Pricing

Same silicon, less than half the bill

NexGPU connects directly to 1,175 compute nodes across 51 countries. An H100 costs $3.582 per GPU-hour here — 47% below the AWS on-demand rate. Billed hourly, no minimum spend, no contract.

Entry tier per GPU-hour: $0.083
H100 cheaper than AWS on-demand: 47%
GPU models in stock: 68
Live rentable nodes: 1,175

01 — Rate card

Every GPU we carry

Datacenter and consumer classes listed separately. Rates are per GPU-hour — the lowest rentable price for that model.

Datacenter class

A100, H100, L40S and other professional accelerators — for large-scale training, high-concurrency inference and long stable runs.

Datacenter class
Model	VRAM	Nodes	From / GPU-hr
Quadro P4000up to 2x	8GB	6	$0.110
RTX A2000up to 2x	6GB	3	$0.137
RTX A4000up to 8x	16GB	31	$0.145
Tesla V100up to 4x	32GB	21	$0.188
Tesla P40up to 2x	24GB	2	$0.214
Q RTX 6000up to 4x	23GB	5	$0.274
RTX 4000Adaup to 2x	20GB	4	$0.274
Tesla T4up to 2x	15GB	3	$0.298
RTX A5000up to 2x	24GB	5	$0.355
RTX PRO 4000up to 4x	24GB	28	$0.408
A10up to 2x	22GB	4	$0.414
Q RTX 8000up to 2x	48GB	4	$0.508
RTX PRO 4500up to 2x	32GB	5	$0.625
L4up to 2x	22GB	6	$0.629
RTX 5000Adaup to 2x	32GB	3	$0.682
RTX A6000up to 4x	48GB	8	$0.817
A100 PCIEup to 4x	40GB	22	$0.824
RTX 5880Adaup to 4x	48GB	8	$0.829
A100 SXM4up to 8x	80GB	22	$1.088
RTX 6000Ada	48GB	8	$1.109
RTX PRO 5000up to 2x	48GB	17	$1.125
L40Sup to 4x	45GB	8	$1.639
RTX PRO 6000 Sup to 8x	96GB	19	$1.852
RTX PRO 6000 WSup to 4x	96GB	32	$2.055
H100 SXMup to 8x	80GB	16	$3.582
H100 PCIEup to 8x	80GB	8	$4.182
H100 NVL	94GB	3	$5.201
H200up to 8x	140GB	17	$6.660
H200 NVLup to 2x	140GB	7	$7.342
B200up to 8x	179GB	16	$10.797
B300up to 8x	269GB	3	$15.742

31 models

Consumer class

RTX 5090, 4090 and 3090 gaming cards — the cheapest compute per dollar, ideal for image generation, LoRA fine-tuning and small-to-mid model inference.

Consumer class
Model	VRAM	Nodes	From / GPU-hr
GTX 1660 Sup to 8x	6GB	11	$0.083
GTX 1080up to 6x	8GB	10	$0.094
GTX 1070up to 2x	8GB	3	$0.095
RTX 3060 Tiup to 2x	8GB	12	$0.097
GTX 1070 Tiup to 4x	8GB	2	$0.099
RTX 3060up to 6x	12GB	48	$0.100
GTX 1080 Tiup to 7x	11GB	13	$0.108
RTX 3070up to 8x	8GB	25	$0.109
RTX 4060up to 4x	8GB	8	$0.112
RTX 2060Sup to 2x	8GB	5	$0.117
RTX 2060up to 2x	12GB	3	$0.120
Titan Xpup to 7x	12GB	10	$0.121
RTX 5060up to 2x	8GB	5	$0.123
RTX 4060 Tiup to 4x	16GB	27	$0.126
RTX 2080 Tiup to 8x	11GB	9	$0.135
RTX 5060 Tiup to 8x	16GB	70	$0.136
RTX 3060 laptopup to 2x	12GB	12	$0.137
GTX 1660 Ti	6GB	4	$0.138
RTX 4070Sup to 8x	12GB	20	$0.139
RTX 4070up to 4x	12GB	19	$0.149
RTX 3080up to 2x	10GB	19	$0.161
RTX 3080 Tiup to 4x	12GB	24	$0.163
RTX 4070 Tiup to 4x	12GB	12	$0.174
GTX 1060up to 2x	3GB	2	$0.176
Titan Vup to 2x	12GB	4	$0.178
RTX 3070 Tiup to 2x	8GB	5	$0.187
RTX 5070up to 2x	12GB	18	$0.191
RTX 3090up to 8x	24GB	66	$0.193
RTX 5070 Tiup to 4x	16GB	29	$0.229
RTX 4080up to 4x	16GB	13	$0.242
RTX 4080Sup to 8x	16GB	25	$0.255
RTX 4070S Tiup to 8x	16GB	15	$0.265
RTX 5080up to 8x	16GB	33	$0.283
Titan RTX	24GB	2	$0.304
RTX 3090 Tiup to 2x	24GB	4	$0.389
RTX 4090up to 14x	24GB	130	$0.540
RTX 5090up to 8x	32GB	107	$0.723

37 models

Prices are the lowest available per-GPU hourly rate and move with supply and demand. Multi-GPU nodes rent whole, so the amount charged is the per-GPU rate times the GPU count.

02 — Head to head

Against three mainstream GPU clouds

Same card, same hourly billing, same on-demand terms — no reserved discounts or annual commitments. Competitor rates come from each vendor's public pricing page; the check date is noted below the table.

