Last updated: July 2026. Every price below was re-checked against the provider's own public pricing page in July 2026.
The cheapest VPS with GPU (also called VPS server with GPU or VPS hosting with GPU) with a publicly posted rate in July 2026 starts at $0.34/hr for RTX 4090 (RunPod Community Cloud), $1.19/hr for A100 80GB, and $1.99/hr for H100 80GB PCIe. Marketplaces like Vast.ai and TensorDock go lower, but their prices float by host and cannot be quoted as a fixed number. This guide ranks the real options with per-second-vs-hourly billing notes, on-demand vs interruptible pricing, and the point where a monthly dedicated box beats hourly rental outright.
For free GPU options first, see Free GPU Compute. For buying hardware instead, see Best GPU for AI 2026.
A note on where these numbers come from
GPU pricing moves faster than most comparison articles get updated, and marketplace providers do not have a fixed price to quote at all. So two rules for this page:
- Every figure in the tables below is the provider's own published on-demand rate, checked in July 2026. Where a provider does not publish a rate for a given card, the cell says so instead of carrying a guess.
- Vast.ai publishes live prices only, and TensorDock's are set by independent hosts. Their cells say "live" rather than a number. Treating a floating marketplace price as a fixed headline is the single most common error in GPU comparison posts, including an earlier version of this one.
VPS GPU pricing (July 2026, per hour, published on-demand rates)
| GPU | Vast.ai | RunPod Community | RunPod Secure | Lambda | TensorDock | SynpixCloud |
|---|---|---|---|---|---|---|
| RTX 3090 24GB | live | n/a | n/a | n/a | live | $0.32 |
| RTX 4090 24GB | live | $0.34 | $0.69 | n/a | from $0.35 | $0.48 |
| RTX 5090 32GB | live | $0.69 | $0.99 | n/a | not listed | $0.79 |
| RTX Pro 6000 96GB | live | n/a | $1.99 | n/a | n/a | n/a |
| RTX A6000 48GB | live | n/a | n/a | $1.09 | live | $0.69 |
| L40 48GB | live | $0.69 | $0.82 | n/a | n/a | n/a |
| L40S 48GB | live | $0.79 | $0.99 | n/a | n/a | n/a |
| A100 40GB | live | n/a | n/a | $1.99 | live | $0.79 |
| A100 80GB PCIe | live | $1.19 | $1.39 | n/a | n/a | n/a |
| A100 80GB SXM | live | $1.39 | $1.49 | $3.99 | from $1.80 | $2.26 |
| H100 80GB PCIe | live | $1.99 | $2.89 | $3.29 | n/a | n/a |
| H100 80GB SXM | live | $2.69 | $2.99 | $3.99-$4.29 | from $2.25 | $2.79 |
| H200 141GB | live | $3.59 | $4.39 | n/a | n/a | not offered |
| B200 | live | $5.89 | $5.89 | $6.69-$6.99 | n/a | not offered |
Lambda also lists GH200 at $2.29/GPU-hr, which is the cheapest Hopper-class rate of any provider here that owns its own hardware.
For a public reference point on the expensive end: DigitalOcean (which now owns Paperspace) publishes H100 at $5.95/hr on-demand. Hyperscaler on-demand rates for the same card sit in that neighbourhood and vary by region and commitment, which is why the specialists win on raw cost for pure GPU rental.
Vast.ai - cheapest, marketplace model
Vast.ai is a marketplace where individual GPU hosts (homelabs, mining farms, small datacenters) list their hardware. Buyers rent; prices float based on supply. Result: usually the cheapest GPU prices anywhere, and no number this page could print would still be true next week.
- Catalogue: 68+ GPU types, RTX 3060 through B200.
- Tiers: On-Demand (per-second billing, no interruption), Interruptible (50%+ cheaper, can be outbid and reclaimed), Reserved (up to 50% off for 1, 3 or 6-month commitments).
- Trade-offs: Reliability varies by host. Some are excellent, some are not. Network throughput varies. You are renting from a stranger's datacenter, and the vetting is on you.
- Best for: Prototyping, batch training with checkpointing, hyperparameter sweeps, anything where cost matters more than uptime.
Check the live board before committing; that is the only honest way to price Vast.
RunPod - best UX, per-second billing, the reference price
RunPod is the convenience pick, and because it publishes a full public rate card it is the most useful yardstick in this comparison. Two tiers:
- Community Cloud: peer-provisioned hosts managed by RunPod. $0.34/hr RTX 4090, $1.19/hr A100 80GB PCIe, $1.99/hr H100 PCIe.
