Rent a whole NVIDIA T4, A100 or H100 for AI, machine learning and rendering. Full root access, NVMe storage and a GPU that is 100% yours — no time-slicing, no card, no KYC. Just an email and a wallet.
ApexVPS GPU servers give you a whole physical NVIDIA accelerator — a T4, A100 or H100 — attached to a truly dedicated VPS. There is no shared GPU pool and no time-slicing: the VRAM and compute you rent are exclusively yours for the full term. That makes them a good fit for AI and machine-learning teams, researchers, studios and solo builders who need reliable acceleration without a long procurement process.
Like the rest of our range, GPU plans are priced in US dollars and settled through a crypto-only checkout, so you can pay from your own wallet in Bitcoin, Ethereum, USDT or 30+ other coins. You still get full root access, NVMe storage, IPv4 and IPv6, DDoS protection and 24/7 monitoring. If you only need CPU for now, compare the standard tiers on the VPS pricing page or read how our crypto VPS checkout works before you deploy.
Each plan is one dedicated NVIDIA card plus dedicated vCPU, RAM and NVMe. Pay monthly or save about 15% yearly.
1× NVIDIA T4 • 16 GB VRAM
1× NVIDIA A100 • 80 GB VRAM
1× NVIDIA H100 • 80 GB VRAM
All GPU plans include full root access, KVM virtualization, IPv4 + IPv6, DDoS protection, 24/7 monitoring, backups and a 30-day money-back guarantee refunded in USDT. Prefer to pay by the hour? See on-demand GPUs. Need CPU-only instead? See standard VPS pricing.
Need a GPU for a few hours, not a whole month? Rent any card by the hour from a prepaid balance — spin one up for a training run or a batch of renders, then shut it down. You only pay for the time it runs.
Rates are per GPU, per hour, drawn from your prepaid balance. More cards (A100 SXM, H100 NVL, B300) available on request. On-demand pricing is indicative and reviewed regularly.
Add a minimum of $20 in BTC, ETH, USDT or 30+ coins. No card, no KYC.
From an RTX A5000 to an H200 — whatever your job needs, deployed in minutes.
Usage draws from your balance. Stop anytime; unused balance stays yours.
New here? Create an account or log in — then add a minimum $20 balance. Prefer a dedicated card for the full month? See the monthly GPU plans.
A dedicated NVIDIA card unlocks workloads a CPU-only VPS cannot handle. These are the jobs people deploy most.
Serve open-weight language models or fine-tune them on your own data. Larger VRAM tiers like the A100 and H100 give the headroom bigger models and longer contexts need.
Run diffusion pipelines for text-to-image, upscaling and batch generation, with the GPU handling the heavy sampling steps so your throughput stays consistent.
Train and iterate on neural networks in PyTorch, TensorFlow or JAX. Dedicated compute means no queue and no noisy neighbours competing for the card.
Offload GPU-accelerated rendering from engines and renderers, keeping your local workstation free while long render jobs run in a data center.
Use hardware encoders to transcode and process video at scale for streaming, archiving or media pipelines, all on storage-backed NVMe.
Accelerate notebooks, dataframes and large-scale numerical work with CUDA-aware libraries, turning batch jobs that took hours into minutes.
Because you have full root, you can install any CUDA-compatible framework and pin the exact driver, CUDA or container runtime your project needs. Many teams pair a GPU box with a CPU VPS for self-hosting supporting apps and dashboards, and explore the full lineup on our all services overview.
The same principles as the rest of our platform — dedicated hardware, real control and crypto-native billing — applied to accelerated compute.
You get a whole physical NVIDIA card. The VRAM and compute are yours for the full term, with no sharing or throttling from other tenants.
Complete root access on KVM virtualization. Pick your OS and pin the exact NVIDIA driver, CUDA or container runtime your framework expects.
NVMe SSD keeps datasets, checkpoints and model weights close to the GPU, so data loading rarely becomes the bottleneck in your pipeline.
Pay in BTC, ETH, USDT or 30+ coins through OxaPay. No credit card, no bank account and no KYC — just an email for your access details.
Deploy close to your users or data in one of our 39 data centers, from Frankfurt and London to Singapore, Tokyo, New York and São Paulo.
Every GPU plan is backed by a 30-day money-back guarantee. Refunds are issued in USDT to a wallet you provide, whatever coin you paid with.
Compare the three NVIDIA tiers side by side to match the card to your workload.
| Specification | GPU T4 | GPU A100 | GPU H100 |
|---|---|---|---|
| NVIDIA GPU | 1× T4 | 1× A100 | 1× H100 |
| GPU memory (VRAM) | 16 GB | 80 GB | 80 GB |
| vCPU (dedicated) | 8 | 16 | 24 |
| System RAM | 32 GB | 120 GB | 200 GB |
| NVMe storage | 200 GB | 500 GB | 1 TB |
| Bandwidth | 4 TB | 8 TB | 12 TB |
| Best suited for | Light inference, dev & testing | LLM inference & fine-tuning | Training & the largest models |
| Monthly price | $249 | $1290 | $1990 |
All GPU plans include full root access, IPv4 + IPv6, DDoS protection, 24/7 monitoring and backups. GPU hardware specifications reflect NVIDIA's published data-center GPU lineup.
The GPU is fully dedicated to your server — there is no time-slicing or sharing with other tenants. Each plan gives you a whole physical NVIDIA card (T4, A100 or H100) plus dedicated vCPU cores, RAM and NVMe storage, so the VRAM and compute you see are yours alone.
For lighter inference and smaller models, the GPU T4 with 16 GB of VRAM is a cost-effective starting point. For serving mid-to-large LLMs and fine-tuning, the GPU A100 with 80 GB of VRAM is our most popular inference tier. For the heaviest training and the largest context windows, the GPU H100 offers the most headroom. Frameworks and model weights from communities like Hugging Face run on any of them.
Yes. Every GPU plan is purchased through a crypto-only checkout powered by OxaPay, accepting Bitcoin (BTC), Ethereum (ETH), USDT (Tether) and 30+ other coins. There is no credit card, no bank account and no KYC — signup is email-only so we can send your access details. See the full flow on our crypto VPS page.
Yes — you choose what fits the workload. Dedicated T4, A100 and H100 plans are billed monthly or yearly (yearly saves about 15%) and reserve a whole card for the full term — best for always-on servers. On-demand GPUs are billed by the hour from a prepaid balance: top up a minimum of $20 and pay only for the hours you run, ideal for bursty jobs like a training run or a batch of renders.
You get a clean server with full root access and your choice of Ubuntu, Debian, CentOS, Rocky, Windows Server or a custom ISO. GPU images ship with recent NVIDIA drivers and the CUDA toolkit ready to go, and because you have root you can install a specific driver, CUDA or cuDNN version, or a container runtime, to match your framework.
Common workloads include LLM inference and fine-tuning, Stable Diffusion and other image generation, model training, 3D rendering, video encoding and transcoding, and general data-science work. You have full root access, so any CUDA-compatible framework such as PyTorch or TensorFlow is supported.
Pick a T4, A100 or H100, pay from your wallet, and get root on a dedicated NVIDIA server. 30-day money-back guarantee, refunded in USDT.