Nvidia is making money from the chips built to replace it
Its $3.5 billion MediaTek investment buys a promise: the custom accelerators MediaTek designs for hyperscalers will run on Nvidia's fabric, memory and racks.
Nvidia has found a way to profit from the chips built to replace it.
On 31 August it announced it had bought $3.5 billion of convertible bonds from MediaTek, the Taiwanese firm that designs custom accelerators for cloud companies that would rather not buy Nvidia's own. In the same announcement, MediaTek agreed to build those accelerators on NVLink Fusion, which means they will plug into Nvidia's interconnect, sit in Nvidia's racks and, as of last month, use Nvidia's memory controller. The customer walks out of the GPU showroom and finds Nvidia holding the door.
Only one of those two facts made it into Nvidia's post on X, which talked about "expanding our work" with MediaTek. The investment is in the press release.
What a rival chip still has to rent
NVLink Fusion is everything around the part a hyperscaler wants to design itself. Nvidia says its sixth-generation NVLink joins 72 accelerators with a third of the latency of off-the-shelf Ethernet. NVLink-C2C connects the chip to the CPU at what Nvidia rates as six times the energy efficiency of PCIe. Then come the racks, the cooling, the switch trays and the software that schedules the whole thing.
Since 26 August there is memory too. NVHBM moves Nvidia's memory controller off the accelerator and into the memory stack, which Nvidia says frees up to a quarter of the chip's area for compute. Amazon's chip unit, Annapurna Labs, is the first taker and will support NVLink Fusion from Trainium4.
Why MediaTek is the right partner to pay
MediaTek's own executive gave the game away a week before the money arrived. Vince Hu, who runs its data-centre business, said customers "can deploy their rack-level solution with the NVIDIA GPU, and then they can decouple the development of their XPU and put it at a different pace." Buy the rack first. Bring your own silicon later. Nvidia is paid at both ends.
So who is paying whom? The convertible is the price of a preference. When a hyperscaler goes looking for a team to build its custom chip, Nvidia wants the team that answers to have built its whole flow around NVLink, and it wants a claim that turns into equity if that works. Broadcom, which designs the same kind of chips and has no fabric of its own to sell, is the company that has to respond.
The fair objection
A toll road is still cheaper than a wall. NVLink Fusion makes it easier, not harder, to run something other than an Nvidia GPU, and the partners around it, Intel's data-centre silicon group among them, are not captives. Nvidia's own pitch is that operators can commit to buildings and power now and "defer the precise silicon mix". That freedom is worth paying for, and Nvidia has decided it would rather sell it than fight it.
Watch for a named hyperscaler chip built with MediaTek on NVLink Fusion by the end of 2027. If the only one anyone can point to is still Trainium4, the $3.5 billion bought an option, not a business.
