Live spot offers
AIVory · GPU Marketplace
Live spot pricingThe spot marketplace for AI compute.
GPUs sit idle in data centres all over the world. We scan the global spot market every few seconds and surface the single cheapest rate — so you stop overpaying. Deploy in one click, pay by the second.
The concept
Cheap GPUs, explained.
Spot pricing is renting the capacity a cloud isn't using right now — at a fraction of the sticker price.
-
What spot pricing is
Every cloud has GPUs sitting idle. They sell that slack cheap to keep it earning — often 2–5× below on-demand. The trade-off: a spot instance can be reclaimed at short notice. For batch jobs, training runs, and fault-tolerant serving, that's a price worth paying.
-
What a GPU actually does
Thousands of small cores doing matrix math in parallel. That's the engine that trains and runs AI models, generates images and video, and powers rendering and scientific simulation. More VRAM means bigger models fit on the card.
-
Where it's used & why the rate matters
LLM inference and fine-tuning, image and video generation, 3D rendering, scientific compute. In all of them, GPU hours are the bill. Shave the hourly rate and you cut the single biggest line item in any AI budget — without changing a thing about the work.
Why it's cheaper
The same card costs wildly different prices.
One data centre's idle H100 is another's premium rate. We read the whole global market and quote the lowest.
Real offers we're tracking right now across every connected provider — refreshed continuously.
Below the typical rate
How far the cheapest provider sits below the typical going rate for the same card. That gap is the money we save you.
Providers compared
Every major GPU cloud and marketplace we track, worldwide — one board instead of a dozen pricing pages.
GPU specs
Every card we track, at a glance.
Datacenter GPUs for training and large-model inference. Consumer GPUs for cost-effective inference at smaller scales.
| GPU | VRAM | Cheapest spot | Use case |
|---|---|---|---|
| RTX 4090 | 23 GB | $0.29/hr | Cost-effective inference |
| RTX 5090 | 31 GB | $0.39/hr | Cost-effective inference |
| L40S | 44 GB | $0.57/hr | Inference & rendering |
| A100 80GB | 80 GB | $1.53/hr | Training & large-model inference |
| MI300X | 192 GB | $2.35/hr | Training & large-model inference |
| H100 | 80 GB | $2.35/hr | Training & large-model inference |
| H200 | 141 GB | $4.71/hr | Training & large-model inference |
| B200 | 180 GB | $6.48/hr | Training & large-model inference |
| B300 | 288 GB | $8.19/hr | Training & large-model inference |
These are real spot prices we observed at build time. The live table below updates automatically — by the time you read this, a cheaper offer may already be available. Datacenter GPUs like the H100 and H200 have the VRAM to run 70B+ parameter models and handle large-batch training. Consumer cards like the RTX 4090 and RTX 5090 offer strong inference performance at a fraction of the hourly rate — ideal if your model fits in 24 GB of VRAM.
All cards are available as spot instances with per-second billing and no commitment. The marketplace aggregates offers from RunPod, Vast.ai, AWS Spot, Azure Spot, Crusoe Cloud, and other providers so you never need to check a dozen pricing pages yourself.
Live spot pricing
Every GPU, cheapest provider first.
Pulled live from the marketplace, biggest savings first. When cheaper capacity appears, the number moves.
Each row shows the lowest live rate for that card, how many providers have it, and the interruption risk on the cheapest offer. Honest numbers — no markups, no invented stock tickers.
Savings = how far below the most expensive provider the cheapest offer sits.
Spot capacity can be reclaimed — set a max price and we restart you elsewhere. Browse the full marketplace.
How it works
Pick a card. Ship in minutes.
No quotas to request, no contract to sign. Find the price, click deploy, get an endpoint.
-
Pick a GPU
Browse the live board, sorted cheapest first. Filter by the VRAM your model needs and read the interruption risk before you commit.
RTX 4090 · 24 GB · cheapest live -
Deploy in one click
Pre-baked images for Llama, Mixtral, Qwen and Stable Diffusion boot in around 90 seconds. Or bring your own container.
image: llama-3.3-70b · boot ~90s -
Pay per second
Billing is per-second with a max-price guard rail. Idle instances shut themselves down, so you never pay for a GPU that's doing nothing.
max-price cap · auto-shutdown on idle
Global reach
Cheapest anywhere — or exactly where you need it.
By default we hunt the lowest price across the whole market, wherever the capacity happens to sit. Need a specific region? Pin it.
The cheapest card might be in Helsinki one minute and Virginia the next — we follow it. But when latency, data residency, or compliance means location matters, pin a region or datacenter and we only deploy there. You still get the lowest price available inside it.
- North America
- Europe
- Asia-Pacific
- Cheapest, anywhere
FAQ
Common questions.
What is the cheapest GPU cloud right now?
It changes every few seconds. We scan providers like RunPod, Vast.ai, AWS, Azure, and Crusoe Cloud and surface the single cheapest live rate. An RTX 4090 can go as low as $0.29 per hour on spot.
What is a spot GPU?
Spare GPU capacity that a cloud provider sells cheaply when it is not reserved. The trade-off is the instance can be reclaimed at short notice. For batch training, inference, and fault-tolerant workloads, spot pricing saves 50 to 80 percent versus on-demand rates.
Which GPUs are available?
H100, H200, B200, B300, RTX 4090, RTX 5090, A100 80GB, L40S, MI300X, and more. We track over 70 live offers across multiple providers at any given time.
How does billing work?
Per-second billing with no minimum commitment and no reservation. Set a max-price cap and instances auto-shutdown when idle so you never pay for a GPU doing nothing.
Can I pick a specific region?
Yes. By default we show the cheapest offer worldwide. You can pin a region such as North America, Europe, or Asia-Pacific and we deploy only there while still finding the lowest price inside that region.
What if my spot instance gets reclaimed?
Set a max price and we restart you on the next cheapest available GPU. For inference workloads, you can also pair spot instances with Smart Inference for automatic API-level failover.
Do I need to manage the GPU myself?
For hands-on GPU access, yes — you get a container or a pre-baked image. If you prefer a fully managed API where you never touch a server, use Smart Inference instead. It routes requests and manages infrastructure behind the scenes.
Don't want to manage GPUs at all?
Pricing
Pay per second. Nothing else.
No subscription, no reservation, no minimum. Spin a GPU up, spin it down, pay for the seconds in between.
The same card can cost very different rates depending on who's selling spare capacity. We always quote the cheapest live provider.
— · typical rate
Median across live providers
—
— · via AIVory
Cheapest live spot price
—
Same card, same hour. We pick the cheapest provider live, every time.
- Per-second billing
- No commitment
- Set your own max price
- Auto-shutdown on idle
Cheap GPUs. Live prices. Deploy now.
Find the cheapest card, click deploy, pay by the second.
The marketplace is live and the board updates every 30 seconds. Grab the rate you see — spot capacity moves fast.