Partial GPU Rental
Rent 1 or 2 GPUs out of an 8-GPU rig on Clore.ai and pay only for what you take
You do not have to rent the whole machine. On about 60% of listed servers you can take a single GPU out of a multi-GPU rig and pay a proportional share of the price. At the 24 August 2026 snapshot that was 1,331 of 2,202 servers and 2,153 individual GPU slots free — many of them cards you would otherwise only reach by renting a $5/hr eight-GPU box.
Almost every guide in this collection runs on one GPU. Fine-tuning a 7B model, serving Qwen3.8-27B at Q4, generating video with LTX-2.5 — one card each. Partial rental is what makes those workloads land on the good hardware instead of only on the leftover single-GPU machines.
How it works
A rig with 8× RTX 5090 lists at a daily rig price. When you rent 2 of those GPUs:
your price = rig_price × (gpu_count / total_gpus)Nothing else changes. Billing is still per minute, the order still gives you a Docker container with SSH, and the GPUs you paid for are bound to your container by their hardware UUID — so the allocation survives a reboot or PCI re-enumeration, and no co-tenant can touch them.
8× RTX 5090 @ $4.80/hr
$4.80/hr
1 GPU
$0.60/hr
8× RTX 5090 @ $4.80/hr
$4.80/hr
2 GPUs
$1.20/hr
4× RTX 4090 @ $0.80/hr
$0.80/hr
1 GPU
$0.20/hr
Rules worth knowing
On-demand only
Spot orders always take the whole rig
Same-model rigs only
A mixed rig (2× 4090 + 6× 3090) cannot be split
Hosts can opt out
Some rigs are whole-machine only, by the host's choice
GPUs are exclusive
Your UUIDs are yours for the order; other tenants get different cards
CPU, RAM and disk are not split
The rig's system resources are shared with co-tenants — check that the machine has enough RAM for everyone, not just for you
Overclock profiles
A partial order cannot set an OC profile — the API returns oc_not_supported_for_partial
Check system RAM before you take one GPU from a big rig. GPU allocation is exclusive, system memory is not. A 8× rig with 32GB of RAM is a poor host for a workload that wants 32GB by itself — look for rigs whose RAM comfortably exceeds the sum of what its tenants will use. Median RAM across the fleet is 32GB, and the top decile is 129GB.
Renting one GPU from the web UI
Open the Marketplace and pick a multi-GPU server.
On the server card, choose how many GPUs you want. Servers that support partial rental show the free-slot picker; servers that do not are whole-rig only.
The price updates proportionally as you change the count.
Configure the image and ports as usual, then rent.
Renting one GPU from the API
The listing tells you what is available. partial_gpu_rental is either false or an object:
Pass gpu_count (and optionally gpu_indices, chosen from free_indices) when creating the order:
gpu_indices is optional — pass gpu_count alone and the backend picks free GPUs for you. Omit both and you rent the whole rig.
With the Python SDK / CLI:
Errors you can get back, and what they mean:
not_enough_free_gpus
Someone took the slots between your read and your write — re-read free_indices
invalid_gpu_count
Below 1 or above total_gpus
gpu_indices_count_mismatch
The list length does not equal gpu_count
gpu_index_out_of_range
An index outside 0 … total_gpus-1
duplicate_gpu_index
The same index twice
partial_not_supported
Mixed-model rig, or the host opted out
Is it cheaper per GPU?
Sometimes, and that is not the main point. Median price per GPU per hour by rig size, from the snapshot:
RTX 4090
$0.165
$0.156
$0.172
$0.156
RTX 3090
$0.083
$0.083
$0.073
$0.083
RTX 3080
$0.079
$0.069
$0.058
$0.052
RTX 5090
$0.270
$0.271
$0.322
$0.443
Older cards get meaningfully cheaper in bulk; RTX 5090 rigs price at a premium. The real gain is availability — partial rental turns roughly a thousand multi-GPU machines into individually rentable cards, which is where most of the free 3090/4090/5090 capacity actually sits.
Next Steps
GPU Pricing & Availability — what is live right now
GPU Comparison — which card for which workload
Multi-GPU Setup — when you do want the whole rig
Links
Last updated
Was this helpful?