GPU Rent HubGPU for fine-tuning

Rent a GPU for fine-tuning, sized by arithmetic you can check.

GPU Rent Hub rents dedicated NVIDIA and AMD GPUs for fine-tuning by the week or month, from $190 a month for a 24 GB card that runs QLoRA on a 7B model to $12,170 a month for an NVLink-connected 8× H100 SXM node, paid in Bitcoin, Ethereum, USDT, USDC, Monero, Solana or Litecoin with no identity verification.

Updated 2026-09-01

112 GB7B full fine-tune, Adam, before activations

10 GBSame model under QLoRA

900 GB/sNVLink on H100 SXM and H200

72 hGrace period if a balance runs dry

01 Definition

How much GPU memory does fine-tuning actually need?

This is the part people get wrong, and it is the only part that decides which card you rent.

Fine-tuning memory in one paragraph. A full fine-tune with the Adam optimiser costs about 16 bytes per parameter: 2 for the BF16 weights, 2 for the gradients, 4 for the FP32 master copy and 4 each for Adam's two moment states. A 7B model therefore needs roughly 112 GB before a single activation is stored, which does not fit a 24 GB card and does not fit an 80 GB card either. LoRA freezes the base weights and trains small adapter matrices, so the bill falls to the frozen base at 2 bytes per parameter plus a few hundred megabytes of adapter and optimiser state: about 21 GB for a 7B. QLoRA quantises that frozen base to 4 bits, 0.5 bytes per parameter, and trains the same adapters: about 10 GB for a 7B and about 51 GB for a 70B.

Those three figures, and every one in the table below, are the GPU VRAM guide's, on its assumptions: gradient checkpointing on, sequence length 2,048, batch 1, rank-16 adapters at about 0.5% of parameters, and a runtime overhead allowance already included. Raise the sequence length or the batch and activations grow on top.

Where the 16 comes from

Nothing in a training loop stores the model once. Weights, a gradient of the same shape, an FP32 master copy, and Adam's mean and variance per parameter. No framework negotiates that away.

What LoRA removes

The gradient and both optimiser states, for everything but the adapters. The frozen base still has to be resident, which is why LoRA on a BF16 7B is still 14 GB before anything else. QLoRA attacks residency, not the optimiser.

02 Sizing

Which GPU do I need to fine-tune a 7B, 13B or 70B model?

Model against method against a card that holds it, with the VRAM quoted from the VRAM guide. Per GPU-hour is monthly divided by 720.

Model and method	VRAM needed	Card that fits	Card VRAM	Weekly	Monthly	Per GPU-hour
7B, QLoRA4-bit base, rank 16	10 GB	RTX A5000	24 GB	$55	$190/mo	$0.26
13B, QLoRA4-bit base	14 GB	RTX 3090	24 GB	$65	$230/mo	$0.32
7B, LoRABF16 base held resident	21 GB	RTX 4090	24 GB	$110	$390/mo	$0.54
34B, QLoRA4-bit base	28 GB	A40	48 GB	$68	$240/mo	$0.33
13B, LoRABF16 base	33 GB	L40S	48 GB	$148	$520/mo	$0.72
70B, QLoRA35 GB of 4-bit weights	51 GB	A100 SXM	80 GB	$280	$990/mo	$1.38
34B, LoRABF16 base, one card	79 GB	MI300X	192 GB	$450	$1,590/mo	$2.21
7B, full fine-tuneAdam, 16 bytes per parameter	129 GB	2× A100 SXMNVLink 600 GB/s, the cheapest CUDA path	160 GB	$532	$1,881/mo	$1.31
70B, LoRABF16 base, one card	156 GB	1× MI300Xno sharding needed	192 GB	$450	$1,590/mo	$2.21
70B, full fine-tuneZeRO-3 across the node	1,288 GB	8× MI300XInfinity Fabric 896 GB/s	1,536 GB	$3,240	$11,450/mo	$1.99