- Secure Cloud: RunPod-owned hardware in proper datacenters. $0.69/hr RTX 4090, $1.39/hr A100 80GB PCIe, $2.89/hr H100 PCIe.
Why teams choose RunPod:
- Per-second billing. Short experiments don't get billed for a full hour.
- Templates. One-click deploy of PyTorch, ComfyUI, Stable Diffusion WebUI, Ollama and others.
- Persistent volumes. Keep data between sessions.
- API + CLI. Automate spin-up and shutdown from CI.
Note that RunPod's A100 line is 80GB only; there is no 40GB SKU. If you see a 40GB RunPod price quoted somewhere, it is stale or invented.
Best for: Solo developers, small teams, anyone who wants GPUs to just work without per-instance reliability variance.
Lambda - ML team default, priced accordingly
Lambda Cloud targets ML engineers and owns its hardware. It is no longer the value pick it once was: A100 80GB SXM is $3.99/GPU-hr, nearly 3x RunPod Community for the same card, and H100 SXM runs $3.99-$4.29.
- Where it is competitive: GH200 at $2.29/GPU-hr, A6000 at $1.09, and reserved contracts for sustained multi-node training.
- UX: Clean Jupyter and SSH workflow, strong docs for ML practitioners, billing by the minute.
Best for: Funded teams doing multi-node training where reliability, support and reserved pricing matter more than the hourly rate. If you are renting a single card, you are overpaying here.
TensorDock - marketplace, now part of Voltage Park
TensorDock was acquired by Voltage Park in March 2025 and continues to operate as a marketplace of independent hosts who set their own prices. Entry points published on the site: RTX 4090 from $0.35/hr, A100 SXM4 from $1.80/hr, H100 SXM5 from $2.25/hr, older consumer cards from $0.12/hr. Minimum deposit to start is $5, and billing runs continuously while a server is deployed.
Best for: A second marketplace to price against Vast.ai, and useful when Vast inventory is tight.
SynpixCloud - small operator, sharp entry prices
Aggressive entry pricing on consumer and Ampere cards: RTX 4090 at $0.48/hr, RTX 3090 at $0.32, A6000 at $0.69, A100 40GB at $0.79. Prices climb steeply for multi-GPU nodes.
Two caveats worth stating plainly. Its A100 80GB rate ($2.26/hr) and H100 rate ($2.79/hr) are well above RunPod for the same silicon, so the "cheap" reputation only holds at the low end. And it does not offer H200 or B200 at all.
Best for: Cost-conscious single-card work on Ampere and consumer GPUs, without marketplace bidding.
Paperspace, now DigitalOcean
Paperspace is part of DigitalOcean. Published on-demand rates: H100 $5.95/hr, A6000 $1.89/hr, V100 $2.30/hr, RTX 5000 $0.82/hr, RTX 4000 $0.56/hr. Three-year commitments drop H100 to $2.24/hr and A100 80GB to $1.15/hr.
At on-demand rates this is the most expensive way to rent an H100 in this comparison. The strength is persistent storage, the notebook UX, and sitting inside DigitalOcean's wider platform if you already live there.
Best for: Teams already on DigitalOcean, or those who value persistent storage and notebook workflow over hourly cost.
Hourly rental vs a monthly dedicated GPU - the break-even
Almost every GPU comparison, including earlier versions of this one, silently assumes you rent by the hour. That assumption quietly loses money for anyone running an always-on workload: an inference endpoint, a bot, a render queue, an internal tool. A month is 730 hours. At $1.19/hr, a "cheap" A100 costs $869/month if it never sleeps.
Monthly dedicated GPU hosting is a different market with different economics. Converting plans to an hourly equivalent (price divided by 730) makes them directly comparable.
