Cloud infrastructure was built for software. Robots are not software.

A machine acting in the world runs a loop: sense, perceive, plan, actuate. The whole loop has to complete inside the time the physical process allows — a grasp adjusting to a slipping object, a vehicle threading moving pedestrians, an arm tracking a part on a belt. Every stage spends from one budget, and the budget is set by physics, not by product requirements.

A network round trip does not fit in that budget. It is not merely slow; it is unbounded — queueing, transit, load balancing, retries, and someone else’s congestion, none of it under the machine’s control. Worse, its variance is the killer: a control loop can be engineered around a known delay, but not around one that is sometimes fine and occasionally enormous. Determinism, not average speed, is the requirement — and determinism is precisely what a shared wide-area network cannot sell you.

Robot fleets generate continuous, high-bandwidth sensor data: multiple camera streams per machine, depth, force, joint telemetry, all day, every day. Moving that offsite is expensive at a scale that compounds monthly — and that is the easy problem.

The hard problem is that this data is a complete operational record of the facility. Video shows the process, the people, and the layout. Telemetry encodes rates, yields, and methods. For a defense site, a hospital, or a plant whose process is the competitive advantage, exporting that stream to third-party infrastructure is prohibited by policy, by contract, or by law — often all three. The data has gravity. The compute has to come to it.

Cloud architectures assume the connection is always there. The physical economy does not cooperate: ports sit behind cranes and containers that shred radio coverage, mines and ships operate beyond terrestrial networks entirely, forward sites treat an emitting uplink as a targeting beacon, and large stretches of an ordinary factory floor are RF dead zones behind steel and concrete.

For a machine, losing connectivity cannot mean losing capability. A fleet that stops thinking when the uplink drops is not autonomous — it is remote-controlled with extra steps. Whatever intelligence the machine needs to do its job safely has to live where the machine lives.

Modern robotics improves by updating models, and every update changes how a machine behaves near people. That makes the update channel itself safety-relevant: a tampered model is not a data breach, it is a machine doing something nobody intended, with mass and velocity.

So the questions an operator must be able to answer are unforgiving. Where did this model come from? Who approved it? Is this machine running exactly what was approved — and can it prove that, on demand, to an auditor or an incident investigation? Answering those questions requires provenance and attestation built into the distribution path itself, not reconstructed from logs after the fact.