VRAM needed is the figure the GPU VRAM guide derives for that model and method, at sequence length 2,048 and batch 1, with gradient checkpointing on and the runtime overhead allowance already inside it. The two-card price includes the 5% multi-GPU discount; the eight-card figure is the node price. Per GPU-hour is the monthly total divided by 720 hours and by the card count. The two MI300X rows are DFW-1 Dallas only and run ROCm rather than CUDA. These are the cards we would run the job on; the VRAM guide's lookup names the cheapest card that merely holds it, so a row here is sometimes the more expensive card. Two rows diverge and both have a reason. The 7B LoRA row takes an RTX 4090 over the guide's RTX A5000 at the same 24 GB because 1,008 GB/s of bandwidth against 768 shortens every step. The 7B full fine-tune row takes two A100 SXM as the cheapest CUDA and NVLink path, where one 192 GB MI300X holds the same 129 GB for $1,590 a month on ROCm. Specs on the GPU catalogue.

Read row eight against row six. A full fine-tune of a 7B costs about twice what QLoRA on a 70B costs. For most people it is not the job they need doing.

03 The pick

Which GPU should you rent for training, if you had to pick one?

Sorted by what we would deploy for a multi-week run, not by price. Stock is across DFW-1 Dallas, IAD-1 Ashburn and PDX-1 Hillsboro.

GPU	VRAM	Interconnect	Weekly	Monthly	Per GPU-hour	Free now
H100 SXM PickHBM3, 3.35 TB/s, FP8	80 GB	NVLink 900 GB/s	$475	$1,690/mo	$2.35	63
A100 SXMHBM2e, 2.04 TB/s, no FP8	80 GB	NVLink 600 GB/s	$280	$990/mo	$1.38	78
H200 SXMHBM3e, 4.8 TB/s	141 GB	NVLink 900 GB/s	$640	$2,290/mo	$3.18	19
MI300XROCm 6.4, 5.3 TB/s	192 GB	Infinity Fabric 896 GB/s	$450	$1,590/mo	$2.21	8
L40SGDDR6 ECC, FP8	48 GB	PCIe Gen4 peer-to-peer	$148	$520/mo	$0.72	74
RTX A6000GDDR6 ECC, 768 GB/s	48 GB	NVLink bridge 112 GB/s, pairs	$82	$290/mo	$0.40	65
RTX 3090GDDR6X, 936 GB/s	24 GB	NVLink bridge 112 GB/s, pairs	$65	$230/mo	$0.32	186
RTX A5000GDDR6 ECC, 230 W	24 GB	NVLink bridge 112 GB/s, pairs	$55	$190/mo	$0.26	51

Free now is unreserved cards across the three sites at the last catalogue update. Every price includes power, cooling, a 100G uplink with a 10G guarantee, IPv4 and IPv6, and 20 TB of outbound a month. Multi-card and prepaid discounts are on the pricing page.

The H100 SXM is the pick: the cheapest card we rent that puts 80 GB of HBM3, FP8 tensor cores and 900 GB/s of NVLink on one part number. The A100 SXM 80 GB is the value answer at $990, no FP8, same NVLink topology, 59% of the price.

04 Interconnect

Why does NVLink matter above one card?

Because a sharded run moves gradients between cards on every step, and the link is the floor on how fast that happens.

On one GPU the interconnect is irrelevant. The moment you shard, with ZeRO-2, ZeRO-3 or tensor parallelism, the cards exchange gradients or parameter shards every step, and that traffic is a fixed multiple of the model size. PCIe Gen4 x16 has a line rate of about 32 GB/s in one direction, and Gen5 about 63. NVLink on an H100 SXM is 900 GB/s. That is the difference between a communication step that hides behind compute and one that becomes the step.

The rule: single-card LoRA or QLoRA runs on anything with enough VRAM, data-parallel training tolerates PCIe, and anything that shards a model wants NVLink. Above four cards it wants NVSwitch, which the 8-way HGX nodes on the multi-GPU cluster rental page carry.

The RTX 4090 and RTX 5090 have no NVLink. NVIDIA dropped it from the consumer line after Ampere. They are strong single-card training GPUs and a real constraint above one card, and we would rather say so here than after you have bought eight.

The A6000, A40, RTX 3090 and RTX A5000 take an NVLink bridge, but only in pairs and only at 112 GB/s. Useful for pooling two 48 GB cards into 96 GB. Not a fabric.

