Voyager Cloud

Limits and failure modes

Read this before a long job. It lists when a machine stops by itself, what a stop keeps, when work can be lost, and what each plan and computer can run.

Checked against the service on November 18, 2026. For agents: this page as Markdown.

Plans and GPU caps

Cloud GPUs need a paid workspace, Starter or above; a Starter trial counts. A lease from a Free workspace is refused with a link to upgrade.

GPUs per machine and machines at once, by plan
PlanGPUs per machineMachines at once
FreeNoneNone
Starter11 at a time
CreatorUp to 216 per workspace member
UltraUp to 416 per workspace member
GigaMax™Up to 416 per workspace member
  • GPUs per machine

    Your plan sets how many GPUs one machine can have (table above). render-2x and render-4x need a plan with 2 or 4.

  • Machines at once

    Starter runs one machine at a time. Creator and above run up to 16 per member of the workspace. Paused machines don’t count; resuming one does.

  • No queue for a card

    A start waits for a free card, never for a place in line. If no host has your card free, Voyager may try a bigger card at your card’s price. A start that isn’t ready after 10 minutes fails, bills nothing, and lists up to three other cards to try.

When a machine stops by itself

Every stop below pauses the machine: its disk is saved as a snapshot, the GPU is released and billing stops. voyager cloud resume <machine> (or any command that needs it) starts it again from that snapshot.

  • Idle for 15 minutes

    A machine pauses after 15 minutes idle: no running job, no command or API call, no file transfer and no GPU busy above 10%. Set another limit with voyager cloud lease --idle-timeout <seconds> (30 seconds to 7 days).

  • A busy GPU counts as work

    On current machine images, training or a render started from a shell keeps the machine up while any of its GPUs is above 10% busy, even with no Voyager job and no calls. A machine that only waits on the network or the CPU still pauses.

  • About 24 hours per start

    A machine pauses after about 24 hours running, like any other stop. Resuming it starts a new 24 hours from its disk, so long work should save its state to disk as it goes.

  • Budget and credits

    With --budget the machine pauses just before it would pass that amount. Without one it draws on the workspace’s credits and pauses when they run out.

  • Unresponsive

    A machine whose calls have all failed for 10 minutes, with no job running, pauses the same way.

When you can lose work

  • Only the disk is kept

    A pause or a snapshot keeps files, never memory, running processes or GPU state. After a resume, start your processes again; save checkpoints to disk if a job must survive a stop.

  • A failed machine loses work since the last snapshot

    If the GPU host ends a machine (out of memory, a hardware fault, the host reclaiming it), Voyager starts a replacement from its last snapshot. Files written after that snapshot are gone, and with no snapshot at all /project starts empty. Jobs that were running end as lost. Take a snapshot after any step you would not want to repeat: voyager cloud snapshot <machine>.

  • Release deletes the disk

    Pull what you need before voyager cloud release. After it, only your snapshots and volumes remain.

  • Snapshots expire

    Snapshots last 30 days, and a workspace keeps at most 100 it asked for (the ones stops take don’t count). voyager cloud snapshot list shows when each one expires.

  • Volumes are separate, and slow for many small files

    A volume (/persist, attached with voyager machine lease --volume in the desktop app or through the API) outlives machines, but a snapshot doesn’t include it and restoring one doesn’t roll it back. A freshly attached volume reads many small files very slowly: importing a Python environment from one took 16 minutes. Use it for data, outputs and checkpoints; keep Python environments and caches on the machine’s own disk. A machine with a volume can start on fewer kinds of hosts, so a busy card has fewer fallbacks.

Cards and software

  • Vork needs ray tracing

    Vork runs on the ray-tracing cards: render, render-lite, render-2x and render-4x. compute (H100) has no ray tracing, so it can’t run Vork; use it for CUDA, simulation and training.

  • Network-blocked machines

    A machine leased with its network blocked (an API or desktop-app option) moves only files up to 128 KB in or out.

Where the CLI runs

  • Supported

    The voyager CLI runs on Apple silicon Macs, x64 Linux (glibc; on Alpine install gcompat first) and 64-bit Windows (Windows on Arm runs the x64 build).

  • Intel Macs and Linux on ARM

    There is no voyager build for these. curl -fsSL https://voyager.so/cloud/install.sh | sh installs voyager-cloud, an older standalone CLI that is no longer updated; use voyager-cloud <verb> where the docs say voyager cloud <verb>. Newer commands are missing from it. The machines MCP works from any computer.

  • Stale installs

    An old CLI can lack commands the docs describe. voyager update --check says whether a newer release exists and voyager update installs it. A CLI too old to have update is updated by running the install line again.

More

  • Prices

    Each preset’s all-in price per hour, and a measured example.

  • Cloud GPUs dashboard

    Running machines, live prices and spend.

  • Agent guide

    Install the CLI and render, step by step, for an AI agent.