Why real-time work never comes to the network

Teams reading about on-demand GPU compute sometimes assume the network is a place to run everything, including the models that keep a machine moving. It is not, and the reason is worth setting out plainly, because it shapes every other choice in the design.

E31 Network is being built as hourly compute for batch and asynchronous AI jobs. Real-time work is excluded, and the exclusion is not a policy that could be relaxed later. It comes from what the network is.

What real-time means here

Real-time work is anything that must answer now. The examples are familiar: a control loop that keeps a robot's arm where it should be, a perception model that tells a moving machine what is ahead of it, a safety function that has to trip before something goes wrong, an operator tool that has to respond while a person is standing in front of it, an assistant that has to reply while someone waits.

What these share is not a particular kind of model or a particular amount of compute. Some of them are tiny. What they share is a deadline measured in moments, and a cost of missing it that is paid in the physical world or in a person's time.

Why a shared, hourly, queued pool cannot carry it

Three properties define the network, and each one rules out real-time work on its own.

The first is that the pool is shared. Compute held for one job is not available to another, and a job that arrives when the pool is busy waits. A control loop cannot wait. To serve it, compute would have to be held in reserve permanently, which is ownership with extra steps and defeats the purpose of renting.

The second is that the meter runs by the hour, for the hours compute is held. That is a sensible unit for a training run or a footage backlog. It is a meaningless unit for a loop that needs a fresh answer many times a second and never finishes.

The third is that every job passes through a queue when it must. Queues are the mechanism that lets the network schedule mixed job sizes and keep the pool busy without keeping it oversubscribed. A workload that cannot enter a queue would have to be scheduled around every other job, which is to say it would not be a network job at all.

There is also a plainer reason. A real-time path that crosses a network link depends on that link. A robot whose perception runs somewhere else stops perceiving when the connection does. Work with that failure mode belongs on hardware next to the machine, not in a pool rented by the hour.

What the network is instead

Because real-time work is kept off, the network can be described precisely. It is on-demand GPU compute for batch and asynchronous jobs: model training, fine-tuning, retraining, batch inference, large-model inference run asynchronously, video and image processing, large dataset processing, simulation, rendering, synthetic data generation, evaluation, data transformation, and analytics.

It is not a control plane for a facility, not an autonomous-system runtime, not an edge control platform, and not a serving pool inside a building. Those are different systems with different requirements, and the network is not being built to be one of them.

The split that follows

Keeping real-time work local does not mean the network has nothing to do with it. The model that runs on a robot still has to be trained, and later retrained on new data. The footage the robot's cameras collected still has to be processed to find out what went wrong on the days something did. The policy that steers the machine still has to be evaluated across a large sweep of simulated scenarios before anyone trusts a change.

Every one of those is a batch job that can wait for a queue. The design intent is that the network handles them while the real-time path stays where it is. A team ends up with two kinds of compute: dedicated local hardware for the work that must answer now, and rented hourly compute for the work that can run tonight. The workloads page lists which jobs fall on the network side, and the examples page shows what the batch half of that split would look like as submitted jobs.