Solutions · AI & GPU compute
AI GPU servers for inference and fine-tuning, billed monthly.
Dedicated NVIDIA GPUs with 24 to 96 GB of memory, a whole server to yourself and a 10 Gbps unmetered port for datasets and model weights. No hourly meter, no egress bill.
Recommended for large models
RTX PRO 6000
- NVIDIA RTX PRO 6000 · 96 GB
- AMD EPYC 9355P · 32C/64T
- 384 GB DDR5 ECC
- 2 × 3.84 TB NVMe Gen5
96 GB of GDDR7 on one GPU: 70B-parameter models at 8-bit, or LoRA fine-tuning of mid-size models.
$1,499 −40%$899/mo
Ready in 72 h
- 10 Gbps unmetered port
- Always-on DDoS protection
- IPMI / KVM access
- IPv4 + routed /64 IPv6
- $0 setup, cancel anytime
Why PowerDedi
Built for GPU workloads that run all month.
Monthly, not hourlyFrom $0.23 to $1.23 per hour equivalent when run around the clock, with nothing metered on top.
The whole machineNo time-slicing, no noisy neighbors: full PCIe access, your drivers, your CUDA version.
Data moves freePull weights and datasets at 10 Gbps and serve inference traffic without watching an egress meter.
Private by designYour models, prompts and data stay on hardware only you access, paid for with an email address and crypto.
Sizing
What fits on each GPU.
Largest model class that fits for inference with a moderate context length. Weights take about 2 bytes per parameter at FP16/BF16, 1 byte at 8-bit and about 0.5 byte at 4-bit, plus 10–30% for the KV cache.
| GPU | Memory | FP16 / BF16 | 8-bit | 4-bit | Per hour, 24/7 |
|---|---|---|---|---|---|
| NVIDIA RTX PRO 4000 Blackwell | 24 GB | ~8B | ~14B | ~32B | $0.23 |
| NVIDIA L40S | 48 GB | ~20B | ~32B | ~70B | $0.75 |
| NVIDIA RTX PRO 6000 Blackwell | 96 GB | ~40B | ~70B | ~120B (MoE) | $1.23 |
Rules of thumb for single-GPU inference. Long contexts, large batches and training need more memory; fine-tuning with LoRA or QLoRA fits well on 48 and 96 GB cards.
The GPUs
Three NVIDIA GPUs, one per server.
Public NVIDIA specifications. Every GPU server includes the host CPU, memory and NVMe listed, plus a 10 Gbps unmetered port.
| GPU | Architecture | Memory | Tensor formats | Host | Delivery | Price |
|---|---|---|---|---|---|---|
| NVIDIA RTX PRO 4000 Blackwell SFF · 24 GB | Blackwell | 24 GB GDDR7 ECC | FP4 · FP8 · BF16 | AMD Ryzen 9 9900X · 64 GB DDR5 | 24 h | $169/mo |
| NVIDIA L40S · 48 GB | Ada Lovelace | 48 GB GDDR6 ECC | FP8 · BF16 | AMD EPYC 9355P · 256 GB DDR5 ECC | 72 h | $549/mo |
| NVIDIA RTX PRO 6000 Blackwell · 96 GB | Blackwell | 96 GB GDDR7 ECC | FP4 · FP8 · BF16 | AMD EPYC 9355P · 384 GB DDR5 ECC | 72 h | $899/mo |
Runs your stack
- PyTorch
- vLLM
- SGLang
- llama.cpp
- Ollama
- TensorRT-LLM
- Hugging Face TGI
- ComfyUI
- Blender
- FFmpeg (NVENC)
- Docker + NVIDIA Container Toolkit
Recommended servers
Start from a proven configuration.
Every price includes the 10 Gbps unmetered port, DDoS protection and IPMI. The struck price is the market median for the same hardware.
Inference and development
24 h
- NVIDIA RTX PRO 4000 · 24 GB
- AMD Ryzen 9 9900X · 12C/24T
- 64 GB DDR5
- 2 × 1.92 TB NVMe
- 10 Gbps unmetered
Models up to about 14B at 8-bit, embeddings, Whisper and image generation on 24 GB of GDDR7.
Deploy$279−39%$169/moAvailable in Amsterdam, Frankfurt, New York.
Production inference
72 h
- NVIDIA L40S · 48 GB
- AMD EPYC 9355P · 32C/64T
- 256 GB DDR5 ECC
- 2 × 3.84 TB NVMe
- 10 Gbps unmetered
48 GB for quantized 30B to 70B models and batched serving, on a 32-core EPYC host.
Deploy$909−40%$549/moAvailable in Amsterdam, Frankfurt.
Large models and fine-tuningRecommended
72 h
- NVIDIA RTX PRO 6000 · 96 GB
- AMD EPYC 9355P · 32C/64T
- 384 GB DDR5 ECC
- 2 × 3.84 TB NVMe Gen5
- 10 Gbps unmetered
96 GB for 70B models at 8-bit and LoRA fine-tuning, with 384 GB of host memory.
Deploy$1,499−40%$899/moAvailable in Amsterdam, Frankfurt.
FAQ
AI & GPU compute: questions, answered.
Still have a question? Customers with an active server get answers from engineers in a support ticket.
Support ticketsCan I get several GPUs in one server?
Each GPU server has one dedicated NVIDIA GPU, without NVLink. To serve more traffic, run several GPU servers side by side.
Can I train models?
Fine-tuning (LoRA, QLoRA, full fine-tuning of smaller models) works well on the L40S and RTX PRO 6000. Pre-training large models needs multi-GPU clusters with fast interconnects, which we do not offer.
Which software can I run?
Anything that runs on Linux with NVIDIA drivers: PyTorch, vLLM, SGLang, llama.cpp, Ollama, TensorRT-LLM, ComfyUI and more. Install the driver and CUDA from NVIDIA’s repositories, or use Docker with the NVIDIA Container Toolkit.
Is GPU mining allowed?
No. GPU servers are reserved for compute, rendering and AI workloads.
Where are GPU servers located?
The RTX PRO 4000 is available in Amsterdam, Frankfurt and New York (24 h delivery); the L40S and RTX PRO 6000 in Amsterdam and Frankfurt (72 h).
Guides
Related guides
- gpt-oss-120b and gpt-oss-20b hardware requirements: VRAM and GPUsWhat OpenAI’s gpt-oss-120b and gpt-oss-20b need to run: weight sizes from the model card, the memory each token of context takes, single-GPU serving with vLLM and Ollama, documented GPU support, and which 24, 48 and 96 GB cards fit which model.
- RTX PRO 6000 rental in 2026: hourly cloud prices vs a monthly serverWhat an NVIDIA RTX PRO 6000 costs by the hour at the main GPU clouds and marketplaces in October 2026, the break-even point between hourly billing and a monthly server, the costs the hourly rate leaves out, and when hourly still wins.
- The 2026 DRAM squeeze: what it means for dedicated server buyersServer DRAM contract prices rose at record quarterly rates from late 2025 into 2026. The data, why it happened, what it does to dedicated-server pricing and how to rent only the memory you need.
Deploy RTX PRO 6000 for $899 a month.
Built to order — ready within 72 hours. 10 Gbps unmetered, no setup fee, cancel anytime.
Ready in 72 hCancel anytimeBTC, ETH, XMR, USDT or SOL