GPU rental · Europe
Dedicated European GPU capacity for AI inference, development, agents and model workloads. Every rented machine is dedicated to you for the rental period, with full SSH access. It's the same hardware behind Managed GPU and GPU VM.
Hardware
| System | Availability | Memory | Price | |
|---|---|---|---|---|
| NVIDIA DGX Spark (GB10) | Available | 128 GB unified | from €0.85/hour | Check availability |
| NVIDIA RTX 6000 Pro | Request capacity | 96 GB GDDR7 | €1.58 / GPU-hr | Ask for this GPU |
| NVIDIA RTX 5090, 4-GPU systems | Available | 32 GB GDDR7 per card | €1.26 / GPU-hr | Ask for this GPU |
| NVIDIA RTX 3090, 4-GPU systems (budget tier) | Request capacity | 24 GB GDDR6X per card | €0.62 / GPU-hr | Ask for this GPU |
| NVIDIA RTX 3060, one card or the dual-GPU machine, the starter tier | Available | 12 GB GDDR6 per card | €0.41 / hour · one card €0.24 | Check availability |
| NVIDIA RTX 4080 | Request capacity | 16 GB GDDR6X | On request | Ask for this GPU |
| NVIDIA H100 | Request capacity | 80 GB HBM3 | On request | Ask for this GPU |
| NVIDIA H200 | Request capacity | 141 GB HBM3e | On request | Ask for this GPU |
| NVIDIA B200 | Request capacity | 192 GB HBM3e | On request | Ask for this GPU |
| NVIDIA DGX Station (GB300) | Request capacity | 748 GB coherent (252 GB HBM3e) | On request | Ask for this GPU |
€1.19/hour on demand · €0.95/hour by the week · €0.91/hour by the month · €0.85/hour by the year for a dedicated DGX Spark. Check availability or Reserve capacity for a scheduled start. RTX 3060 machines rent the same way, one card or both, from €0.29/hour. Check availability. Request-capacity systems show their launch price per GPU-hour; the deployment is confirmed with your request. Rentable systems are available as Managed GPU (we deploy and operate) or GPU VM (SSH, your stack).
Dedicated DGX Spark rentals run from Málaga (eu-es-1); RTX 3060 machines from Stockholm (eu-se-1), where AxForge serverless inference also runs.
How to read the table
Available
Rentable capacity. DGX Spark (GB10): book directly in the console by the hour, week, month or year, from €0.85/hour, hosted in Málaga, Spain (eu-es-1). Take the next available machine or reserve a start time. Labelled community benchmark data on the GB10 page.
Available
Starter tier, rentable. RTX 3060 machines, one 12 GB card or both, from €0.29/hour for the pair. Rent the same way as the GB10: check availability in the console and take the next machine or reserve a start time.
Request capacity
Capacity on request. RTX 6000 Pro, RTX 5090, RTX 3090, RTX 4080, H100, H200, B200 and the DGX Station: request dedicated capacity through the console at the listed launch price. Registered demand is used to prioritize new capacity, and machines are sized to the workloads that asked for them.
Rather have tokens?
If you want an API rather than a machine, the shared OpenAI-compatible API serves Qwen3.8 27B and more from Stockholm. See all models.
Community data
The llama.cpp CUDA scoreboard runs the identical workload (Llama 2 7B Q4_0, tg128, full GPU offload) across NVIDIA cards, the cleanest apples-to-apples comparison there is. Cards AxForge offers are linked.
