Companies designing custom AI accelerators can now use Nvidia's NVLink Fusion to connect those chips, or XPUs, to Nvidia's wider data-center platform. The pitch is that builders can focus on their processor while reusing established networking, CPU links, rack designs and operating software.
At large scale, a fast chip is only one part of an AI system. Power, cooling, storage and communication between hundreds of processors determine how much useful work the cluster delivers. Nvidia says sixth-generation NVLink supports a 72-accelerator scale-up domain, with lower transfer latency and a higher packet rate than off-the-shelf Ethernet alternatives. NVLink-C2C can also connect an XPU to Nvidia Vera CPUs or other compatible processors.
The practical trade-off is dependence on Nvidia infrastructure around a non-Nvidia accelerator. That may reduce engineering and supply-chain risk, but customers still need to validate cost, interoperability and workload performance. The cited comparisons are Nvidia's own results, not an independent assessment of every competing architecture.