Cheap machines that finish the job.

Spot capacity is cheaper because it can vanish. slipa turns that into a bounded interruption: save the work, release the machine, resume from a verified checkpoint.

Illustrative interactive continuity trace.

The machine can leave. The job should not.

The right comparison is not spot versus on-demand by hourly rate. It is the cost of finishing a job after the machine has been reclaimed.

One eviction, two ways of handling it

Same job, same interruption, same replacement card.

Spot, by handno checkpoints Queued
Spot with slipacheckpoint, move, resume Queued

A checkpoint manifest. The point a run can return to.

Work discarded. Everything the eviction threw away.

Work repeated. That same work, run a second time. With no checkpoint to return to, it is the whole run.

No GPU held. The wait for replacement capacity. Nothing is billed here.

Both jobs queued for the same card.

One durable run. Machines are replaceable.

The lifecycle is shared by every supported workload. A provision failure, eviction, or missing heartbeat returns to capacity selection; a completed run delivers its result and releases the machine.

Run lifecycle

Seven phases, shared by every workload type.

StepPhase
01Validate job
02Wait for price
03Find capacity
04Provision
05Monitor run
06Deliver
07Release

Return paths

Eviction and rebid return through the same control point.

TriggerWhat happens
Eviction or heartbeat lost Return to find capacity and resume from the newest manifest that verifies.
Cheaper option wins Checkpoint, release the current card, and re-provision only when remaining cost clears the rebid threshold.

The recovery loops are the product: market volatility becomes a managed state transition rather than an operator incident.

See the hourly market. Score the completed run.

A quote is only useful if the card can meet the job’s VRAM floor, is actually available, and is likely to finish the work inside the cap. The board shows the first input; the scheduler uses the full run estimate.

Live hourly snapshot

The same GPU across multiple providers.

Reading

ProviderCardPer hour
—————.——
—————.——
—————.——

The lowest hourly quote is not automatically the choice for the whole job.

Notice the loss. Trust the last complete save.

The worker reports a heartbeat every ten seconds. For supported provider states, slipa also checks provider liveness every thirty seconds rather than waiting through the full heartbeat timeout.

Loss detection

Same interruption. Different response time.

PathTimelineTime to notice
Heartbeat timeoutfallback path Up to 300 seconds. Silence may mean a slow step, so this timeout must be conservative.
Provider livenesssupported terminal states About 30 seconds. An explicit terminal state starts recovery without treating unknown states as evictions.

Unknown states, missing data, and provider errors remain alive and fall back to heartbeat health. A false eviction is worse than a delayed recovery.

Checkpoint integrity

Manifest last. Newest verified checkpoint wins.

UploadWhat it means
Complete upload adaptertokenizerstatemanifest ✓ The manifest lists every expected file and size. It is written last, so its presence is evidence that the upload reached a complete state.
Interrupted upload adaptertokenizerstatemanifest No manifest means the directory is skipped. Half-written files never become a resume target.

Manifest walk

Newest first, until one reconciles.

CheckpointFindingVerdict
step-1400/ Manifest missing after an interrupted write. skip
step-1200/ Manifest exists, but a size does not match. skip
step-1000/ Every file reconciles with the manifest. resume here

Writing the manifest last keeps resume work bounded. A torn checkpoint is skipped rather than trusted.

We are currently in private beta.

Available in the product

One control flow across workloads.

Shipped
  • Finetune, batch inference, RL rollouts, RL finetune, and hello-GPU workload types.
  • Per-job budget caps, heartbeats, cost events, and resumable storage paths.
  • Routing across multiple GPU providers and public bring-your-own images.

Current limitations

Not every job belongs on spot.

Not offered today
  • Interactive or latency-sensitive work should use continuous capacity.
  • Private registry credentials and multi-node training are not offered today.
  • Savings bands are modelled; customer-run performance data is not claimed here.