One thing to watch before you do that arithmetic: monthly GPU hosts commonly display the price for the longest prepayment term, not the month-to-month rate. GPU Mart's public pages show its two-year price by default; the month-to-month figure is 20-25% higher. Both are in the table below, because which one applies to you changes the answer.
| Plan | VRAM | Month to month | $/hr equiv | On 2-year prepay | $/hr equiv |
|---|---|---|---|---|---|
| GPU Mart RTX Pro 2000 VPS | 16GB | $119 | $0.16 | $99 | $0.14 |
| GPU Mart RTX Pro 4000 VPS | 24GB | $199 | $0.27 | $159 | $0.22 |
| GPU Mart RTX Pro 5000 VPS | 48GB | $349 | $0.48 | $269 | $0.37 |
| GPU Mart RTX 5090 VPS | 32GB | $449 | $0.62 | $399 | $0.55 |
| GPU Mart RTX Pro 6000 VPS | 96GB | $599 | $0.82 | $479 | $0.66 |
| GPU Mart A100 80GB dedicated | 80GB | $1,699 | $2.33 | $1,559 | $2.14 |
| GPU Mart H100 80GB dedicated | 80GB | $2,599 | $3.56 | $2,099 | $2.88 |
The break-even against hourly rental, in hours per month, month-to-month first and two-year prepay in brackets:
- RTX Pro 4000 24GB vs RunPod RTX 4090 at $0.34/hr: ~585 hours, about 80% of the month (on 2-year prepay: ~468 hours, 64%).
- RTX Pro 6000 96GB vs RunPod's RTX Pro 6000 at $1.99/hr: ~301 hours, about 41% of the month (on 2-year prepay: ~241 hours, 33%). Same card on both sides, so this is the cleanest comparison on this page.
- RTX 5090 32GB vs RunPod RTX 5090 at $0.69/hr: ~651 hours, about 89% of the month (on 2-year prepay: ~578 hours, 79%).
The rule: there is no single duty-cycle threshold, because it depends on how aggressively each side prices that specific card. The Blackwell Pro 6000 pays for itself at 41% of the month; a 24GB card needs 80%; a 5090 needs 89%. Run the division for the card you actually want before assuming either model is cheaper. What is consistent: below roughly a third of the month, hourly always wins, and the trap is renting by the hour for something that never turns off.
Storage is the other line item hourly pricing hides. RunPod bills persistent volumes separately at $0.10/GB per month while a pod runs and $0.20/GB per month while it sits idle, so keeping 400GB around adds $40-$80 a month on top of the GPU. Monthly plans generally bundle the disk: GPU Mart's Pro 6000 VPS includes 400GB SSD in the price above.
This cuts both ways. On frontier chips the monthly plans are not competitive at all: GPU Mart's A100 80GB is $1,699/mo month-to-month, or $2.33/hr, against RunPod Community's $1.19/hr for the same card. Its H100 at $2,599/mo works out to $3.56/hr against $2.99/hr for RunPod Secure H100 SXM, and only on a two-year term ($2,099/mo, $2.88/hr) does it draw level. The monthly advantage lives in the Blackwell RTX Pro tier, not at the top of the stack.
GPU Mart - monthly Blackwell, US-only
GPU Mart is the GPU brand of Database Mart LLC, a US host that has run its own hardware since 2005; the GPU line launched in 2021. Hardware sits in company-owned datacenters in Dallas, Texas and Kansas City, Missouri. The catalogue is 25+ GPU configurations across GPU VPS, Blackwell, dedicated and multi-GPU lines.
Where it is genuinely strong:
- The Blackwell RTX Pro VPS tier. RTX Pro 6000 with 96GB VRAM works out to $0.82/hr month-to-month, or $0.66/hr on a two-year term, if you keep it busy. RunPod rents the same card for $1.99/hr on Secure Cloud, so the monthly box is roughly 2.4x cheaper per hour, with the disk included, as long as it genuinely stays busy.
- Dedicated, not shared. GPU VPS instances get the physical NVIDIA card via PCIe passthrough with dedicated vCPUs, so there is no preemption and no bidding.
- Support. 24/7 in-house, with a claimed 99.9% uptime SLA and sub-5-minute live chat response.
Where it is not:
- Hourly rates are uncompetitive. A100 40GB runs $0.88-$1.11/hr against $0.79 at SynpixCloud and far less on the marketplaces. If your usage is spiky, rent elsewhere.
- US only. Two datacenters, both American. Latency-sensitive workloads in Europe or Asia should look elsewhere.
- Dedicated servers take about two hours to provision (VPS is about five minutes). This is not a platform for spinning something up for twenty minutes.
- No H200 or B200. The top of the range is H100 and RTX Pro 6000.
Best for: Always-on inference, hosted LLM endpoints, render queues and internal tools that run more than half the month, where a predictable monthly bill and a card nobody else can preempt is worth more than the lowest hourly rate. A free hour is available to test before committing.
On-demand vs interruptible - the savings math
Interruptible and spot instances cost 50%+ less than on-demand but can be reclaimed with little notice. Vast.ai's interruptible tier is the most aggressive; instances can be outbid and taken back.
