Nvidia's moat just moved from silicon to fabric

The custom-AI-silicon race no longer runs around Nvidia. It runs through it. Every time a hyperscaler designs an accelerator of its own, that chip still has to talk to something. Nvidia is now being paid to be that something.

The $3.5 billion convertible bond into MediaTek is the first capital move of that strategy. MediaTek is not a phone-chip maker crossing into data centers; it is a phone-chip maker Nvidia needs to keep on its side of the rack. The partnership is built on NVLink Fusion, Nvidia's interconnect fabric, and it extends to PC and automotive silicon. Read the deal in plain terms and the message is straightforward: "custom AI chips" and "Nvidia-free" are no longer the same thing. Jensen Huang calling MediaTek "one of the world's great semiconductor companies" is the courtesy; the corporate action is the argument.

The mechanism is portable. As long as custom silicon has to integrate with the dominant training and inference fabric, the fabric is the moat, and capital is how you widen it. Hyperscalers will keep designing their own chips. Nvidia will keep paying to make sure those chips still need its plumbing. The race was never only about chips. It is about who owns the layer everything plugs into.