ARCHITECTURE
The facility is the datacenter
A working fleet generates its most valuable training signal on the wrong side of the boundary for any cloud to use. The case for treating the facility itself as the unit of compute, coordination, and learning.
· 7 min read
Ask where a robot fleet's intelligence should live and the industry's default answer is inherited, not derived: in the cloud, because that is where AI infrastructure already is. Derive the answer instead — from latency, from data, from learning — and it comes out differently. The natural unit of physical AI is not the region or the machine. It is the facility.
Three derivations, one answer
Coordination is facility-shaped. A fleet is not a set of independent machines; it is a system with shared state. Machines allocate tasks among themselves, deconflict paths, and act on a common picture of a space that changes by the minute. That shared world model has a natural scope — this floor, this yard, this site — and a natural latency requirement: decisions about the next few seconds, made continuously, for every machine at once. Put that state in a remote region and every decision inherits a round trip plus the failure modes of the path. Put it on the site's own switches and coordination runs in local time under local control.
Data is facility-bound. A fleet's sensor output is continuous, high-bandwidth, and confined — video of the process and the people, telemetry that encodes rates and methods, layouts that map the operation. Moving it offsite is expensive every month and prohibited in exactly the industries that need fleets most. The data's boundary is the facility's boundary. Compute that wants to touch the data has to stand inside the fence with it.
Learning is facility-specific. This is the derivation that gets missed. The gap between a robotics demo and a robotics deployment is the long tail of this site — its lighting at shift change, its pallet dialects, its seasonal product mix, its floor's particular chaos. The training signal for closing that gap is generated fresh every shift, by the fleet, inside the boundary. It is the most valuable data that will ever exist for making this fleet better at this facility's work, and it is unusable by any learning loop that requires it to leave.
Three independent constraints; the same conclusion. The facility is where the compute belongs.
The overnight loop
Taking that conclusion seriously produces a rhythm rather than a diagram.
During the shift, edge nodes keep their machines inside their control budgets, and the facility plane coordinates: allocating work, maintaining the shared world model, answering the inference queries too heavy for the edge. The plane also does what nothing outside the boundary can do — it keeps the day. Episodes, edge cases, near-misses, corrections: the raw material of improvement, retained at full fidelity because it never has to survive an uplink.
After the shift, the facility plane turns that material into capability: scheduled adaptation on the day's data, evaluation against the site's own scenarios, and staging of the result for the next shift — all inside the fence. The fleet gets measurably better at this facility's work, and the facility's data participates in exactly none of the infrastructure it distrusts.
Across sites, what travels is the governed minimum: signed models and policies inbound, attestation evidence and audit records outbound. Improvements can propagate between facilities as artifacts — reviewed, signed, distributed — without any site's operational data pooling anywhere. Sites share capability, not exposure.
What this is not
This is not edge computing with a bigger box, and not a datacenter with a shorter cable. The facility plane differs from both in what it refuses to assume: no runtime dependency on external networks, no unsealed general-purpose substrate accumulating configuration drift, no learning loop that launders data across the boundary. It is being designed as sovereign infrastructure in the specific sense that the facility's owner holds every key that matters — over the data, over the models, over what runs.
The datacenter was the right unit for software. For machines that work in buildings, the building is the unit. E31 Network is the infrastructure that takes that sentence literally.