Nvidia's NVL72 has two miles of copper inside it

In 1969, a young engineer named Bob Metcalfe was staring at a problem nobody wanted to solve. The ARPANET could move packets between machines that sat in different buildings, but inside those buildings, the actual wires connecting computers to each other were a mess of incompatible standards, proprietary cables, and hand-rolled kludges. The glamour work in computing was CPU design. Nobody wanted to build cables. Metcalfe eventually invented Ethernet, cofounded 3Com, and made a fortune solving the boring problem everyone had ignored.

We are living through the exact same moment right now, except the wires cost thousands of dollars each and there is a physical shortage of the people who know how to design them.

Here is what has changed in the last eighteen months. When Jensen Huang stood on stage at GTC 2024 and unveiled the GB200 NVL72 rack, most of the coverage focused on the compute. 72 Blackwell GPUs in a single rack, 1.4 exaflops of AI performance, the usual Jensen theater. The number almost nobody in the mainstream press wrote about was two miles. That is the length of copper cabling packed into a single NVL72 rack. Roughly 5,000 individual cables. Each one carrying 224 gigabits per second of signal integrity across distances where the laws of physics start to actively fight you.

This is the part of the AI buildout that gets waved away as 'infrastructure' or (worse) 'plumbing.' It is neither. It is a specific engineering problem, and it is getting harder every quarter.

When an Nvidia GPU wants to talk to another GPU one rack away, the signal has to survive a gauntlet: the die-to-package interconnect, the package-to-board interconnect, the board-to-connector interconnect, the connector-to-cable interconnect, the cable itself, then the whole sequence in reverse. At every step, the signal loses integrity. At 224 Gbps PAM4 signaling, the margin for error is measured in picoseconds and millivolts. A cable that is 3% out of spec doesn't cause a slow network. It causes a training run on a $500 million cluster to silently corrupt gradients for six hours before someone notices.

The hyperscalers figured this out around eighteen months ago. Meta, Microsoft, Google, and Amazon all restructured their component qualification teams around a single question without saying much publicly about it: who can build this stuff at the tolerances we need, at the volumes we need, at the reliability we need?

The answer is a very short list of companies. Most of them are not household names. Some of them are 90 years old. All of them share a peculiar characteristic. They spent decades building unglamorous components for defense, aerospace, industrial equipment, and old-school telecom, and it turns out the paranoid engineering culture required to make a connector that works reliably inside a fighter jet is exactly the paranoid engineering culture required to make a connector that works reliably inside an AI training cluster.

There is a second-order effect that almost nobody is pricing in. When Nvidia moves to Rubin in 2027, the signaling rate goes from 224 Gbps to 448 Gbps. When they move to the generation after that, it goes to 896 Gbps. At those speeds, copper starts to lose the fight against physics for anything beyond about a meter, which means optical interconnects have to eat more of the rack. Which means the connector companies also need to be optical companies. Which means the small number of firms that already do both, at scale, with the qualification history to get through hyperscaler procurement, become genuinely irreplaceable.

The third-order effect is the one I find most interesting. Robotics is arriving at the same problem from the opposite direction. A humanoid robot needs hundreds of sensor interconnects operating in a vibration environment that would destroy most data center hardware, at power efficiencies that would horrify a hyperscaler, and at unit volumes that are about to go from tens of thousands to tens of millions. The engineering culture that solved the data center problem is the same engineering culture that will solve the robotics problem. The customer list overlaps. The revenue compounds.

So when you look at the AI compute buildout, do not look at the GPU. Everyone is looking at the GPU. Look at what connects the GPU to everything else. Look at who builds the specific physical objects that carry the signal between the silicon and the world. The market cap of the pure interconnect winner in the last cycle, Cisco in the late 1990s, briefly exceeded General Electric.

There is one company that sits at the exact intersection of every trend I have described above. Ninety-four years old. Boring name. Boring products. Approximately 300,000 different SKUs. Just closed one of the largest acquisitions in its history. And its AI-related revenue is compounding at a rate that would make a software company blush.