Nvidia Opens Racks to Competitor, Betting on Toll-Collector Revenue

Nvidia is opening its infrastructure ecosystem to d-Matrix, a Microsoft-backed inference chip startup. The move allows d-Matrix's Raptor processors to plug directly into Nvidia-powered systems using NVLink Fusion interconnect. This partnership is expected to result in rack-scale products by 2027.

The NVLink Fusion fabric enables high-speed communication between racks of GPUs and allows other silicon to share memory bandwidth with Nvidia's GPUs and CPUs. For d-Matrix, this connection is crucial for inference accelerators, which rely on quick access to model weights and pass tokens back to the host system.

Nvidia's licensing of NVLink Fusion turns a proprietary bus into an industry standard, potentially increasing revenue from licensing fees. The company has already partnered with Marvell and Groq through NVLink Fusion, with CEO Jensen Huang stating that 'AI has reached its inflection point' and compute is now revenue-driven.

Rack economics reinforce the idea that Nvidia earns more as a toll collector for AI data centers than it would by trying to win every chip sale. The company's data-center revenue per gigawatt is estimated at $18 billion, $25 billion, and $40 billion for Hopper, Blackwell, and Vera Rubin respectively.