Nvidia put $3.5 billion into MediaTek to make custom chips that are not Nvidia chips

Nvidia has invested $3.5 billion in convertible bonds issued by MediaTek, and MediaTek will adopt NVLink Fusion so its customers can build custom accelerators that plug into Nvidia racks. Both facts are in the press release of August 31st. Only the second one is in the tweet.

NVIDIA@nvidia

We’re excited to expand our work with @MediaTek. 🤝

AI factories will be built around differentiated compute — but custom silicon needs a path from chip design to production-scale systems. With MediaTek adopting NVIDIA NVLink Fusion, customers can focus on what makes their XPUs unique and connect them to NVIDIA rack-scale AI infrastructure.
t.co/ey9r7NacA2

What NVLink Fusion sells is everything around the chip you wanted to design yourself. Sixth-generation NVLink across a 72-accelerator domain, with end-to-end XPU-to-XPU latency Nvidia says is three times lower than off-the-shelf Ethernet at ten times the packet rate. NVLink-C2C for the CPU attach, rated at six times the energy efficiency of PCIe. MGX racks, cooling, power, switch trays, NCCL, Dynamo, Mission Control. Build the differentiated part; rent the rest.

And since August 26th, the memory too.

NVHBM moves Nvidia's custom memory controller off the accelerator and into the HBM base die. Traditional stacks put the controller on the XPU, where it eats silicon that could have been compute. Read that 25% figure against a reticle-limited die of roughly 858 mm² and you are talking about a couple of hundred square millimetres handed back — more area than some shipping inference chips have in total. Amazon's Annapurna Labs is first in, and will support NVLink Fusion from Trainium4.

So who is the customer here?

Not MediaTek, exactly. MediaTek is the ASIC house, and its own executive describes the value in the most revealing sentence in Nvidia's August 24th blog: 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." Which is to say the rack is bought first, the alternative silicon arrives later, and Nvidia is paid at both ends. Vince Hu, who runs MediaTek's data centre business, was quoted saying this a week before Nvidia bought $3.5 billion of his employer's paper.

We would read the convertible as the price of a preference, not a rescue. MediaTek has plenty of business. What Nvidia gets is that when a hyperscaler goes shopping for a custom accelerator team, the one that answers has already built its flow around NVLink, NVHBM and MGX — and Nvidia holds a claim that converts into equity if it works. Broadcom, which has no such fabric to sell, is the company that has to answer this.

The honest counter-argument is that NVLink Fusion makes it cheaper to not buy an Nvidia GPU, and that is real. A toll road is a better deal than a wall, and the ecosystem list around it (Intel's data centre silicon group, GUC, Quanta, Annapurna) is not a list of captives. Nvidia's own pitch admits the logic: operators want to "defer the precise silicon mix" while power and buildings get committed. That freedom is worth paying for.

One detail worth filing away. The press release describes NVLink-C2C as connecting XPUs to "NVIDIA Rosa CPUs" — a processor that appears nowhere in the Vera launch material published four days earlier. Watch that name.

Our expectation is that a named hyperscaler XPU built with MediaTek on NVLink Fusion is announced before the end of 2027. If a year from now the only NVLink Fusion silicon anyone can point to is still Trainium4, then this is a defensive standard rather than a business, and the $3.5 billion was an option premium.