GPU-1
RTX 4090 — inference & LoRA
119 $/month
- 1× NVIDIA RTX 4090 (24 GB GDDR6X)
- 16 dedicated AMD EPYC vCPUs
- 64 GB DDR4 RAM
- 1 TB NVMe Gen4
- 1 Gbps unlimited
- CUDA 12 + pre-installed drivers
GPU line — offshore AI hosting
GPU LINE EST. 2023
Offshore GPU for AI
Training, fine-tuning, and inference on RTX 4090 and H100, in jurisdictions that ask you for neither name nor ID. CUDA 12 preinstalled, PyTorch and TensorFlow images ready to use, crypto-only billing.
No KYC
Crypto only
CUDA 12 + NVIDIA drivers
24/7 NOC support
Catalog
Three configurations, zero paperwork
From solo 4090 for inference to H100 SXM5 for serious training. Each machine is dedicated: the GPU you rent is the GPU you get, at 3% or 97% load.
GPU-2
2× RTX 4090 — parallelization & 70B+
229 $/month
- 2× NVIDIA RTX 4090 (48 GB total)
- 32 dedicated AMD EPYC vCPUs
- 128 GB DDR4 RAM
- 2 TB NVMe Gen4
- 10 Gbps unlimited
- PyTorch / TensorFlow / JAX images
GPU-3
H100 SXM5 — frontier training
On quote
- 1× NVIDIA H100 SXM5 (80 GB HBM3)
- High-density AMD EPYC vCPUs
- Up to 512 GB RAM
- NVMe Gen4 up to 8 TB
- 10 Gbps+ unlimited
- Priority provisioning within 48 h
Provisioning within 24 business hours after first crypto confirmation — dedicated machines, no noisy neighbors, no KYC.
0
GPUs rented since 2023
0
GPU line uptime — Q2 2026
24/7
Human NOC support
0
GPU jurisdictions (NL · RO · CH)
Use cases
What our clients run at 100% VRAM
The GPU line was born in 2023 from a simple demand: doing serious AI without linking a bank card, a named cloud account, or a phone number to an inference provider.
Self-hosted LLM inference
Serve Llama, Mistral, Qwen, or DeepSeek from your own hardware with vLLM, TGI, or Ollama. Zero third-party requests, zero per-token billing, long contexts without queues.
Stable Diffusion & image generation
ComfyUI or AUTOMATIC1111 on 24 GB of VRAM: SDXL, Flux, ControlNet, style LoRA, and permanent rendering queues. Your checkpoints and grids stay with you, on your NVMe.
Fine-tuning & LoRA
QLoRA on RTX 4090 24 GB, full fine-tuning on H100 80 GB. Personal datasets, checkpoints stored locally, nothing goes through a shared GPU rented by the minute.
24/7 AI Agents
Autonomous workers, task orchestration, tool queues, LLM-driven automations: a dedicated machine that runs while you sleep, monitored and rebooted by our NOC.
Rendering & hybrid computing
Blender Cycles, OctaneRender, CPU+GPU simulations, and heavy batches. The 10 Gbps of the GPU-2 swallows large scenes and datasets without breaking a sweat.
Deployment
From delivery to first token in four lines
Each GPU comes with Ubuntu 22.04/24.04, CUDA 12, NVIDIA drivers, and the NVIDIA Container Toolkit. Real example on a freshly provisioned GPU-1 in Rotterdam:
bash — GPU-1 · Ollama
# SSH connection — credentials shown only once after payment
$ ssh [email protected]
$ nvidia-smi -L
# GPU 0: NVIDIA GeForce RTX 4090 (24 Go) — pilote 560.35, CUDA 12.6
$ docker run -d --gpus all -p 11434:11434 --name ollama ollama/ollama
$ docker exec ollama ollama run llama3.1:8b-instruct-q8_0
$ curl -s http://localhost:11434/api/generate \
-d '{"model":"llama3.1:8b-instruct-q8_0","prompt":"Pourquoi héberger offshore ?","stream":false}'
The same flow works with vLLM, TGI, or a custom PyTorch container. Factory reset (Ubuntu + CUDA) in one click from the dashboard, no ticket, no justification.
Technical specifications
Detailed specifications
Values measured at the card level (board power), not marketing aggregates. What is written here is what nvidia-smi
| Specification | GPU-1 | GPU-2 | GPU-3 |
|---|---|---|---|
| Accelerator | 1× NVIDIA RTX 4090 | 2× NVIDIA RTX 4090 | 1× NVIDIA H100 SXM5 |
| GPU memory | 24 GB GDDR6X | 2× 24 GB GDDR6X | 80 GB HBM3 |
| Memory bandwidth | 1,008 GB/s | 2× 1,008 GB/s | 3.35 TB/s |
| CUDA cores | 16 384 | 32 768 | 16 896 |
| TDP (board power) | 450 W | 2× 450 W | 700 W |
| Host processor | 16 vCPU AMD EPYC | 32 vCPU AMD EPYC | Up to 64 vCPU AMD EPYC |
| Host RAM | 64 GB DDR4 | 128 GB DDR4 | Up to 512 GB |
| Storage | 1 TB NVMe Gen4 | 2 TB NVMe Gen4 | 4–8 TB NVMe Gen4 |
| Network | 1 Gbps unmetered | 10 Gbps unmetered | 10 Gbps+ unmetered |
| Anti-DDoS | Included (L3–L7) | Included (L3–L7) | Included (L3–L7) |
| Locations | NL · RO · CH | NL · RO · CH | NL · RO · CH (NL priority) |
| Price | $119/month | $229/month | On request |
Short answers
Questions asked before every GPU order
Do you offer hourly billing?
Not yet. The GPU line is billed monthly, with no commitment, cancellable at any time from the dashboard. Hourly billing is planned for the second half of 2026 — current GPU customers will get priority access at the preferential rate.
Do you offer spot instances or burst capacity?
No, and that's intentional. Our GPUs are dedicated: no one can claim your card mid-training, and your performance doesn't depend on the current neighbor. The displayed price is the price you pay, from the first to the last day of the month.
Which frameworks and environments are supported?
PyTorch 2.x, TensorFlow 2.x, JAX, ONNX Runtime, vLLM, TGI, Ollama, ComfyUI — with CUDA 12 and NVIDIA drivers pre-installed. Base images are Ubuntu 22.04 and 24.04; you can install any distribution or Docker image via NVIDIA Container Toolkit.
Is the Russia location available for GPU?
No. The GPU line is only offered in the Netherlands, Romania, and Switzerland — only these three sites have the power and cooling suited for high-consumption cards. The six jurisdictions remain available for CPU VPS and RDP.
Related pages
Your next run can start this week.
Dedicated GPU, crypto-only, zero KYC. Choose a configuration, pay, receive your credentials — and let the 24 GB heat up.
Choose my GPU