Card	Link between cards	Bandwidth	Sharded training
B2002, 4 or 8 way	NVLink 5 with NVSwitch	1.8 TB/s	Yes, to 8 cards
H100 SXM, H200 SXM	NVLink with NVSwitch	900 GB/s	Yes, to 8 cards
A100 SXM	NVLink with NVSwitch	600 GB/s	Yes, to 8 cards
MI300X	Infinity Fabric	896 GB/s	Yes, 8-way only
H100 PCIe	Optional NVLink bridge	600 GB/s	Pairs
A6000, A40, 3090, A5000	NVLink bridge	112 GB/s	Pairs
L40S, RTX 6000 Ada, L4	PCIe Gen4 peer-to-peer	~32 GB/s	Data parallel only
RTX 4090, RTX 5090	PCIe peer-to-peer, no NVLink	~32 / ~63 GB/sGen4 x16 / Gen5 x16	Single card

Aggregate per-GPU link bandwidth as published by the vendor. The PCIe figures are line rates for an x16 link in one direction, Gen4 on the 4090 and Gen5 on the 5090, not measured throughput, which lands lower.

05 Term

Why is a monthly term the right shape for a training run?

A fine-tune is the workload where the billing model and the job agree with each other.

Checkpoint cadence is yours

On preemptible capacity you checkpoint defensively, because the interval is a hedge against eviction. On a term it is a recovery decision: write every thousand steps if the loss curve deserves it, or every ten thousand.

2 TB local NVMe on an H100 SXM

A fine-tune is rarely one run

A learning-rate sweep, a data ablation and a rerun after the evaluation disappoints are three more runs, arriving over a fortnight rather than in one booking. Each costs nothing extra on a flat term.

Same card, same environment, no reprovisioning

It cannot be evicted

An eviction at hour 60 costs the gap since the last checkpoint, the requeue wait and the dataset reload. A term cannot be preempted, because nobody else is queued for the card.

Single-tenant bare metal, root access

The dataset stays resident

Between runs the working set sits on local NVMe, 2 TB on an H100 SXM and 1 TB on an A100, so sweep two starts at the first step rather than at a download.

No cold start after the first one

The crossover arithmetic against a meter belongs to monthly versus hourly GPU rental, and it is worth reading before you take a term. What is specific to training is that the sweeps and the reloads count as GPU-hours too, so the utilisation figure you should test the crossover with is the fortnight, not the single run.

06 Checkpoints

Where do checkpoints go, and what survives the end of a term?

A 70B checkpoint with optimiser state is over a terabyte. Decide where it lands before step 4,000.

Local NVMe: Included: 2 TB on H100 SXM and H200, 1 TB on A100 SXM, 500 GB on L40S and A6000. LUKS2-encrypted per instance, wiped seven days after term end.
Extra local NVMe: $9 per TB a month, up to 30 TB on a node. Same host, same encryption.
Network volume: $14 per TB a month, triple-replicated in the region, attachable to several instances, persistent after release.
Snapshots: $6 per TB a month. Restorable to any instance in the region, which is how a run moves from an A100 to an H100 without re-uploading the dataset.
Pulling it out: 20 TB of outbound included per instance per month, then $4 per TB. An adapter set never touches that limit.

If the run outlives the balance

Terms renew from your balance 24 hours early. If it is short the instance stops and the slot is held for 72 hours while you top up in BTC, ETH, USDT, USDC, XMR, SOL or LTC. The disk is untouched. After that the slot is released and the NVMe key destroyed.

Write every checkpoint twice: local NVMe for speed, a network volume for survival. The volume is the only copy that outlives the term.

BTCETHUSDTUSDCXMRSOLLTC

Hardware sits in DFW-1, IAD-1 and PDX-1. An account is a username and a password: see no KYC GPU hosting and how to pay with crypto.

07 Honesty

When should you not rent a fine-tuning GPU from us?

Three cases, common enough that we would rather write them down.

If the job is one experiment that runs for six hours, use an hourly provider. Our shortest term is seven days. A month of H100 SXM for one afternoon of QLoRA costs $1,690, and you use about $18 of the $1,690 you paid.

