NVLink Fusion addresses enterprise demand for flexibility. NVIDIA is evolving from GPU vendor to data center architect.
The NVLink Fusion announcement was a strategic inflection that’s been telegraphed for two years. NVIDIA has spent its entire AI cycle preserving the integrity of its end-to-end stack — GPU, NVLink, NVSwitch, software, networking — and pushing customers toward a single-vendor architecture.
NVLink Fusion is the opposite move. It opens NVLink to third-party silicon, starting with Marvell, and signals NVIDIA’s willingness to be a layer in a heterogeneous AI infrastructure rather than the only layer.
Why now? Because the hyperscalers are not going back to single-vendor. The captive XPU programs are real, the AMD GPU pipeline is real, and the in-rack custom silicon programs are accelerating. NVIDIA’s choice is between watching its share of in-rack interconnect collapse to Ethernet/Infiniband alternatives, or making NVLink the open standard for in-rack accelerator-to-accelerator fabric.
By opening the protocol, NVIDIA reframes the competitive question. Instead of ‘NVIDIA full-stack vs. AMD/in-house silicon’, the question becomes ‘whose interconnect is at the center of the heterogeneous rack?’ That’s a position NVIDIA can win — and it’s the position of a datacenter architect, not just a chip vendor.
