AI A100
Fine-tuning & inference
$1,199.00/mo billed monthly — cancel anytime
- NVIDIA A100
- 80 GB HBM VRAM
- 16 vCPU cores
- 120 GB RAM
- 1 TB NVMe Gen4
- CUDA + cuDNN pre-installed
- PyTorch & TensorFlow ready
- Full GPU isolation
NVIDIA GPU — CUDA ready
Infrastructure built for AI workloads
NVIDIA A100 and H100 with CUDA ready, PyTorch and TensorFlow pre-installed and NVLink between GPUs — start training your models in minutes, not days. Your GPU, fully isolated.
CUDA-ready on first boot
1,979 TFLOPS FP16 per H100
0 GB HBM VRAM per GPU
0 GB/s NVLink GPU interconnect
24/7 AI-savvy expert support
Choose the right GPU power for your workload
From lightweight inference to massive LLM training — dedicated GPUs, never shared, never time-sliced.
AI H100
Serious model training
$2,499.00/mo billed monthly — cancel anytime
- NVIDIA H100 SXM
- 80 GB HBM VRAM
- 26 vCPU cores
- 200 GB RAM
- 2 TB NVMe Gen4
- NVLink interconnect
- Everything in AI A100
- HBM3 — 3.35 TB/s bandwidth
- NVLink-ready for scaling
AI Cluster
Multi-GPU NVLink fabric
Price on request Custom clusters, sized with our engineers
- 4× NVIDIA H100 SXM
- 320 GB HBM VRAM
- 96 vCPU cores
- 768 GB RAM
- 8 TB NVMe Gen4
- NVLink interconnect
- 4× H100 over NVLink 600 GB/s
- Distributed training tuned
- Dedicated solutions engineer
Every AI Server includes
- CUDA Toolkit + cuDNN pre-installed
- PyTorch + TensorFlow ready
- Full GPU isolation
- High-speed NVMe for datasets
- Full root access
- Snapshots for checkpoints
- DDoS-protected endpoints
- 24/7 specialist support
Plan comparison
Same software stack everywhere — pick by VRAM, TFLOPS and interconnect.
| Feature | AI A100 | AI H100 | AI Cluster |
|---|---|---|---|
| GPU | |||
| GPU model | A100 80GB | H100 80GB SXM | 4× H100 SXM |
| VRAM | 80 GB HBM2e | 80 GB HBM3 | 320 GB HBM3 |
| FP16 compute | 312 TFLOPS | 1,979 TFLOPS | 7,916 TFLOPS |
| NVLink interconnect | ✓ | ✓ | |
| Compute | |||
| vCPU cores | 16 | 26 | 96 |
| System RAM | 120 GB | 200 GB | 768 GB |
| Storage & network | |||
| NVMe Gen4 storage | 1 TB | 2 TB | 8 TB |
| Network | 10 Gbps | 10 Gbps | 25 Gbps |
| Software | |||
| CUDA + cuDNN pre-installed | ✓ | ✓ | ✓ |
| PyTorch & TensorFlow ready | ✓ | ✓ | ✓ |
| Docker + NVIDIA toolkit | ✓ | ✓ | ✓ |
| Management | |||
| Full root access | ✓ | ✓ | ✓ |
| Snapshots | ✓ | ✓ | ✓ |
| Arabic & English panel | ✓ | ✓ | ✓ |
| Dedicated engineer | ✓ | ||
The hardware — NVIDIA A100 & H100
The GPUs the AI industry standardized on — dedicated to you, never fractioned, never time-shared.
- 80 GB HBM2e VRAM
- 312 TFLOPS (FP16)
- 2 TB/s memory bandwidth
Ideal for: fine-tuning, production inference, mid-scale training
- 80 GB HBM3 VRAM
- 1,979 TFLOPS (FP16)
- 3.35 TB/s memory bandwidth
Ideal for: large LLM training, multi-GPU clusters, research
What can you build?
Five workloads our GPU fleet runs every day — with the stack already in place.
Benchmark performance on every metric
The numbers that decide your training time — counted, not promised.
0 TFLOPS FP16 — H100
0 GB VRAM per GPU
3.35 TB/s Memory bandwidth — H100
0 GB/s NVLink interconnect
0 MB/s NVMe dataset reads
Your stack is already installed
No driver hunting, no CUDA version hell — boot, activate the environment, start training.
