Goldman Sachs estimates the AI networking TAM could rise from about $15B for GB300 NVL72 to $154B for Rubin Ultra NVL576 Spec B.
That is a massive change in where AI infrastructure dollars may go.
The obvious AI infrastructure story has been GPUs. But as clusters get larger, the networking layer starts to matter much more. More chips need to talk to each other. More racks need to move data with lower latency. More bandwidth means more optics, more copper, more fiber, more CPO components, and more complexity inside the data center.
The mix also changes.
In Goldman’s estimate, optical modules are still a large pool at roughly $32B in the Rubin Ultra NVL576 Spec B case. But the biggest swing comes from CPO-related components and fiber, which Goldman estimates at about $91B in that configuration.
Copper also does not disappear. It grows from about $6.7B in the GB300 case to about $30.8B in the high-end Rubin Ultra case.
The main point: AI infrastructure spending is not only about the accelerator anymore. If these rack architectures play out, networking becomes one of the largest incremental value pools in the AI buildout.