NVIDIA's NVLink Fusion is designed to let hyperscalers drop XPUs into the same high-speed interconnect network that powers its GPUs, turning a mixed cluster of chips into a single, scalable AI factory. The company frames the technology as a way to combine custom silicon with proven infrastructure, rather than forcing a choice between the two.
One network for mixed chips
XPUs are the catch-all term for specialized processors that aren't CPUs or GPUs — things like inference accelerators or custom ASICs. NVLink, NVIDIA's high-bandwidth interconnect, has long linked its own GPUs together in tight clusters. Fusion extends that same network to XPUs, so a hyperscaler can wire its own silicon alongside NVIDIA hardware and treat the whole thing as one computing pool.
That matters because building a separate network for a few specialty chips is expensive and slow. With Fusion, the interconnect already in place carries the XPU traffic too. The practical upshot: a company can add a custom chip for one specific workload without redesigning the entire data center network.
Why hyperscalers want that
Hyperscalers have been increasingly designing their own silicon for the AI tasks they know best. But those chips rarely live in isolation. They sit in data centers full of other vendors' hardware. The old way of connecting them meant extra switches, more cabling, and a second set of operational headaches.
NVLink Fusion changes the math. It gives a hyperscaler a common, proven networking layer that spans its whole fleet. The company behind it says this lets customers combine specialization with proven infrastructure — meaning they get the cost and performance benefits of their own silicon without giving up the reliability of a mature interconnect.
The AI factory blueprint
The announcement fits squarely into the broader push toward what the company calls AI factories — large-scale data centers built around moving huge amounts of data through GPUs and other accelerators. These aren't ordinary server farms. They're designed from the ground up for training and running AI models, with the interconnect as the backbone.
NVLink Fusion makes that backbone more open. It's a way to let a factory floor hold both standard NVIDIA cards and specialized XPU without forcing the operator to run two separate facilities. For a hyperscaler, that could mean the difference between building a whole new site and simply adding a row of servers.
The real test will come when someone actually stands up a mixed cluster with NVLink Fusion and keeps it running at scale. That's when the promise either holds or doesn't. Until then, it's a piece of paper — or a chip design — waiting for the data center.
