Choose a GPU plan
Compare RTX 2060 SUPER, L4, RTX 4090 and A100 instances, size VRAM for the model you want to run, and understand provisioning.
GPU & LLM hosting gives you a whole card, not a slice. No MIG partitioning, no shared queue, no per-token billing.
Plans
| Plan | GPU | vCPU | RAM | Storage | Best for |
|---|---|---|---|---|---|
| RTX 2060 SUPER | 1× RTX 2060 SUPER, 8 GB | 10 | 18 GB | 500 GB SSD | Small models, embeddings, dev and staging endpoints |
| L4 Inference | 1× NVIDIA L4, 24 GB | 8 | 32 GB | 100 GB NVMe | Always-on inference, low power draw |
| RTX 4090 | 1× RTX 4090, 24 GB | 8 | 32 GB | 200 GB NVMe | Embeddings, RAG, fine-tuning, fast inference |
| A100 80GB | 1× A100, 80 GB HBM2e | 12 | 96 GB | 500 GB NVMe | Training, large models, long context |
Annual variants are the same monthly rate billed once for twelve months. Provisioning takes about 48 hours because the machine is prepared for you.
Sizing by VRAM
VRAM decides which models fit. Rough guidance for quantised inference:
| VRAM | Comfortable |
|---|---|
| 8 GB | 1B–3B quantised, embedding and reranking models, dev and staging endpoints |
| 24 GB | 7B–8B models at 4-bit or 8-bit, embeddings, small vision models |
| 48 GB+ | 13B–34B quantised, larger context windows |
| 80 GB | 70B quantised, bigger batches, training and fine-tuning |
These are starting points, not limits. Context length, batch size and the inference engine all change memory use. A model that fits may still be slow if the KV cache does not.
Which card
- RTX 2060 SUPER — the cheapest way onto a dedicated card. 8 GB of VRAM covers small models, embedding and reranking pipelines, and any workload where you need a real GPU but not 24 GB of it — staging replicas of a production endpoint, for example.
- L4 — efficient for steady API traffic. Lower power, lower heat, predictable. Good default for a production endpoint serving one model.
- RTX 4090 — the best price/performance for single-GPU work: embeddings, RAG pipelines, LoRA fine-tunes and interactive inference.
- A100 80GB — choose it when the model does not fit in 24 GB, when you need large batches, or when you are training rather than serving.
If you are unsure, start with the RTX 2060 SUPER or the L4 for inference and move up only when you hit memory or throughput limits. You can order a second machine rather than rebuild.
Software
- The instances ship with CUDA 12 support.
- You have root, so you can install any framework.
- Nothing is pre-installed beyond the driver stack. Install Ollama or vLLM yourself; the Run Ollama and Run vLLM guides take about fifteen minutes.
- The card is private to your instance: weights, prompts and logs stay on the machine.
Ordering
- Open zoophost.com/gpu.
- Pick a plan and term.
- Check out.
The machine is handed over by our team with root access and a plain CUDA image. It appears under Servers in the console with the usual connection details. Power, rescue and reinstall actions work exactly as they do for a VPS.
What is not included
- No bundled backup service. Keep model weights and datasets somewhere durable — Z2 Object Storage is a good fit.
- No managed ML platform. You install and run the serving stack.
- No per-hour billing. Plans are flat monthly; the card is yours for the term.
Next
- gpu
- rtx 2060 super
- l4
- rtx 4090
- a100
- vram
- inference
- training
Ready to use this
A whole card with root access, not a shared slice or a per-token meter.
- 8 GB RTX 2060 SUPER from the entry tier
- CUDA 12 ready on every plan

