Nvidia has spent years convincing the world it’s a chip company. The reality is more complicated — and more lucrative. A deep dive into the company’s NVLink strategy reveals that even when customers buy silicon from AMD, Broadcom, or custom ASIC makers, Nvidia often still collects a check. NVLink, the high-speed interconnect originally designed to stitch together Nvidia GPUs, has quietly become a licensing goldmine. Server vendors and cloud providers building non-Nvidia accelerators increasingly find themselves paying for the privilege of using Nvidia’s interconnect IP to make their systems competitive.
The mechanics are straightforward: modern AI clusters live and die by how fast chips can talk to each other. Nvidia spent years perfecting NVLink to solve that exact problem, and the patents surrounding it are now impossible to engineer around. A hyperscaler wanting to build a custom accelerator has two choices — pay Nvidia for the interconnect, or accept a performance penalty that makes the whole project pointless. Most choose the former. That’s why Nvidia’s data center revenue increasingly includes licensing fees that never show up in flashy GPU shipment numbers, but show up very clearly in the company’s margins.
What makes this notable isn’t just the revenue. It’s the strategic positioning. Nvidia has effectively turned its interconnect technology into the toll booth for the entire AI infrastructure buildout, regardless of whose chips sit in the rack. Competitors like AMD have their own interconnect story with Infinity Fabric, but the ecosystem around NVLink — the tooling, the software support, the sheer installed base — creates a switching cost that’s brutal to overcome. It’s the same playbook Qualcomm ran with cellular patents: sell the hardware, license the IP, and make sure everyone building anything adjacent has to pay up.
For customers, the implication is uncomfortable. The industry’s desperate attempt to reduce Nvidia dependency by diversifying silicon suppliers doesn’t actually reduce Nvidia dependency at all. It just changes the line item. You might buy fewer H100s, but the licensing fees baked into your alternative hardware still flow to Santa Clara. Jensen Huang has built a business where Nvidia wins whether you buy its GPUs or not — and that’s a position no other company in the semiconductor industry currently holds.
