How it works

How it works

How work would move from a business to idle devices and back. Nothing here is running yet.

The design

The path of a job

Six stages. Four are designs we intend to build; two still have open questions.

The path of a job: submit, split, match, run in sandbox, verify, settle.1Submit2Split3Match4Run in sandbox5Verify6Settle
  1. Stage 1: Submit

    Planned

    A business would describe a batch job: the software, the inputs, the hardware it needs, and the window it must finish in.

  2. Stage 2: Split

    Planned

    The job would be cut into many small, restartable tasks, each short enough that losing one device costs little.

  3. Stage 3: Match

    Planned

    Each task would go to a device that meets its needs and is available under its owner's limits, in any countries the buyer has allowed.

  4. Stage 4: Run in sandbox

    Planned

    The task would run in an isolated virtual machine with no access to the owner's files or home network, and pause the moment the owner needs the device.

  5. Stage 5: Verify

    Conceptual

    Results would be checked by replication, known-answer tasks or re-execution before anything is accepted. Some workloads have no cheap check.

  6. Stage 6: Settle

    Conceptual

    The buyer would pay in ordinary currency for verified work, and the contributor would be paid from that, less a platform fee. Terms are not set.

Two sides

Two views of the same network

If you contribute a device

  1. 1You would install Jetpack on a supported device and set the hours, power and temperature limits you are comfortable with.
  2. 2Jetpack would benchmark the device, so it is only offered work it can actually do.
  3. 3In the hours you allowed, while you are not using it, the device would take small tasks in a sandbox.
  4. 4You would see what ran, the electricity it used, and what it earned after costs.
  5. 5Payouts would be in ordinary currency. You could stop or uninstall at any time.

If you bring a workload

  1. 1Describe the workload, the hardware it needs, the data involved and when it must finish.
  2. 2We would assess whether it fits, and say so plainly when it does not.
  3. 3If it fits, a small pilot on non-sensitive data, with results you can check yourself.
  4. 4Pricing per verified unit of work, agreed before anything runs.
  5. 5Results delivered as tasks are verified, with job-level progress you can follow.

Failure is normal

When things go wrong

Home devices are unreliable by nature, so the design treats failure as the normal case.

  • A device disconnects

    Its task would return to the queue after a missed heartbeat and run on another device, resuming from a checkpoint where the task has one.

    The design and its limits
  • A result is wrong

    It would fail verification, go unpaid, and count against that device's reputation. The task would be run again elsewhere.

    The design and its limits
  • The owner comes back

    Jetpack would pause the task immediately. The owner's use always wins, and the task would be rescheduled.

    The design and its limits
  • A job misses its window

    Slow tasks near a deadline would be duplicated and the first verified result kept. Until pilots measure real behaviour, there are no completion-time guarantees.

    The design and its limits

Questions

Frequently asked

Is it running?
No. There is no network, no Jetpack app and no marketplace yet. This website and its forms are the only parts that exist.
Is this crypto?
No. There is no token, no coin and no mining. Buyers would pay, and contributors would be paid, in ordinary currency.
What could my device earn?
Often very little, and sometimes less than nothing once electricity is counted. A gaming PC with a recent GPU would do best. CPU desktops, thin laptops, tablets and phones would net about nothing. Try your own numbers in the earnings calculator.
Will it slow my computer?
It is designed not to. Work would only run in hours you allow, within limits you set, and would pause as soon as you use the device. A busy GPU is still warm and audible, so the limits matter.
What kinds of jobs would run?
Interruptible, parallel batch work: AI inference over a queue of inputs, rendering frames, media transcoding, simulations and data preparation. Nothing urgent, stateful or sensitive. The compute page lists what fits and what does not.
Is my data safe as a buyer?
Not safe enough for anything confidential. A device owner controls their own hardware, and consumer GPUs do not offer confidential-computing protection. Early workloads should use data you could afford to have seen. The architecture page explains why.
When does the beta start?
There is no date. A closed pilot with a handful of devices and one buyer would come first, and a beta only if the pilot works.

Which side are you on?

You have a device

Join the early-access list and tell us what you would contribute.

You have a workload

Tell us about it. We intend to say plainly whether it fits.