Nvidia is opening its rack architecture to a competing accelerator vendor, and d-Matrix is the first inference specialist through the door.
The Santa Clara company said it will integrate its next-generation XPUs, starting with Raptor, into Nvidia MGX rack-scale systems using NVLink Fusion. A multi-year roadmap gives d-Matrix silicon a slot alongside Vera CPUs, NVLink switches, BlueField-4 DPUs, ConnectX-9 SuperNICs and Spectrum-X Ethernet.
This collaboration with NVIDIA is a defining moment on our journey to infinite inference, said founder and chief executive Sid Sheth. Jensen Huang framed NVLink Fusion as a path for partners to blend custom silicon with Nvidia’s packaging, rack and networking stack.
Astera Labs is part of the arrangement, supplying connectivity the company says turns novel compute into working AI factories.
The pitch is the premium token economy. As agentic workloads strain inference capacity, service providers want mixed architectures tuned for latency-sensitive jobs such as coding assistants, real-time chatbots and voice agents, where users pay for speed.
For d-Matrix, adopting Nvidia’s supply chain and rack reference design is faster than building an alternative ecosystem from scratch. For Nvidia, letting rivals inside the rack keeps its networking and switching layers at the centre of the AI factory.
