d-Matrix on Thursday said it will license Nvidia’s NVLink Fusion interconnect and ship its next-generation Raptor inference XPUs inside Nvidia MGX rack reference designs, The Register and The Next Platform reported. The partnership puts specialized inference silicon onto the same liquid-cooled NVL144-class MGX racks, NVSwitch fabric, and supply chain used for Nvidia’s own AI factory stack, rather than requiring a separate scale-up network for each accelerator vendor.
Raptor is due to tape out by year-end and reach systems in the fourth quarter of 2027, according to The Next Platform’s account of a company briefing. d-Matrix plans to pair Raptor trays with Nvidia Vera CPUs, BlueField-4 DPUs, ConnectX-9 SuperNICs, and Spectrum-X Ethernet, and said a single MGX rack can hold 144 Raptor XPUs on an all-to-all NVLink domain. Nvidia’s blog framed NVLink Fusion as letting XPU makers plug into validated rack, power, and cooling designs while still differentiating on silicon; d-Matrix CEO Sid Sheth said the stack gives customers a faster, lower-risk path to ultralow-latency inference.
The move adds another inference player to Nvidia’s growing NVLink Fusion licensee list alongside names such as Qualcomm, Arm, Marvell, Amazon, Fujitsu, and MediaTek, The Register noted. Raptor racks can also sit beside Vera Rubin GPU systems for disaggregated prefill and decode workloads. This brief covers the September 10 licensing and rack partnership; it does not benchmark Raptor against GPUs or audit customer deployment timelines.