A whole DGX Spark.
€0.99 an hour.
Rent a DGX Spark AI workstation with NVIDIA Grace Blackwell, 128 GB of memory and a 4 TB SSD. Exclusively yours, hosted in the EU, booked by the hour from 2 hours.
First booking: your first day is 30 % off — €0.693/h excl. VAT for up to 24 h.
Why GPUhire.eu
Your workloads and data never leave the EU.
The machine sits in Czechia, run by a Czech company under EU law. No hyperscaler in between, nobody else on the box.
Your own machine, not a slice
The whole box for your booking. No sharing, no queue.
Fixed price, no surprises
Prepaid. No subscription, no per-token billing, no egress fees.
Models up to ~200B
Quantized, in 128 GB shared by CPU and GPU.
Clean-room handover
Every booking starts from a fresh image; the SSD is wiped after.
Root over SSH
Paste your public key once. No passwords.
By the hour, from 2 hours
2 hours to a year, starting at any hour. Extend from the console.
What you can run
Not only LLMs. Anything that wants a CUDA GPU, lots of memory or Arm cores, for as many days as it takes.
Run and fine-tune open models up to ~200B.
Quantized models up to ~200B parameters in 128 GB of unified memory. Serve them, test them, or LoRA fine-tune over days instead of hours.
- vLLM
- Ollama
- llama.cpp
- PyTorch
root@spark:~# vllm serve [model] --max-model-len 32768
Generate big. Batch bigger.
Image and video models that don’t fit on a 24 GB gaming card, at high resolution and in large batches.
- ComfyUI
- Stable Diffusion
- Flux
- video models
root@spark:~# python main.py --listen # ComfyUI
Transcribe a whole archive overnight.
Speech-to-text across thousands of hours, text-to-speech and voice models. Book a day, queue it, collect the results.
- Whisper
- faster-whisper
- TTS models
root@spark:~# whisper ./archive/*.mp3 --model large-v3
Dataframes on the GPU.
GPU versions of pandas and scikit-learn on datasets that fit in 128 GB. No cluster to set up.
- RAPIDS cuDF
- cuML
- Jupyter
root@spark:~# python -c "import cudf; cudf.read_parquet('data/')"
Train on large image sets.
Detection, segmentation and classification models trained and evaluated on your own data, which stays in the EU.
- PyTorch
- Ultralytics YOLO
- OpenCV
root@spark:~# yolo train data=[dataset].yaml epochs=100
A 20-core Arm build machine.
Build and test arm64 software, Docker images and edge targets on native Arm cores with 4 TB of disk. No GPU needed.
- Docker buildx
- CI runner
- cross-compilers
root@spark:~# docker buildx build --platform linux/arm64 .
Single-precision science, for days.
Molecular dynamics, Monte Carlo and other FP32 GPU workloads that need to run uninterrupted.
- GROMACS
- CUDA
- Python
root@spark:~# gmx mdrun -deffnm [run] -nb gpu
Simulate and train robot policies.
NVIDIA’s robotics simulation and reinforcement-learning stack on the platform it targets.
- Isaac Sim
- Isaac Lab
root@spark:~# ./isaaclab.sh -p [train].py
Hardware
GIGABYTE AI TOP ATOM — an NVIDIA DGX Spark system, 4 TB model.
GB10 Grace Blackwell · 20 Arm v9 cores · 1 PFLOP FP4
128 GB LPDDR5x unified · 273 GB/s · 4 TB NVMe Gen 4 · 1 Gbps, no data limit
Full specs +Hide specs −
Hardware
GIGABYTE AI TOP ATOM — an NVIDIA DGX Spark system, 4 TB model.128 GB LPDDR5x unified · 273 GB/s · 4 TB NVMe Gen 4 · 1 Gbps, no data limit
- Superchip
- NVIDIA GB10 Grace Blackwell
- AI compute
- 1 petaFLOP (1,000 AI TOPS, FP4)
- CPU
- 20 Arm v9 cores (10 Cortex-X925 + 10 Cortex-A725)
- GPU
- Blackwell, 5th-gen Tensor Cores, 4th-gen RT Cores
- Memory
- 128 GB LPDDR5x unified, 256-bit, 273 GB/s
- Storage
- 4 TB NVMe SSD, PCIe Gen 4
- Internet
- 1 Gbps up and down, no data limit
- Software
- NVIDIA DGX OS (Ubuntu) with NVIDIA AI stack
- Model size
- Up to ~200B parameters (quantized)
Figures as published by GIGABYTE.
Pricing
Your first day is 30 % off.
Every hour of your first booking is 30 % off, up to 24 hours. Booked by the hour, from 2 hours up to a year, starting at any whole hour. Extend any time in the console. Paid from prepaid credit. Nothing renews.
Once per person or company. 30 % applies to bookings of up to 24 hours; longer bookings save less (see the table).
| Booking | Standard | First booking | Save |
|---|---|---|---|
| 2 hours | €1.98 | €1.39 | 30 % |
| 1 day | €23.76 | €16.63 | 30 % |
| 3 days | €71.28 | €64.15 | 10 % |
| 7 days | €166.32 | €159.19 | 4 % |
Prices excl. VAT. Consumers pay 21 % VAT on top.
You save €2.38 (30 %, excl. VAT) — 8 h at 30 % off.
All prices are shown excl. VAT. Individuals and unverified companies pay 21 % Czech VAT on top. Businesses in other EU countries with a VIES-verified VAT ID pay no VAT (reverse charge). Businesses outside the EU: write to us and we’ll set your account up without VAT. EUR accounts top up by bank transfer or card; CZK accounts by QR bank transfer, card, Google Pay or Apple Pay.
How it works
- 1
Sign up
With Google or email. Calendar and prices show right away.
- 2
Billing & top up
Pick EUR or CZK and add credit: card or bank transfer, plus Google Pay, Apple Pay or QR for CZK.
- 3
Book
Pick a start and how many hours, 2 or more. Paid from credit at once. Extend it later in the console.
- 4
Add your SSH key
Paste your public key. You can add it after booking.
- 5
Connect
When your booking starts, SSH in. It’s yours until it ends.
Booking and SSH keys live in the console at console.gpuhire.eu.
For businesses
An EU company in the Czech Republic, invoicing the way your finance team expects.
- Proper VAT invoices from a Czech company
- Reverse charge for VIES-verified EU VAT IDs
- Pay in EUR or CZK
FAQ
How do I connect to the machine?
How long can I book?
Is the machine shared?
How does the first-booking offer work?
Do EU businesses pay VAT?
Your first day is 30 % off.
€0.693/h excl. VAT for the first 24 hours of your first booking, then €0.99/h.
Book now