increasingly open to specialized accelerators.