The offers are compared
Smart Inference reads live offers across providers such as AWS, Google Cloud, Azure, Lambda, Crusoe, Hyperstack, RunPod, Vast.ai and others, and prices every offer per GPU, not per machine.
AIVory Smart Inference · GPU rentals
You need a lot of GPUs. See who has them this week.
High-end GPUs are never in one place. Smart Inference reads live offers from a dozen clouds and marketplaces and rents from the cheapest one that has what you need. Pick the GPU, the count and the duration below, and see every offer that is live right now, priced for your whole job.
B300 capacity is thin and moves within days. The quote shows what is live at this moment, and stock is confirmed when you rent.
Live offers from our own API, on this page.
GPU quote
Every live offer, priced for your job.
Pick the GPU, how many and for how long. The list loads live from the offers Smart Inference can rent from right now, priced per GPU and for the whole job.
Loading live offers…
Cheapest per GPU hour–
Cheapest for the job–
Matching offers–
| Offer | Capacity | Interruption risk | Memory per GPU | Region | $ / GPU hour | Job total |
|---|
Each rental is one machine; a large count may be split over several machines, and stock is confirmed when you rent. Billing is per GPU, per second, after a 15-minute minimum. Offer names are routing labels, not cloud brands.
How a rental works
One API, many clouds, billed per GPU.
You do not open an account with each cloud. Smart Inference rents the machine, hands you SSH access, and bills you per GPU.
You rent one machine
A rental is one machine with the GPU count you ask for, up to the full node a provider sells, such as an 8-GPU H200 node. You get SSH with your own keys and up to ten open ports.
You pay for what runs
Billing is per second after a 15-minute minimum, per GPU. A rental that sits idle for 15 minutes stops, so a forgotten machine does not run all weekend.
What you get back
A quote you can check against the market.
Every line of the quote is something the rental API reads directly, so the quote and the rental match.
Price per GPU per hour
Per GPU, spot and on-demand side by side, from each provider that has stock. Not a per-machine price you have to divide yourself.
Interruption risk
Spot offers carry an interruption estimate. Training with checkpoints can take the risk. A demo next Tuesday probably should not.
Region
Where each offer runs, when the provider says so. Many offers do not name a region until the machine starts.
Machine shape
GPUs per machine, GPU memory, and whether the GPUs in one machine are linked. This decides whether a large model fits on one node.
Access
SSH with up to ten of your public keys, and the ports you need open. Rent and release through the same API we quote from.
A credit hold, not a contract
Rentals run on prepaid credits. We hold credit when the machine starts and bill the actual seconds. There is no reservation fee and no term.
Set your expectations
We are not a cluster builder.
Here is the deal in full, both halves of it.
What you do get
What you do not get
What happens next
From the quote to a running machine.
-
You pick an offer
Above. Spot where your job survives an interruption, on-demand where it does not.
-
You sign up and add credit
Prepaid credits. Credit is held when the machine starts, and the actual seconds are billed.
-
You rent through the API
POST /v1/gpu/rentwith the GPU, the count and your SSH keys. The GPU docs have the call. -
You release it
Delete the rental when the job is done, or let the 15-minute idle stop catch it.
Questions
Before you rent
What does a quote cost?
Nothing, and you need no account. The quote runs in your browser against our public offers API.
Can I rent B300s?
Sometimes. B300 offers appear in the marketplaces we read, but stock is thin and changes within days. The quote shows what is live right now, and stock is confirmed when you rent. The B300 price page tracks the spot price.
Can I rent 32 GPUs at once?
As separate machines, yes, if the stock is there. Each rental is one machine, up to the full node a provider sells, such as 8 GPUs. We do not build multi-node clusters or the networking between machines.
How is a rental billed?
Per GPU, per second, after a 15-minute minimum. Credit is held when the machine starts and the actual time is billed. A rental that sits idle for 15 minutes stops on its own.
Spot or on-demand?
Your choice per rental. Spot is cheaper and can be reclaimed by the cloud. Every spot offer carries an interruption estimate, so you can set the risk you accept.
Can I see the prices myself?
Yes. The GPU price pages show live spot prices per GPU, and the rental API lists the offers through GET /v1/gpu/offers. The quote on this page reads the same list.
Find your GPUs.
Every live offer for the card you need, priced for the whole job, on one page.
See live offers