CUDA + cuDNN The foundation for every GPU framework — version-matched to the driver. Pre-installed 
PyTorch Pre-installed with full CUDA support — torch.cuda.is_available() is already true. Pre-installed 
TensorFlow GPU support enabled out of the box, XLA ready. Pre-installed 
Hugging Face Transformers + Diffusers — pull a model and run within minutes. Pre-installed 
JupyterLab Research and experiment in the browser, against the real GPU. One command away 
Docker + NVIDIA Container Toolkit for isolated, reproducible training environments. One command away Storage & network at GPU speed
A starved GPU is a wasted GPU — the data path keeps up.
High-speed NVMe 7,000 MB/s Dataset reads never bottleneck your epoch time.
10Gbps+ network 10–25 Gbps Move checkpoints and datasets in minutes, not hours.
NVLink interconnect 600 GB/s GPU-to-GPU bandwidth that makes multi-GPU training scale.
Your models stay yours
Training data and weights are business secrets — the platform treats them that way.
GPU capacity in European regions
AI Servers deploy in our European data centers. The ones below are on sale today.
Teams shipping models on Hostrena
Fine-tuned Llama 3 70B on A100 in under 8 hours — the NVLink setup made multi-GPU training genuinely seamless.
Running production inference at 10k requests/min — latency never exceeded 180ms even at peak.
Switched from a cloud giant — saved 35% monthly on our training runs with identical TFLOPS.
Frequently asked questions
Is CUDA pre-installed?
Yes — CUDA Toolkit and cuDNN ship pre-installed and version-matched to the driver. PyTorch and TensorFlow see the GPU on first boot.
What is the difference between A100 and H100?
H100 is the newer generation: ~6× FP16 compute (1,979 vs 312 TFLOPS), faster HBM3 memory (3.35 vs 2 TB/s) and a stronger transformer engine. A100 remains excellent value for fine-tuning and inference.
Can I use Docker with the GPU?
Yes — Docker plus the NVIDIA Container Toolkit are ready, so `--gpus all` works immediately for reproducible training environments.
How do I transfer large datasets to the server?
Over the 10–25 Gbps network via rsync/scp or object storage tooling. A 1 TB dataset lands in well under half an hour on the larger plans.
Can I connect multiple GPUs?
Yes — the AI Cluster tier wires 4× H100 over NVLink at 600 GB/s, tuned for distributed training. Talk to sales for larger fabrics.
Is Jupyter ready out of the box?
JupyterLab installs with one command and is documented in the panel quick-start — many teams run it behind the built-in firewall rules.
What VRAM is available per GPU?
80 GB HBM on both A100 (HBM2e) and H100 (HBM3) — enough for 70B-parameter inference with quantization, or training with model parallelism.
Is the GPU really dedicated?
Fully — no MIG fractions, no time-slicing, no sharing. The GPU appears in nvidia-smi as completely yours, because it is.
How do I store model checkpoints safely?
Snapshot the server at checkpoint milestones, or sync checkpoints off-box to object storage — both paths are documented in the panel.
Which regions have GPU capacity?
Frankfurt (FRA-1) and Helsinki (HEL-1) today — both GDPR-friendly EU regions. New regions are announced on the data centers page.
Can I run inference APIs publicly?
Yes — expose your ports behind the cloud firewall and DDoS protection. Many clients serve production traffic straight from their AI Server.
Is there an annual discount?
Yes — annual billing saves 15% on every GPU plan.
What does "price on request" mean for clusters?
Multi-GPU fabrics are sized with our engineers — GPUs, interconnect, storage and network are quoted as one tailored package.
Do you support spot/interruptible pricing?
Not yet — all GPU servers are reserved and persistent. For burst workloads, monthly billing with cancel-anytime keeps you flexible.
How fast can I start?
A100 and H100 servers are typically online within minutes of ordering; clusters are provisioned with an engineer within 1–2 business days.
Can I fine-tune open models like Llama or Mistral here?
That is the most common workload we see. A single A100 80GB fine-tunes 7B–13B parameter models comfortably with LoRA or QLoRA; full-precision tuning of larger models wants the H100 or a multi-GPU setup with NVLink. Datasets stage on local NVMe scratch space, checkpoints push to any S3-compatible storage, and the prebuilt PyTorch environment takes you from tokenizer to training inside the first hour.
Start training your models today
CUDA ready, PyTorch installed, a GPU that is entirely yours. Minutes to first epoch — no setup marathon.
A100 from $1,199/mo — annual billing saves 15%.