BlogApplications4 min read
What goes wrong at the rack, and how touch fixes it
AI racks are packed with cables, fiber and coolant lines, and most of the trouble starts at a single connection. Here is what goes wrong, why cameras miss it, and what a sense of touch changes.
A modern AI rack is a dense bundle of connections. One NVIDIA GB200 NVL72 rack carries about 5,000 copper NVLink cables, roughly two miles of them.1 Add fiber to every network port and, increasingly, a coolant line and quick-disconnect at every server. Each of those connections is made by a person, and each one can go wrong in ways that are hard to see.
This post walks through the problems we hear about most, and what a robot that can feel changes about each one.
The problems at the connection
Connectors that look seated but aren’t. A connector can sit flush with its latch never engaged. It passes a visual check, the rack powers on, and the fault shows up later as a flapping link or a dead port. Somebody then has to trace it back to one connection among thousands.
Ports nobody can see. Racks are packed tight. Latches sit behind connector bodies, ports sit behind other cables, and the person doing the work often goes by feel anyway. A camera mounted on a robot has the same problem, only worse.
Fiber and coolant lines that punish the wrong force. Push fiber too hard and you damage the end-face or the transceiver. Leave a coolant quick-disconnect short of fully engaged and it can leak. Both mistakes come down to force, and both are invisible from outside.
Dirty end-faces. Connector vendors citing an NTT Advanced Technology study rank contaminated connector end-faces as the leading cause of network faults.3 Clean handling and inspection at the moment of connection matters as much as the connection itself.
Human error, at scale. None of this is about careless people. It is about repetitive, high-precision work done thousands of times per rack, under schedule pressure. Uptime Institute reports that human error plays a part in roughly two-thirds to four-fifths of major outages.2
No record of how the work was done. Once a rack ships, there is no evidence of how each connection was made. If a fault appears, there is nothing to check against.
Why a camera isn’t enough
Cameras are good at finding the port and planning the route. They are poor at the moment that decides the outcome: contact. The difference between seated and half-seated is a click and a drop in force, often behind the connector housing. The difference between a good insertion and a damaging one is how hard you pushed. Neither shows up in an image.
| Task | What a camera sees | What touch tells you | If it’s missed |
|---|---|---|---|
| Seating a connector | Looks in place; the latch is hidden | The click: a sharp drop in force | A half-seated connector that fails later |
| Inserting fiber | The port and the plug | Insertion force, and when to stop | A damaged end-face or transceiver |
| Coupling a coolant line | The fitting, roughly aligned | Full engagement of the quick-disconnect | A slow leak |
| Routing a cable | The path | Tension, snags and bend | A strained or pinched cable |
What touch changes
A robot with fingertip touch can work the way a skilled technician does: by feel.
It confirms every connection. Our fingertips sense force and vibration. When a latch engages, there is a click the fingers can feel and a characteristic drop in force. The robot looks for both, then confirms with a gentle pull. If the click never came, it backs off and re-seats before moving on.
It finds ports where it can’t see them. Touch lets the robot feel the edge of a port and slide into it, the same way a person does when the port is hidden behind a bundle of cables.
It controls force on fragile parts. Grip and insertion force are measured and controlled in real time, so fiber is seated without being crushed and quick-disconnects are pushed until they are fully engaged.
It leaves a record. Every connection comes with its own force trace. A missing click is caught at install, not at the next outage, and there is evidence to check if a fault appears later.
It adds capacity where hiring is hard. Repetitive connection work moves to robots, and skilled technicians spend their time on the jobs that need judgment.
Where we are
We are a research-stage company. Our fingertip sensors are built, and our touch-aware model is in development. In simulation, we have shown the core loop on an ethernet plug: given the same depth error from vision, the vision-only controller moves on with the latch unengaged, while the touch-aware one notices, re-seats and verifies with a pull. We have not yet deployed in a data center, and we will say so plainly until we have.
If you build racks or run data halls and connection work is where your quality and staffing problems meet, we would like to see how it is done on your floor.