SPOT BIDS
Bid low to run cheap. Bid high to run now. Volatility works in your favor.
TWO WAYS TO USE SPOT
Match the bid to the job
Same GPUs, same auction, same market price. The only thing you choose is your limit.
01
Run workloads as cheaply as possible
Training runs, batch inference, sweeps, and work that’s only worth doing when compute is cheap. Set a low bid and let it sit. When the market dips to your bid, the job runs.
12-hr A100 job · low bid · overnight run for under $10 total
expect: some waiting · occasional preemption
02
Run workloads right now, at market rate
Interactive dev sessions, deadline runs, burst capacity for inference spikes. Set a high bid as a safety limit. You start immediately, almost never get preempted, and still pay the market rate, not your bid.
1×H100 dev session · high bid · ~$7 for 3 hrs at market rate
expect: instant start · never pay above your limit
HOW IT WORKS
Wait less or pay less
01
Set your limit
Bid the most you’ll pay per instance-hour. Your bid enters a continuous auction for the capacity.
02
Run when it clears
The market moves. Whenever it clears at or below your limit, you’re running; when it rises above, your bid waits.
03
Pay the market
It’s a blind second-price auction. While you’re running, you pay the clearing price, never more than your bid.
Illustrative · simulated bids against a sample 48-hour window
PREEMPTION
Preempted? You get five minutes, not zero
A spot instance can be preempted when a higher bid claims its capacity. Here’s what makes that cycle a non-event.
5-minute warning
The notice lands in the instance filesystem the moment you’re outbid, and you get five minutes before shutdown. Monitor it to cordon and drain automatically.
Checkpoint and resume
Persistent volumes and your boot disk survive preemption. Save state on the way down, and pick up where you left off when you reallocate.
Automatic reallocation
No re-bidding and no babysitting. Your bid stays in the auction, and when the market comes back to your price, you’re allocated again.
LIVE PRICES
See the market before you bid
The market price and the win price are both public. Watch the tape in the console or pull it from the CLI, then bid with the whole picture.
MULTI-REGION
Different regions, different markets
Each Mithril region has its own GPUs and its own market dynamics. If a workload can run in more than one place, submit bids in several regions and cancel the rest once one wins.
QUICKSTART
Submit a bid from the CLI
Every command is an API call underneath. Place, adjust, and cancel bids programmatically.
# live spot and win prices
$ ml instance list-types
# bid on an A100 on spot, limited to $0.60/GPU-hr
$ ml instance create --instance-type a100 --region us-central3-a --max-price-per-hour 0.60
FAQ
Spot, answered
What happens if I get preempted?
You get a five-minute warning, written to the instance filesystem, so you can checkpoint and drain automatically. Persistent volumes and your boot disk survive. When the market returns to your price, your bid reallocates.
Can I pause my bid without losing my environment?
Yes. Deactivate the bid, and your boot disk and persistent storage stay intact; only local scratch space is lost. Unpause the bid whenever you want to re-enter the auction.
How do I pick a bid price?
Bid what the time is worth to you. You pay the market price, never more than your bid. A high bid doesn’t raise your cost when the market price is lower; it just keeps you allocated when the market rises.
What’s the difference between spot and a reservation?
Spot pays the live market rate and can be preempted. A flexible reservation guarantees capacity at a fixed rate (as short as a few hours) and credits back the hours you pause or return.
Can I run multi-node distributed training on spot?
Yes. Submit one bid for multiple nodes and it allocates when capacity clears at your price, on the same InfiniBand fabric as reserved capacity.
Can I serve inference on spot?
Yes. Most requests finish well inside the five-minute preemption warning, so a preempted node drains cleanly: stop sending it new requests and let the in-flight ones complete. Bid high to hold burst capacity when traffic spikes.