| GPU | VRAM | Decode | vs RTX 3060 | Prompt processing |
|---|---|---|---|---|
| RTX 4060 Ti | 8 GB | 63.9 t/s | 85% | N/A |
| RTX 3060, our starter tier | 12 GB | 75.6 t/s | 100% | 2,138 t/s |
| RTX 5060 Ti | 16 GB | 90.9 t/s | 120% | N/A |
| RTX 4070 Ti SUPER | 16 GB | 132.3 t/s | 175% | 6,925 t/s |
| RTX 4080 | 16 GB | 142.5 t/s | 188% | N/A |
| RTX 4080 SUPER | 16 GB | 148.3 t/s | 196% | 8,125 t/s |
| RTX 3090 | 24 GB | 158.2 t/s | 209% | 5,175 t/s |
| RTX 3090 Ti | 24 GB | 171.2 t/s | 227% | N/A |
| RTX 6000 Ada | 48 GB | 176.1 t/s | 233% | N/A |
| RTX 5070 Ti | 16 GB | 176.9 t/s | 234% | N/A |
| RTX 5080 | 16 GB | 182.0 t/s | 241% | 8,297 t/s |
| RTX 4090 | 24 GB | 186.2 t/s | 246% | 11,993 t/s |
| RTX 6000 Pro (PRO 6000 Blackwell) | 96 GB | 274.2 t/s | 363% | 14,855 t/s |
| RTX 5090 | 32 GB | 290.0 t/s | 384% | 14,073 t/s |
| DGX Station (GB300) | 252 GB | ~300 t/s* | ~400% | ~10,500 t/s* |
Four times, not ten
The top card is about 4× an RTX 3060 for ordinary single-stream decode, not ten times. What the expensive cards actually buy is capacity and throughput under larger workloads, not just one user's tokens per second.
Capacity beats clocks
The moment a model no longer fits in VRAM and starts spilling, decode collapses: a 96 GB RTX 6000 Pro runs a 27B dense model ~6× faster than a 5090, not because of faster silicon but because it never offloads.
Community measurements (2026-09), not ours; results vary by runtime, quantization, context and serving configuration. GB10 / DGX Spark is a different architecture that doesn't belong in this ladder. Its community data lives on the GB10 page. * DGX Station: no public entry yet, estimated from the B200 (298 t/s) and GH200 (312 t/s) entries of the same test.
Community data · scaling out
The same single-stream workload, scaled to the multi-card systems AxForge offers and to the datacenter parts you can request. Multi-card rows are linear extrapolations of the single-card community figures; * marks estimates for parts without a public llama.cpp entry. Total memory is what the whole system can hold.
| Configuration | Total memory | Aggregate decode | vs 1× RTX 3060 |
|---|---|---|---|
| 1× RTX 3060 | 12 GB | 75.6 t/s | 1.00× |
| 2× RTX 3060 | 24 GB | 151 t/s | 2.00× |
| 4× RTX 3060 | 48 GB | 302 t/s | 4.00× |
| 1× RTX 3090 | 24 GB | 158 t/s | 2.09× |
| 2× RTX 3090 | 48 GB | 316 t/s | 4.19× |
| 4× RTX 3090 | 96 GB | 633 t/s | 8.37× |
| 1× RTX 4090 | 24 GB | 186 t/s | 2.46× |
| 2× RTX 4090 | 48 GB | 372 t/s | 4.93× |
| 4× RTX 4090 | 96 GB | 745 t/s | 9.85× |
| 1× RTX 5090 | 32 GB | 290 t/s | 3.84× |
| 2× RTX 5090 | 64 GB | 580 t/s | 7.67× |
| 4× RTX 5090 | 128 GB | 1,160 t/s | 15.34× |
| 1× RTX PRO 6000 Blackwell | 96 GB | 274 t/s | 3.63× |
| 2× RTX PRO 6000 | 192 GB | 548 t/s | 7.25× |
| 4× RTX PRO 6000 | 384 GB | 1,097 t/s | 14.51× |
| 1× GB10 | 128 GB | 60 t/s | 0.79× |
| 2× GB10 | 256 GB unified | ~120 t/s | 1.59× |
| 4× GB10 | 512 GB unified | ~240 t/s | 3.17× |
| 1× A100 80GB | 80 GB | 191 t/s | 2.52× |
| 2× A100 80GB | 160 GB | 382 t/s | 5.05× |
| 4× A100 80GB | 320 GB | 764 t/s | 10.10× |
| 1× H100 80GB | 80 GB | 268 t/s | 3.54× |
| 2× H100 | 160 GB | ~536 t/s | 7.08× |
| 4× H100 | 320 GB | ~1,071 t/s | 14.17× |
| 1× H200 | 141 GB | 308 t/s | 4.07× |
| 2× H200 | 282 GB | ~616 t/s | 8.14× |
| 4× H200 | 564 GB | ~1,231 t/s | 16.29× |
| 1× MI300X | 192 GB | ~233 t/s | 3.08× |
| 2× MI300X | 384 GB | ~466 t/s | 6.16× |
| 4× MI300X | 768 GB | ~932 t/s | 12.33× |
| 1× B200 | 192 GB | 298 t/s | 3.94× |
| 2× B200 | 384 GB | ~595 t/s | 7.88× |