The second is over-buying the card. Plenty of people rent an H100 to fine-tune a 7B with QLoRA, which needs 10 GB. That job runs on a $190 RTX A5000, slower, but not nine times slower. Size from the table in section 02, not from the model you wish you were training.

The third is a run needing 32 or 64 cards for two days and never again. We sell nodes up to eight cards; beyond that a metered cluster provider fits better.

Rent by the hour if

Your total is under roughly 480 to 680 GPU-hours a month, which is where $1,690 for an H100 SXM month meets an illustrative $2.49 to $3.49 meter, or you want an hour each on an A100 SXM and an H100 SXM to time your step.

Rent a term if

The run is measured in weeks, re-uploading the dataset hurts, you have been evicted from spot capacity once already, or you want to pay in crypto without an identity check.

08 Questions

Renting a GPU for fine-tuning, asked plainly.

How much VRAM do I need to fine-tune a 7B model?

About 10 GB under QLoRA, 21 GB under LoRA on a BF16 base, and 129 GB for a full fine-tune with Adam once activations are counted, or 112 GB for weights, gradients and optimiser state alone. The first two fit a 24 GB card, so an RTX A5000 at $190 a month. For the third, two 80 GB cards with NVLink at $1,881 a month is the cheapest CUDA path, and one 192 GB MI300X at $1,590 holds the same 129 GB for less if a ROCm stack suits you. Figures from the GPU VRAM guide.

Can I fine-tune a 70B model on one GPU?

Under QLoRA, yes. A 4-bit 70B base is about 35 GB of weights and 51 GB with adapters, optimiser state and overhead, which fits an A100 SXM 80 GB at $990 a month. LoRA on a BF16 70B base needs 156 GB, which is still one card if it is a 192 GB MI300X at $1,590. A full 70B fine-tune needs about 1,288 GB and takes a node.

What is the difference between LoRA and QLoRA in GPU memory?

LoRA freezes the base weights in BF16, 2 bytes per parameter, and trains small adapter matrices, so optimiser state is hundreds of megabytes rather than tens of gigabytes. QLoRA quantises that frozen base to 4 bits, 0.5 bytes per parameter. On a 7B model that is 21 GB against 10 GB.

Do I need NVLink to fine-tune on multiple GPUs?

Not for one card, and not for data-parallel runs on small models. It matters when gradients or parameter shards cross cards every step. H100 SXM and H200 carry 900 GB/s, A100 SXM 600 GB/s, B200 1.8 TB/s. The RTX 4090 and RTX 5090 have no NVLink, so we do not recommend them for sharded training.

How much does it cost to rent a GPU for training per month?

An A100 SXM 80 GB is $990 a month, which is $1.38 per GPU-hour across a 720-hour term. An H100 SXM is $1,690, or $2.35. An eight-card H100 SXM node with NVSwitch is $12,170, or $2.11. A 24 GB card for QLoRA starts at $190, and every card is available weekly.

What happens to my checkpoints when the term ends?

Local NVMe is wiped seven days after term end and its encryption key destroyed, so anything you want to keep belongs on a network volume at $14 per TB a month, triple-replicated in the region and persistent after the instance is released. You can also pull it out over the 20 TB of outbound included each month.

Can I rent a training GPU with no identity verification?

Yes. An account is a username and a password. No email address, no phone number, no identity document, no card on file. You fund a balance in BTC, ETH, USDT, USDC, XMR, SOL or LTC and the term is drawn from it. No agent runs inside your instance and nobody here reads your dataset.

Is monthly GPU rental better than hourly for a training run?

For a run measured in weeks, yes. A fine-tune keeps the card busy nearly all the time, which is where a flat term wins: $2.35 per GPU-hour on an H100 SXM against $2.49 to $3.49 an hour, an illustrative range across US on-demand providers rather than a quote from a named vendor, with no eviction and no queue. For one afternoon, use an hourly provider.

Sizing on the GPU VRAM guide, topologies on multi-GPU cluster rental, serving the result on GPU for LLM inference, images in the documentation.

Size the card, take the term, start the run.

QLoRA on 24 GB from $190 a month, an NVLink H100 SXM from $1,690, an 8-way node from $12,170. Minimum deposit $20 equivalent, no ID, no card.

Create an account →