SANTA CLARA, 1 SEP 2026 — MediaTek has become the latest licensee of NVLink Fusion, Nvidia's chip interconnect, joining Amazon, Fujitsu, Qualcomm, Arm and Marvell. Nvidia is not selling licences to make money from licensing. It is making its fabric the substrate that competitors' accelerators plug into.
What is being licensed
NVLink Fusion comes in two forms. The chip-to-chip variant connects CPUs to GPUs or other accelerators in a memory-coherent fabric. The switched-fabric variant stitches many accelerators into what functions as a rack-scale processor.
Nvidia also offers NVHBM, a custom high-bandwidth memory base with an integrated memory controller, claiming it frees 25 per cent of die area for compute, cuts high-bandwidth memory power by 15 per cent and raises bandwidth by 30 per cent. Those are vendor figures with no independent verification available.
Licensing an interconnect is a standards play
A licensing empire implies revenue from licences. The strategic value here is somewhere else.
Custom accelerators are the main threat to Nvidia's position. Every hyperscaler building its own silicon is a customer buying fewer GPUs. Nvidia cannot stop that, and licensing NVLink means those accelerators connect using Nvidia's fabric. A rack full of somebody else's chips still runs on Nvidia's interconnect, and where a switched fabric is involved, on Nvidia's switches.
The quarterly networking revenue above US$15bn is what that looks like once it works. Networking was a rounding error in Nvidia's business a few years ago; it is now a large division. It grows with total accelerator deployment, not just Nvidia's share of it.
Dion Harris, Nvidia's senior director of HPC and AI hyperscale infrastructure, describes the platform as providing flexibility and fungibility to meet customers where they are. That is the point: fungibility at the accelerator layer, with the layer underneath fixed.
Why the interconnect is the layer worth owning
Interconnects sound like plumbing, and this one is strategically valuable.
Training a large model means many accelerators working on one problem, which means constantly moving parameters and gradients between them. Past a certain scale the limiting factor stops being how fast each chip computes and becomes how fast they can talk to each other. A cluster with faster chips and a slower fabric loses to the opposite arrangement.
Memory coherence is the specific property being sold. In a coherent fabric, a processor can address memory attached to another processor as though it were local, and the hardware keeps the copies consistent. That removes explicit copying from the programmer's problem and is what allows a rack of separate chips to be programmed roughly as one large machine.
Standard alternatives like PCIe and Ethernet exist and are slower for this purpose. PCIe was designed for peripherals and Ethernet for networks between machines, not inside one. Both work; neither was built for the traffic pattern that rack-scale training produces, which is why proprietary fabrics emerged at all.
The MediaTek investment is not a licensing fee
Nvidia put US$3.5bn into MediaTek convertible bonds. Read the direction of that money carefully.
Ordinarily the licensee pays the licensor. Here the licensor has invested billions in the licensee, which makes this something other than a licensing transaction. Nvidia is capitalising the companies that adopt its interconnect, which accelerates adoption and gives it an equity interest in the outcome.
Convertible bonds are the specific instrument and they carry an option. Debt now, with the right to convert to equity later, means Nvidia takes limited downside if MediaTek's accelerator business disappoints and participates if it succeeds. That is a rational structure for funding a partner whose success you are trying to engineer.
The competitive response is the thing to watch
Broadcom is the competitor named against this, and the contest is over which interconnect the industry standardises on.
The Ultra Accelerator Link consortium offers an open alternative. It exists because other firms would rather not concede control of the rack-scale fabric to the dominant accelerator vendor. Whether NVLink Fusion or an open standard wins determines who controls the layer every AI rack depends on for the next decade.
Nvidia's advantage is that its fabric works now, at scale, with a software stack that assumes it. The open alternative's advantage is that nobody has to ask permission. Historically the first has won early and the second has won eventually, and the interval between those has usually been long enough to matter enormously.
What this means for buyers in this region
Data centre operators across Singapore, Malaysia and Indonesia are making rack-level purchasing decisions now that will outlast this contest, and the interconnect is the part that is hardest to change later.
An accelerator can be swapped at refresh. A rack-scale fabric is cabling, switches, power topology and a software stack tuned around it, and replacing it means rebuilding the hall. So a decision that looks like choosing a chip vendor is often a longer commitment to an interconnect than to the silicon it connects.
The practical question for a regional operator is not which accelerator is fastest. It is whether the rack design admits accelerators from more than one vendor over its life, and NVLink Fusion is an answer to that — provided you accept whose fabric it is. We reported that Nvidia's 30x per-megawatt claim is measured against its own previous generation, and the same discipline applies here: read the licensing announcement as a description of where the lock-in moved, not of where it was removed.