| 4× B200 | 768 GB | ~1,191 t/s | 15.75× |
| 1× B300 | 288 GB | ~300 t/s* | ~4.0× |
| 2× B300 | 576 GB | ~600 t/s* | ~7.9× |
| 4× B300 | 1.15 TB | ~1,200 t/s* | ~15.9× |
| 1× DGX Station (GB300) | 748 GB coherent | ~300 t/s* | ~4.0× |
| 2× DGX Station | 1.5 TB (748 GB a box) | ~600 t/s* | ~7.9× |
| 4× DGX Station | 3 TB (748 GB a box) | ~1,200 t/s* | ~15.9× |
Community and extrapolated figures (2026-09), not ours; "~" = approximate, "*" = estimated. Aggregate decode assumes independent streams per card. A single request does not get faster by adding cards, but more concurrent requests do. On this test the fastest cards all land near 300 t/s: one short stream is held back by latency, not memory speed.
Which card for what
| System | Memory | Role |
|---|---|---|
| RTX 3060 starter systems | 12 GB / card | Cheap 7B to 8B inference, embeddings, experimentation. |
| RTX 3090, single card or multi-GPU | 24 GB / card | Exceptional value: 20 to 35B quantized models on ~936 GB/s of bandwidth. |
| RTX 5090, single card or multi-GPU | 32 GB / card | Fastest-per-euro for models and MoEs that fit in 32 GB. |
| RTX 6000 Pro | 96 GB | A different class: large models stay entirely on one GPU, with high bandwidth behind them. |
| DGX Spark | 128 GB unified | The most capacity for the least money, fitting huge models over raw decode speed (~273 GB/s bandwidth). Available now, by the hour, week, month or year. |
| DGX Station (GB300) | 748 GB coherent | Datacenter Blackwell in one box: models up to about a trillion parameters, 252 GB on the GPU at 7.1 TB/s. |
Not sure which fits your model and traffic? Talk to an engineer.
Data & privacy
A rented system serves only you: your model, your traffic, our hardware. Prompts are never persisted. Nothing is logged, retained or used for training by AxForge; only operational metadata is kept for billing and operations. Full policy at axforge.ai/privacy.
Models
The catalogue shows the language models that fit each card, estimated from every model's own files: RTX 3060 (12 GB) · RTX 3090 (24 GB) · RTX 5090 (32 GB) · H100 (80 GB) · RTX 6000 Pro (96 GB) · DGX Spark (96 GB) · H200 (141 GB) · B200 (192 GB) · DGX Station (252 GB). Image, video and embedding models have the same filter.
FAQ
GB10 and RTX 3060 capacity can be booked directly in the console. RTX 6000 Pro, RTX 5090, RTX 3090, RTX 4080, H100, H200, B200 and the DGX Station can be requested through the console.
Hourly, daily, weekly and monthly dedicated GPU rental. Use capacity immediately when available, or reserve a scheduled start in the AxForge Console. Check availability.
A dedicated NVIDIA DGX Spark (GB10) is €1.19/hour on demand · €0.95/hour by the week · €0.91/hour by the month · €0.85/hour by the year. Request-capacity systems are quoted per deployment. Full list on the pricing page.
Request dedicated capacity for these GPUs through the AxForge Console. Registered demand is used to prioritize new capacity. See H100, H200, B200, DGX Station.
Dedicated DGX Spark rentals run from Málaga (eu-es-1); RTX 3060 machines from Stockholm (eu-se-1), where AxForge serverless inference also runs.
Dedicated. A rented system serves only your traffic, with full SSH access: your model, your traffic, our hardware. The shared pay-per-token API is a separate product under /models/.
No. Prompts and completions are processed on your dedicated machine and never persisted, logged or used for training by AxForge. Only operational metadata is kept for billing and operations. See the privacy policy.