d-Matrix Licenses NVIDIA NVLink Fusion to Scale Raptor AI Accelerators
The AI infrastructure startup will integrate NVIDIA's interconnect and rack designs to enable massive scale-up networking for its upcoming chips.
AI infrastructure startup d-Matrix has licensed NVIDIA's NVLink Fusion interconnect technology and MGX rack-scale reference designs. The move allows d-Matrix to integrate its upcoming Raptor accelerators directly into NVIDIA's established AI infrastructure platform.
By adopting these standards, d-Matrix can utilize the same racks and NVSwitch fabrics as NVIDIA GPUs. The company plans to pair its accelerators with a suite of NVIDIA hardware, including Vera CPUs, NVSwitch appliances, BlueField and ConnectX NICs, and SpectrumX Ethernet products. By the end of next year, d-Matrix expects to offer systems featuring up to 144 Raptor accelerators connected via a single all-to-all NVLink fabric. These NVL144 racks are projected to deliver approximately 2.3 TB of memory capacity and a peak aggregate memory bandwidth of 7.2 petabytes per second.
The Shift to IP Licensing
This partnership is part of a broader strategic pivot by NVIDIA toward an IP licensing model. Through "NVLink Fusion," NVIDIA is extending its proprietary sixth-generation NVLink protocol to non-NVIDIA silicon. This allows third-party AI chipmakers to bypass the immense engineering hurdles associated with designing their own scale-up networks and rack systems from scratch.
NVIDIA has aggressively incentivized this transition through significant financial investments. The company has invested $3.5 billion in MediaTek and $2 billion in Marvell to bring them into the ecosystem. d-Matrix now joins a growing list of partners adopting these standards, including Arm, Qualcomm, Amazon, Fujitsu, and the aforementioned Marvell and MediaTek.
Infrastructure as the New Moat
The adoption of NVLink Fusion signals NVIDIA's transition from a GPU vendor to the foundational infrastructure layer for the AI industry. By licensing its interconnects and MGX designs, NVIDIA ensures that even when customers deploy competing AI accelerators, they remain dependent on NVIDIA's high-margin networking, CPUs, and rack ecosystems.
This strategy effectively creates a "walled garden" for AI compute. While d-Matrix provides the processing power via the Raptor chip, the surrounding environment—the "plumbing" of the data center—remains firmly under NVIDIA's architectural control.
Future Outlook
Industry observers will now watch for the first physical deployments of the NVL144 racks to see if the projected 7.2 petabytes per second of aggregate bandwidth can be realized in real-world workloads. While the integration of Raptor accelerators into the NVIDIA ecosystem is confirmed, the specific performance benchmarks of these systems compared to traditional GPU clusters remain to be seen as the hardware moves toward production.