Deploy AI models on H100, A100, H200, and AMD MI300X GPUs with up to 40% cost savings. Lightning-fast machine learning inference on enterprise GPU infrastructure.
Fine-tunes, training runs, batch inference on marketplace GPUs at marketplace prices. Aquanode checkpoints your job on a schedule. When a machine is lost, it rents another, restores the last checkpoint, and re-runs your command. You lose minutes, not the run.
A job that loses its machine resumes on any of them.
Checkpoints & failover
A machine dies mid-run. The run doesn't.
Submit a command, name your checkpoint paths, walk away. Detection, the replacement machine, the restore and the re-run are ours to handle, not yours to babysit.
01 / CHECKPOINT
Your paths, on your interval
Point the job at the directories your code already writes checkpoints into, and pick how often we snapshot them. Only the first snapshot carries everything and later ones carry what changed, so a short interval stays cheap.
02 / DETECT
Gone means the provider says gone
A quiet machine is not treated as a dead one. We confirm with the provider that the box is actually gone before acting, so two machines never run the same job and bill you twice.
03 / RESUME
Another provider, same command
A new machine is rented automatically, skipping the provider that just lost one. The last checkpoint is restored and your command re-runs. Every handoff shows on the run page: what was lost, what was resumed from.
One setupThree datacenters
us-east
H100 80GB
hyperstack
eu-west
A100 80GB
massed compute
us-west
L40S 48GB
vast.ai
training-envv4
restored · models, dataset and packages, ready to train
restored · same environment, different provider
restored · same environment, a cheaper card
If a provider reclaims a running box, the job resumes from your last checkpoint, not from the instant it died. A short interval keeps that window small. With no checkpoint paths, a retry starts from the beginning.
Why the price works
Cheap GPUs are cheap because they can be taken back.
Marketplace supply runs a fraction of the datacenter tiers for one reason: nobody promises the box stays up. Keep the price, hand us the reason it's low.
THE TRADE
A datacenter tier sells you a promise
The same card costs two to three times more with an uptime promise attached, because someone has to absorb the risk of the machine going away. Pay it, and the premium is the product.
OUR SIDE
We absorb the interruption instead
Jobs are built assuming the machine can vanish. Checkpoints on your interval, a confirmed-dead check, a replacement box on another provider. The failure still happens; it just stops being your problem.
Every card and every provider we place on, priced live: see today's prices. Jobs that shouldn't leave a datacenter don't have to: placement can be restricted to datacenter hosts only.
Pods
Need the machine itself? That's a Pod.
Jobs run and finish. A Pod you sit inside: a dev box whose whole environment, packages and CUDA included, pauses, moves between providers, and comes back. Three things a rented machine can't do, below.
Migrate, pause, or resume across any GPU datacenter.
Pause a Setup in one datacenter and resume it in another, on a different card from a different provider. Your files, models and packages come back the way you left them.
aq pause training-env → aq deploy --snapshot 4242 --gpu L40S
PROVIDER A
H100 · paused
PROVIDER B
L40S · running
✓ /root/models · 41.2 GB restored
✓ /root/dataset · 8.7 GB restored
↻ environment · ready to run
See how it works
See a pod move onto another datacenter.
The same console you get on day one, running the three actions a rented box cannot do for you.
Portable
Release
Sync
console.aquanode.io/pods/training-envConnected
training-envpod
RUNNING
GPUH100 (80GB)
vCPU28 cores
Memory180 GB
Storage6 TB NVMe
GPU util88%
VRAM72%
massed compute · eu-central-1
vast.ai
hotaisle
vultr
runpod
41.2 GB portable set in flight
FROM
H100 80GB
hyperstack · canada-1
released
TO
L40S 48GB
massed compute · eu-central-1
running
5.0sCome back on a different provider with the same environment.
Auto-pause
The box you forgot about pauses itself.
Under 5% GPU for 30 minutes and you get a warning. At 60 the Setup saves itself and the machine is released. Your compute bill stops there, and everything comes back when you resume.
Storage for the saved Setup keeps billing; only the compute stops. Turn auto-pause on per Setup.
GPU 84% · ACTIVEGPU 3% · IDLE 30 MIN · WARNINGGPU 3% · IDLE 61 MIN · PAUSED
compute → $2.40/hrcompute → $0.00/hr
Bring your own hardware
Already have the machines? Use them.
If you hold a multi-year lease, the last thing you need is somewhere else to rent. Pods work on hardware we never provisioned.
01 / IMPORT
Turn a box you own into a Pod
aq import surveys the machine, captures it, and registers it as a Pod: versioned, forkable, and launchable on any provider we support.
aq import --dry-run
02 / RUN IN PLACE
Pods on your own machines
Point Aquanode at a box you already pay for. Capture, restore, versions, run and logs, against your own storage bucket, with no account required. Or connect it to the console and get sharing, metrics and jobs on hardware we never rented.
aq host add lab-01 --ssh root@10.0.4.7
Release a machine and it keeps running. We revoke our credentials and drop the record. We never touch your box.
GPU observability
Every reading straight off the card, while the job is still running.
Utilization, VRAM, temperature and power on a running deployment, refreshed every few seconds, and a reading the card cannot give you says so, instead of showing a zero.
console.aquanode.io/vms/llm-finetune-01 · Metrics
Active
This GPU
Utilization84%
Memory used64.2 / 80 GB
Temperature73 °C
Power draw433 / 700 W
SM clock1,845 MHz
Memory clock2,619 MHz
The host it sits on
CPU utilization32%
System memory104 / 180 GB
Utilization · sessionlive window · ~10 min
This card reports every field: utilization, memory, clocks, temperature and power.
