Prices at a glance
Each provider’s entry GPU offer, and what it costs if the machine runs 24/7 (720 h/month).
| Provider | Real VM or server (not a container) | Type | Entry GPU | VRAM / FP32 | Price | ≈ 24/7 per month |
|---|
| Linux GPU iGPU Start | ✓ | KVM VM · time-shared GPU* · dedicated vCPU, RAM and SSD | Intel UHD 770 (shared, Perf 1) | Shared RAM / ≈ 0.74 TFLOPS | €14.99/mo flat | €14.99 |
|---|
| Linux GPU GTX 1060 | ✓ | KVM VM · dedicated GPU | NVIDIA GTX 1060, dedicated | 6 GB / ≈ 4.4 TFLOPS | €49.99/mo flat | €49.99 |
|---|
| Hetzner GEX45 | ✓ | Dedicated physical server · everything is yours | NVIDIA RTX PRO 4000 Blackwell SFF (dedicated server) | 24 GB | €214/mo + €209 setup | €214 |
|---|
| OVHcloud Public Cloud | ✓ | VM · full dedicated GPU | Quadro RTX 5000 (entry) · L4 | 16 GB / ≈ 11 TFLOPS · 24 GB / ≈ 30 TFLOPS | from $0.60/h · from $1.00/h | ≈ $432 · ≈ $720 |
|---|
| Vultr Cloud GPU | ✓ | VM · full dedicated GPU | NVIDIA L40S | 48 GB / ≈ 92 TFLOPS | $1.67/h | ≈ $1,202 |
|---|
| Vast.ai (marketplace) | ✗ | Docker container (VM rarely) · full GPU, but the host’s CPU, RAM, disk and network are shared | RTX 3060 → RTX 4090, on third-party hosts | 12–24 GB | RTX 3060 from ≈ $0.04/h · RTX 4090 ≈ $0.34–0.48/h | ≈ $26–343 + disk (≈ $0.20/GB/month) and traffic |
|---|
* Shared GPU: several customers may use the same physical GPU. The split is fixed by SR-IOV (hardware GPU virtualization): each VM gets its own share of the GPU, sized by the plan you choose (Perf 1, 2 or 3). Public prices checked on 29 October 2026, excluding VAT, on provider pages or price trackers (sources at the bottom of the page). They change often: check with each provider. FP32 TFLOPS: approximate vendor figures.
Powerful GPUs, priced for powerful workloads
An L40S (48 GB, ≈ 92 TFLOPS FP32) or an RTX PRO Blackwell is built to train models, serve 30–70B-parameter LLMs or do heavy rendering. The price follows: from €214 to over $1,200 a month. If what you need is to transcode videos or run Chrome with WebGL, 95% of that power sits idle.
The 24/7 trap: hourly billing
$0.60/h sounds cheap. But a Jellyfin server, a browser fleet or a CI runner runs all month: 720 hours. That is ≈ $432/month for OVHcloud’s Quadro RTX 5000, ≈ $720 for an L4, ≈ $1,202 for Vultr’s L40S. The iGPU Start GPU VPS costs €14.99/month: roughly 25 to 75 times less for an always-on service.
What does a typical workload actually need?
What matters is not the TFLOPS count but the right hardware block:
| Workload | What it really needs | Right plan |
|---|
| Jellyfin / Plex / FFmpeg transcoding (1080p, 4K HEVC) | A hardware encoder (QuickSync or NVENC). TFLOPS are not used. | iGPU Start or Power |
|---|
| Chrome / Playwright / AI agents with WebGL | A GPU for compositing and WebGL, otherwise software rendering (SwiftShader) and a saturated CPU | iGPU Start |
|---|
| Vision, OCR, embeddings, quantized 3–8B LLMs | OpenVINO or Vulkan, 4–8 GB of memory | iGPU Power or Pro |
|---|
| Unity, 3D streaming, CUDA, NVENC | A dedicated NVIDIA GPU | GTX 1060 (€49.99) |
|---|
| Training, 30–70B LLMs, multi-GPU | 24–96 GB of VRAM, Tensor Cores | Hetzner GEX, OVHcloud, Vultr |
|---|
Vast.ai: cheap, but containers on shared machines
Vast.ai can be very cheap: on 30 October 2026, an RTX 3060 from $0.036/h (only 3 offers) and an RTX 4090 between $0.34 and $0.48/h. But these are not VPSs: they are Docker containers. You share the host’s kernel and you are root inside the container, not on the machine: no kernel modules, no regular systemd, forwarded ports, often no static IP. The GPU is yours, but the machine’s CPU, RAM, disk and network are shared with the customers renting its other GPUs. Disk (≈ $0.20/GB/month, billed even while the instance is stopped) and traffic come on top of the hourly price. The cheapest offers are on unverified hosts, often in China, Thailand or Russia. Great for experiments, less so for a 24/7 production service or GDPR-bound data.
Our honest recommendation
Start small: a €14.99/month GPU VPS covers transcoding, GPU browsers and light AI. If you need CUDA or a dedicated GPU, move to the GTX 1060. When you train models or serve large LLMs, the GPUs from Hetzner, OVHcloud or Vultr become the right choice, and they are worth their price.
See the GPU VPS in detail →
FAQ
Why is an integrated GPU so cheap?
The Intel UHD 770 is part of the server’s CPU: no card to buy, a few watts of power, and the GPU is time-shared between a few VMs. The price reflects that real cost.
Is transcoding slower than on an L4?
For one or a few streams the difference is small: the fixed-function encoder does the work, not the compute units. An L4 handles many more parallel streams and AV1 encoding, which matters for a large-scale video platform.
Can I move to a bigger plan later?
Yes: from iGPU Start to Power or Pro, or to the GTX 1060. We help you migrate your setup.
Hetzner GPU matrix · OVHcloud Cloud GPU · Vultr GPU (getdeploying.com) · Vast.ai pricing · Vast.ai docs: billing