NVIDIA's NVLink Powers AI Factory Data Centers

NVIDIA has shipped six generations of NVLink since the interconnect debuted in 2016, and the technology now sits at the center of how the company builds what it calls AI factories: data centers designed to run as single, continuous machines rather than racks of separate servers. That shift matters because the workloads have outgrown any one chip's math. Trillion-parameter models, mixture-of-experts architectures, and long-context reasoning all depend on thousands of accelerators trading data with each other in real time, not just crunching numbers in isolation.

Raw FLOPS on a single GPU stopped being the bottleneck years ago. The bottleneck now is how fast those GPUs can talk to each other, how quickly collective operations finish, and how much of that theoretical compute actually turns into usable throughput.

That's the problem NVLink is built to solve. NVIDIA describes it as the scale-up fabric purpose-built for GPU-to-GPU communication at the density and speed AI factories demand, covering training, inference, and other parallel workloads. With millions of accelerators now depending on it, the design choices behind NVLink's sixth generation carry consequences for every operator betting on this infrastructure.