Nvidia is licensing the wiring between AI chips. Its competitors will pay for it.

A buyer can put a non-Nvidia accelerator next to a non-Nvidia CPU in a non-Nvidia rack and still send a royalty check to Santa Clara. Nvidia's second engine doesn't make chips.

That engine is NVLink, the proprietary chip-to-chip wiring that used to be a moat for Nvidia's own GPUs. At COMPUTEX this spring, Nvidia formally unveiled NVLink Fusion, a licensing program that lets third parties build custom XPUs (industry shorthand for AI accelerators that aren't Nvidia-branded GPUs) and CPUs that plug directly into Nvidia's rack-scale AI factories. The first named adopters: MediaTek, Marvell, Alchip Technologies, Astera Labs, Synopsys and Cadence. Fujitsu and Qualcomm CPUs can also be integrated. Amazon's Annapurna Labs is the first partner on NVHBM, Nvidia's custom HBM (high-bandwidth memory) base die and memory controller.

This is not a marketing round. The receipts are concrete. NVLink 6 plus the NVLink Switch Chip deliver 260 TB/s of aggregate bandwidth in a single 72-accelerator NVLink domain (NVL72), 3.6 TB/s per accelerator, with a roadmap that scales to 1,152 devices and 4x bandwidth efficiency via SHARP FP8 (a collective operation that pushes math into the network switches, here using 8-bit floating-point precision). Nvidia's full networking stack is bundled in: ConnectX-8 SuperNICs (Nvidia's data-center network interface cards), Spectrum-X Ethernet, Quantum-X800 InfiniBand, and co-packaged optics "available soon," delivering up to 800Gb/s of throughput per link. The product page describes the bundle as a complete rack-scale recipe, not a license for one wire.

The licensing perimeter has three layers, and each one extends Nvidia's reach past its own silicon.

The first layer is interconnect. NVLink-C2C extends NVLink as a coherent chip-to-chip link: chiplets (small modular processor tiles packaged together) from different vendors share memory coherently across package boundaries. A custom accelerator from Alchip or a custom CPU from MediaTek can behave, at the system level, as if it were soldered to an Nvidia GPU.

The second layer is memory. NVHBM, Nvidia's HBM base die and memory controller, is now licensable too. Annapurna Labs, Amazon's in-house silicon arm, is the first named partner. The same AWS custom-silicon pitch that competes with Nvidia in the marketplace is now paying Nvidia at the memory layer.

The third layer is the rack. NVLink Fusion adopters plug into Nvidia's MGX rack-scale architecture, the same rack family used by Vera Rubin NVL72. A customer can swap in their own XPU, keep the rack, and reuse the supply chain Nvidia already qualified.

Each layer is its own royalty stream. Stacked, they mean a buyer can order a server with a non-Nvidia accelerator, a non-Nvidia CPU, non-Nvidia memory, and a non-Nvidia NIC and still route revenue to Nvidia.

Jensen Huang called the moment a "tectonic shift": for the first time in decades, data centers must be fundamentally rearchitected around AI. The shift is tectonic for Nvidia's competitors too. Every "Nvidia alternative" pitch in the AI stack now has to clear a quieter Nvidia tax on the wiring, the memory controller, and the rack at the same time.

The Register framed the move as Nvidia building an IP licensing empire on the back of NVLink. The framing is interpretive. Nvidia has not published royalty rates or licensing revenue, and is unlikely to. But the program is public, the partners are named, and the technical specs are concrete enough to defend the mechanism. The strongest falsifier is a quiet rollout: if the named adopters turn out to be marketing chips rather than shipping silicon, the second engine stays a slide deck.

The broader IP posture is consistent. Nvidia and Intel jointly announced development of AI infrastructure and personal computing products, layering x86 IP into the same licensing logic. Cloud providers that pitch sovereign-AI stacks on top of non-Nvidia silicon are now negotiating against the same vendor on the rack fabric that ties everything together.

Nvidia's gravity used to stop at the accelerator. It now extends through the wiring, the memory, and the rack that the rest of the industry builds around.