How is E31 Network different from Czar?

They are two shapes of the same company's answer to one question: where does a team's AI compute live. Czar is the owned shape. It is a sealed appliance that sits in your own rack, sold on its own subscription, for a single team, or joined with others into a Czar Grid for a facility. E31 Network is the rented shape. It is compute by the hour for batch jobs, held for a job from rent to release and then returned to a pool, with nothing owned before or after. This site describes the network and sells neither.

Owned versus rented

With Czar, the compute is yours for the term of the subscription, in your building, available whenever you reach for it. That suits work that is steady, work that must stay inside your walls, and work that must answer now. A control loop or a live perception model can run on compute you own because it is there at every moment.

With the network, compute exists for you only while a job holds it. You describe the work, the data, the request, and the output; compute is rented, the job runs unattended and isolated, the output comes back, and the compute is released. Between jobs you hold nothing and pay nothing. The meter runs for rent and run only. The how it works page shows those five steps; rates for them are not yet published.

Steady versus bursty

The deciding variable is the shape of your demand over time. If your need for compute is roughly the same every week, owning it is the natural fit, and idle time is small because the machine is busy. If it spikes a few times a quarter and sits low in between, owning for the peak means the machine is paid for while it waits. The network is being built for that second shape: a model to retrain this week, a backlog of footage, a sweep of simulations, a dataset to process once.

Many teams have both shapes at once. The steady part, and anything real-time, is a case for owned compute. The spikes above it are the case for renting by the job. The question of whether you need hardware at all is answered at do I need to own any hardware.

What the network will not do

The network never takes real-time work. Robot control, live perception, safety functions, operator tooling, and assistants that must answer in the moment stay on local infrastructure, and an appliance in your rack is one form that infrastructure can take. The network is not a facility control plane or an edge platform, and the program is careful to say so. The workloads page draws the line.

What is not published

For the network: rates, where the capacity comes from, retention specifics, and the scheduling of mixed job sizes. For Czar: this site does not describe its pricing or its internals, and the place to read about it is e31.io/czar. What can be said plainly is the shape of each, and the shape is enough to tell which one a given job belongs on.