Nvidia Puts $3.5 Billion Into MediaTek to Keep Custom Chips Inside Its Own Racks

Nvidia and MediaTek announced an expanded partnership on Monday, and the eye-catching line came near the bottom of the release: Nvidia has bought 3.5 billion dollars of convertible bonds issued by MediaTek. The substance of the deal is that MediaTek will adopt NVLink Fusion, Nvidia’s platform for letting other companies’ custom chips slot into Nvidia’s own rack-scale systems.

The collaboration covers three areas. In data centres, MediaTek will offer NVLink Fusion as a design foundation for customers building custom AI accelerators, generally called XPUs. In local computing, the two continue their work on the chips behind Nvidia’s DGX Spark and RTX Spark, small machines meant to run AI models on a desk rather than in a data centre. In cars, MediaTek’s Dimensity Auto platforms keep integrating Nvidia technology. Jensen Huang’s quote frames it as giving customers “the freedom to create differentiated AI systems,” and MediaTek CEO Rick Tsai points to his company’s position in custom silicon. What NVLink Fusion actually supplies is the unglamorous surrounding infrastructure: the interconnect chiplet, the chip-to-chip links, the memory architecture, the packaging, the rack-scale qualification. A customer brings the compute design; Nvidia and MediaTek supply everything around it.

Read it as defence, not generosity. Amazon, Google, Microsoft, OpenAI and Anthropic are all building their own AI chips, for the obvious reason that buying every GPU from one supplier is expensive and strategically uncomfortable. Nvidia cannot stop that. What it can do is make sure that when a company designs its own accelerator, the thing still lives inside an Nvidia rack, speaks Nvidia’s interconnect, and depends on Nvidia’s software. Losing the chip while keeping the system is a much better outcome than losing the customer. The 3.5 billion dollar investment invites the other reading that analysts have been applying to Nvidia deals all year: money flows out to a partner, and products flow back in the other direction. It is not improper, and it is not unusual in semiconductors, but it does make revenue growth harder to read from the outside, and Nvidia paused a separate 36 billion dollar cloud financing programme last week reportedly over antitrust concerns.

What this means for you. Directly, nothing, unless you buy data centre hardware. Indirectly, two things are worth noting. The local computing part of this deal is the same silicon line that ends up in machines built to run models on your own desk, which is the slow, quiet story behind AI that does not need a subscription. And the broader pattern shapes what everything costs: the more the AI stack consolidates around one company’s interconnect, the less price pressure there is downstream, and eventually that reaches the monthly bill for the tools you actually use. Nothing changes this quarter. It is just useful to know who is holding which end of the cable.

Sources

• NVIDIA and MediaTek Deepen Long-Standing Partnership (NVIDIA Newsroom)
• Nvidia Opens NVLink to Custom XPUs as AI Goes Heterogeneous (Data Center Knowledge)