Server with a dedicated NVIDIA GPU to train and run AI.

It is a server with a dedicated NVIDIA GPU, designed for Artificial Intelligence workloads, deep learning, model training and inference, computer vision and accelerated computing. The GPU is exclusively yours. It comes with CUDA, PyTorch and TensorFlow preinstalled.

The GPU Ada (20 GB VRAM) is ideal for inference, classic machine learning and light fine-tuning of mid-size models. The GPU Blackwell (96 GB VRAM) is for training large models, heavy fine-tuning and workloads that demand a lot of GPU memory.

The GPU Ada costs US$ 649.80/mo with a one-time setup fee of US$ 259.80 (it covers preparing the server, CUDA drivers and the AI environment). The GPU Blackwell is priced on request — as it is limited-stock, high-capacity hardware, the price is set in the quote. No contract.

Provisioning GPU servers takes up to 48 business hours. GPU servers have limited stock and dedicated preparation. The flow is: you request the plan, we confirm availability and delivery, and we provision it.

Upgrade: anytime — from GPU Ada to GPU Blackwell, paying only the pro-rated difference for the time left in the already-paid cycle; what you paid is not lost, it is credited. Because it involves GPU hardware with limited stock, the change is done in a window agreed with the team, preserving your data. Downgrade: scheduled for the next renewal — the current cycle's difference is not refunded in cash; any remaining balance becomes account credit you can use on any service. Reducing disk requires new provisioning and data migration, which we guide you through. The one-time setup fee is not refunded on downgrade. Details in the Refund Policy.

CUDA and cuDNN, PyTorch and TensorFlow. On request, we also configure JAX, distributed-training tools (DeepSpeed, Accelerate) and Jupyter environments. The server arrives ready to run your AI code.

Yes. Datasets, models and code stay on your server — nothing is sent to third parties. That is the difference from cloud AI APIs, where data leaves your infrastructure. Ideal for sensitive data.

You can — but if your focus is specifically hosting LLMs (Llama, Mistral, etc.) with Ollama and vLLM, the LLM Server page is more targeted. The GPU Server is more general: it works for training, deep learning, computer vision and any GPU-accelerated workload.

In most cases, yes. An RTX 4090 alone costs R$ 12-15k in hardware, with no server, cooling, power or redundancy. A monthly rental delivers the server ready, with SLA, support and replacement on failure. Buying only pays off with continuous 24/7 usage for more than 3-4 years.

The GPU Server is generic — it serves for neural network training, fine-tuning, computer vision, scientific computing, rendering. The LLM Server is a specialized variant: same hardware, but with Ollama, vLLM and llama.cpp preinstalled and optimized for serving LLMs (Llama 3, Mistral, DeepSeek).

100% dedicated. The VRAM, CUDA cores and Tensor Cores are exclusively yours. Unlike serverless or multi-tenant services, there is no "neighbor" competing for the GPU — predictable performance in training and inference.

Yes — Rollin Serviços Digitais e Tecnologia LTDA is a Brazilian company with an international Tier III datacenter with CDN in Brazil, NF-e, billing in BRL and human support 24/7. First Brazilian cloud specialized in AI, with dedicated products for GPU, LLM, vector DB and WhatsApp agents.

Yes — human support 24/7, with people who understand CUDA, NVIDIA drivers and tuning. Rollin Host is a Brazilian company (Rollin Serviços Digitais e Tecnologia LTDA).