Do I need to own any hardware to use the network?
Not for the jobs the network is for. Compute is rented by the hour and released when the job finishes. Any real-time work you have still needs something local, because that never goes to the network.
· 3 min read
IN SHORT
No, not for the batch jobs the network is designed for. Compute is held for your job at rent, released at release, and nothing is owned afterwards. You still need hardware of your own for anything that must answer now, because real-time work never goes to the network. Teams that want owned compute in their rack have Czar, the sibling appliance, for that.
No, not for what the network is for. E31 Network is being designed so that a team with a batch job rents compute for that job, runs it, takes the output, and releases the compute. Nothing is owned at the start, and nothing is owned at the end. A team with no GPUs of its own and a model to retrain would, when access opens, describe the job and let the meter run for the hours compute is held.
The longer answer has a condition attached, because not all of a team's work is a batch job.
What you would not need
You would not need a rack, a machine room, or a purchase order for GPUs. The illustrative requests on the examples page are written from the point of view of a team that has none of those: "I need eight GPUs for four hours to process this dataset" asks for compute, not for ownership of it.
You would also not need to keep anything running between jobs. This is the economic case for the network in its plainest form. A team whose need is heavy some of the time and light most of the time, with a model to retrain this week and nothing next week, pays for the compute while it is held and nothing when it is not. Size a rack for the peak and you pay for it through every trough. Rent for the peak and the trough costs nothing. What the bill would look like is covered at what does it cost, with the caveat that no rates are published.
What you would still need
Anything that must answer now. The network takes jobs that can wait for a queue; a control loop, live perception on a moving machine, a safety function, or an assistant that must answer in the moment cannot wait, and none of them are ever submitted. That work runs on your own systems, on edge devices, or on the machine itself, and the hardware for it is yours to provide. The line is drawn on the workloads page and in can I run a robot from the network.
A place for your data and your output. Data is staged for the job and the output is returned to you; where it lives before and after is your side of the exchange.
When owning is the right answer
Some teams want AI compute in their own rack, because their work is steady rather than bursty, or because their data stays where it is. For them Element 31 builds Czar, a sealed appliance sold on its own subscription for a single team, which can be joined into a Czar Grid for a facility. The network is for the jobs that do not need to own compute; Czar is for teams that do. The two are compared directly in how is E31 Network different from Czar.
Neither is sold on this site, and the network has not shipped. But the shape of the choice is clear enough to make now: if the job can run tonight and then stop, you do not need to own the machine it runs on.
Questions this raises
- Do I need a GPU workstation to prepare a job?
- Describing a job, its work, data, request, and output, needs no special hardware. Preparing the data and writing the training or processing code is ordinary engineering work. The heavy compute is what the network would supply.
- If I already own GPUs, is the network still useful?
- It can be, for the peaks. Owned compute sized for the usual week runs out when a retraining run or a backlog arrives, and the network is designed for that overflow. The owned machines keep the steady and the real-time work.
- Does Element 31 sell hardware through e31.network?
- No. This site sells nothing, and the network is for jobs that do not need to own compute. Czar, the appliance, is described on its own site and sold on its own subscription.
KEEP READING