NVLink Fusion: How NVIDIA Turned Its Interconnect Into a Platform

NVLink Fusion connects third-party CPUs and custom ASICs directly to NVIDIA’s NVLink fabric. Image credit: Nvidia This is the fourth article in a series on NVIDIA’s GPU architecture evolution. Earlier pieces: H100 to B200: What Actually Changed covers Blackwell, B200 to Vera Rubin: What NVIDIA Changed Again covers the Vera Rubin platform, and Why NVIDIA Paid $20B for Groq covers the LPU and decode inference. NVIDIA’s acquisition of Groq and its $2B investment in Marvell both point at the same strategy: making NVLink the required interconnect layer for custom silicon.