Use interruptible for:
- Training with frequent checkpointing.
- Hyperparameter sweeps (many short jobs).
- Batch inference where each request is independent.
- Anything you can resume after interruption.
Don't use interruptible for:
- Customer-facing inference.
- Long fine-tuning runs without checkpointing.
- Demos or live workloads with strict uptime needs.
Picking the right GPU for your workload
RTX 4090 24GB ($0.34-$0.69/hr): 7B-13B fine-tuning, Stable Diffusion and video generation, embedding generation, most inference. The sweet spot for solo developers.
RTX 5090 32GB ($0.69-$0.99/hr): 30B at 4-bit dense, 70B MoE with low active params, bigger batch sizes. Worth the premium over the 4090 when 24GB hits the wall.
RTX Pro 6000 96GB ($1.99/hr on RunPod, or $0.82/hr equivalent on a monthly plan): 70B at 4-bit with room to spare, multi-LoRA serving, long-context inference. The monthly route only makes sense if the box stays busy.
A100 80GB ($1.19/hr+): 70B at 4-bit, 30B-40B at FP16, SDXL training, multi-LoRA serving.
H100 80GB ($1.99/hr+): 70B FP16, frontier-class fine-tuning, FP8 training, low-latency production inference.
H200 141GB ($3.59/hr+): When 80GB is not enough. 100B+ inference, longer-context training.
B200 ($5.89/hr+): Frontier training, very large MoE serving. Overkill for 95% of workloads.
Billing models - read the fine print
| Provider | Billing |
|---|---|
| Vast.ai | Per-second |
| RunPod | Per-second |
| Lambda | Per-minute |
| TensorDock | Continuous while deployed, $5 minimum deposit |
| SynpixCloud | Per-hour |
| Paperspace / DigitalOcean | Per-hour |
| GPU Mart | Per-hour, or monthly plans |
Billing granularity matters more than the headline rate for short runs. A five-minute job on a $5/hr per-second instance costs less than a $1.50/hr instance with a one-hour minimum.
Common mistakes when renting GPU VPS
- Renting H100 when RTX 4090 would do. A 6x cost premium for nothing. Right-size the GPU.
- Renting by the hour for an always-on workload. Depending on the card, a monthly box wins somewhere between 300 and 650 hours a month. Do the division for your card instead of assuming.
- Forgetting that hourly excludes storage. Persistent volumes are billed on top, and idle volumes cost more per GB than running ones.
- Reading a monthly host's headline price as the monthly price. It is usually the multi-year prepayment rate. Month-to-month is typically 20-25% higher.
- Skipping checkpointing on interruptible. Interrupted means work lost. Always checkpoint.
- Trusting a stale price table. GPU pricing moved substantially between May and July 2026 on several providers. Check the provider's own page before you commit.
- Forgetting egress. Some providers charge for outbound data. Check before pulling 100GB of weights.
- Hyperscaler-by-default. Unless you need the surrounding cloud, you are overpaying.
- No storage budget. Persistent volumes cost money. Clean up unused ones.
Quick setup recipe (RunPod, 5 minutes)
- Sign up at runpod.io. Add credits.
- Pick a GPU and a pod template (PyTorch, Ollama, ComfyUI and so on).
- Launch. The pod boots in 30-60 seconds.
- Connect via SSH or the web terminal.
- Run the workload. Stop the pod to stop billing.
Repeat the same flow on Vast.ai if cost beats UX, or provision a monthly box if step 5 never happens.
Related guides on this site
- OpenAI Free Credits 2026: Trial, OpenAI for Startups, Grove, Codex Open Source Fund
- Free LLM APIs in 2026: 13 Providers Compared (OpenAI, Claude, Gemini, Groq)
- $500K in Free Cloud Credits 2026: 15 Programs Compared (AWS, Google, Microsoft, Cloudflare)
- Free GPU Compute - Where to Get Hours Without a Credit Card
- Best GPU for AI 2026
- Cheap Dedicated Server in 2026
- Best Open Source LLM 2026 - Model Comparison
- Groq Pricing in 2026
- Stacking Startup Credits - Free Cloud + AI Hub
- HuggingFace Inference API 2026 - free tier, Endpoints, Providers
- OpenRouter Free Tier 2026 - 28+ free models, limits, BYOK
- H100 vs A100 in 2026 - real pricing + workload picks
- RunPod vs Lambda Labs vs Vast.ai - 3-provider GPU rental comparison (2026)