Against three mainstream GPU clouds
Model	NexGPU	AWS	CoreWeave	Lambda
H100 SXM80GB	$3.582	$6.880−47%p5.48xlarge	$6.160−41%HGX H100	$3.990−10%8x H100 SXM
H200141GB	$6.660	No per-GPU rate published	$6.310+6%HGX H200	Full cluster only
A100 SXM80GB	$1.088	$3.431−68%p4de.24xlarge	$2.700−59%A100	$1.990−45%1x A100 SXM
RTX 509032GB	$0.723	Not offered	Not offered	Not offered
RTX 409024GB	$0.540	Not offered	Not offered	Not offered

None of the three mainstream clouds offer consumer GPUs at all. For image generation, LoRA fine-tuning and small-model inference, the RTX 5090 and 4090 have no equal on price-performance — in this tier there is no comparison, only availability.

Competitor figures are each vendor's published US-region on-demand rate, normalised to a per-GPU hour. AWS is us-east-1 on-demand; CoreWeave and Lambda are their listed pricing-page rates. A green − means we are cheaper; an amber + means they are. H200 supply is currently thin and we do not win that tier — it is listed as it stands. Percentages round down throughout: our advantage is never inflated and our disadvantage is never shrunk. AWS · CoreWeave · Lambda

03 — How billing works

Three charges, all of them stated

The most common surprise on a GPU cloud bill is discovering compute was not the only line item. All three are written out here — including the one nobody else mentions.

Compute — metered per second, priced per hour
Starts when the instance reaches running state and stops the moment it is destroyed. Zero compute charge while stopped — we reconciled this against upstream billing second by second.
Storage — keeps running while stopped
Disk is billed per GB-month from the moment the instance is created — including during image download — and it continues after you stop the instance. This is industry standard; most platforms just do not say so. The network-wide median is about $0.414 per GB-month, which means a 200GB instance still costs around $2.76 for every stopped day. If you are done, destroy it — do not merely stop it.
Network — egress metered
Ingress and egress are metered separately and priced per GB, with a median around $0.0081 per GB. For normal training and inference this is typically under 5% of the bill; check the estimate before a large data export.

Why we lead with the storage charge

Because it is the only charge that keeps running when you are not using anything. The industry norm is to put it in clause 14 and never mention it again, letting the customer find it at month end. We put it in the middle of the pricing page, on the order form, on the instance card and in the low-balance email. A bill you can predict is worth more than a bill that merely looks cheaper.

04 — Run the numbers

What common jobs actually cost

Measured single-GPU run times multiplied by the starting rate. Substitute your own duration.

What common jobs actually cost
Job	GPU	Runtime	Cost
100 SDXL images at 1024px	RTX 4090$0.540/hr	~15 min	$0.14
One image-style LoRA fine-tune	RTX 4090$0.540/hr	~2 hours	$1.08
Transcribe 10 hours of audio, Whisper large-v3	RTX 4090$0.540/hr	~30 min	$0.27
QLoRA fine-tune of a 7B model, single GPU	RTX 5090$0.723/hr	~6 hours	$4.34
Serve a 70B model for a full day	A100 SXM4$1.088/hr	24 hours	$26
One week of H100 training, single GPU	H100 SXM$3.582/hr	168 hours	$602

Prices are the lowest available per-GPU hourly rate and move with supply and demand. Multi-GPU nodes rent whole, so the amount charged is the per-GPU rate times the GPU count. Runtimes are typical single-GPU magnitudes and vary with resolution, sequence length, dataset size and step count. Estimates only.

05 — Included

What we do not charge extra for

Line items that are commonly upsold elsewhere are inside the sticker price here.

Public direct ports
Every instance ships with a public IP and a direct port range. SSH, Jupyter and your own services map straight through — no port fees and no mandatory bastion hop.
Web terminal and file manager
Open a root shell in the browser, and optionally install a visual control panel for files, processes and cron with one click. No extra charge.
The whole template library
PyTorch, vLLM, ComfyUI, Whisper and the rest are all available. There is no "enterprise image" tier.
No minimums, no setup fee
No minimum rental period, no monthly floor, no instance creation fee. Run out of balance and it stops; top up and it resumes.

06 — FAQ

Pricing and billing

The six questions asked most often before a first order.

How is GPU rental billed — hourly or by the second?

Metered per second, priced per hour. Ninety minutes of runtime is charged as 1.5 hours; there is no rounding up to a whole hour. Compute charges begin when the instance reaches running state and end when it is destroyed.

Do I still pay after stopping an instance?

Compute charges stop, but disk storage continues to accrue per GB-month, because your data still occupies the host's drive. If you will not need it again soon, destroy the instance. If you need to keep the data, copy the results out first, then destroy.

Is $3.582/hr for an H100 a rate I can actually get?

It is the lowest per-GPU price across all H100 SXM nodes. Nodes of the same model vary by region, bandwidth and host reliability, so the marketplace lists every rentable node with its own live rate and you pay the price you see. Multi-GPU nodes rent whole: an 8-GPU node bills at the per-GPU rate times eight.

Why is this so much cheaper than AWS, CoreWeave or Lambda?

Different supply structure. Traditional clouds price against owned datacenters, so the rate carries facility construction, redundant standby capacity and long depreciation schedules. NexGPU aggregates distributed supply where utilisation is set by market matching rather than one vendor's capacity plan. The H100 is physically the same card — what differs is how the supply reaches you.

Is there a minimum top-up? Any hidden fees?

No minimum spend, no setup fee, no port fee, no image licensing fee. The bill has exactly three lines — compute, storage and network — and the instance detail page shows the running total and unit rate for each.

Can prices rise while my job is running?

No. The rate is locked at order time and holds until that instance is destroyed, even if the market moves up. New rates apply only to newly created instances.

Start building on NexGPU

Enterprise R&D team or solo developer — either way, your first job can be running in minutes.