Building on our coverage from yesterday, NVLink Fusion is being opened so custom AI accelerators can plug into the same rack-scale systems that run today’s largest models, while NVIDIA has backed MediaTek with a major financing commitment to build those chips.
NVIDIA and MediaTek have expanded their partnership so customers can design custom XPUs, integrate them into MGX rack architectures, and keep using NVIDIA’s NVLink-based interconnect. As part of the deal NVIDIA bought convertible bonds in MediaTek totaling $3.5 billion to support chip development and advanced packaging.
Why this matters: it changes the economics and the playing field for hyperscalers. Until now, building at-scale AI hardware meant either buying NVIDIA GPUs or creating a completely separate stack. Opening NVLink Fusion means companies can pursue custom silicon without tearing out their networking and rack design.
How it works, simply: NVLink Fusion is a high-speed rack interconnect and MGX is a pre-validated rack blueprint. Think of it as standard power and plumbing for data centers. If everyone uses the same sockets, different appliance makers can still ship devices that fit and work together. That common plumbing is what makes third-party XPUs practical at scale.
What changes now is concrete. Large cloud providers and frontier-model developers get a clearer, supported path to custom accelerators. Chip designers like MediaTek gain privileged access to those customers. Practical limits remain: these are rack-scale systems that require advanced packaging and manufacturing. This isn’t something you run on a laptop tomorrow.
What to watch next: will these MediaTek-built XPUs match NVIDIA’s performance in independent benchmarks, and will customers actually deploy them in production racks? The answers will determine whether this is ecosystem evolution or just a paper agreement.