A CPU VPS and a GPU VPS can both “run AI,” and that phrase hides the real decision. Some workloads are comfortable on ordinary cores and RAM. Others spend most of their time in matrix math that a GPU finishes in a fraction of the time, or that a CPU will not finish at a usable speed at all.
This is a workload comparison, not a vendor list. It also is not a second Ollama install guide; the Ollama hosting article already covers that runtime. Use this page to decide which kind of machine to rent before you pick a model or a framework.
What each plan actually gives you
A CPU VPS is the familiar product: virtual CPUs, system RAM, disk and a network port. It is the right default for web apps, automation, APIs that call a hosted model, and many small embedding or classification jobs.
A GPU VPS adds a graphics card (or a slice of one) inside the virtual machine, with its own VRAM. The guest needs a working NVIDIA (or other vendor) driver and, for containers, the matching toolkit. If nvidia-smi does not see a card, you do not have a GPU VPS in any useful sense, whatever the plan name says.
The two products also differ in cost structure and in how you operate them. GPU plans are scarcer, often billed at a higher rate, and more sensitive to drivers, CUDA userspace and passthrough quality. CPU plans are easier to move, snapshot and treat like a normal server. The general upgrade path from shared hosting in Shared Hosting vs VPS still applies; a GPU is an extra specialization on top, not a replacement for basic VPS hygiene such as the steps in setting up a VPS and securing it.
When a CPU VPS is enough
| Workload | Usually fine on CPU? | Why |
|---|---|---|
| An app that calls a hosted LLM API | Yes | The heavy inference happens elsewhere. You need RAM, disk, a stable network and a locked-down process, not a GPU. |
| Orchestration and agents that use an external model | Often | The agent loop, tools, queues and databases are ordinary server work. See VPS requirements for AI agents. |
| Very small local models, light embeddings, or batch jobs you can wait for | Sometimes | Quantized 1B–8B class models can run on CPU. Tokens per second will be lower; that may be acceptable for a personal tool or an overnight batch. |
| Interactive chat with a mid-size or large local model | Usually no | Latency becomes the problem. Users will not wait many seconds per token. |
| High-concurrency local inference, image generation, or speech models at useful speed | No | These are the jobs GPU VRAM and GPU memory bandwidth are for. |
If your architecture is “the VPS holds the app, the model lives at an API,” start with a CPU plan and size it like any other application host. The self-host vs API comparison is the decision behind that architecture. Do not rent a GPU “because the project is AI” if the model never runs on your machine.
When you need a GPU VPS
You need a GPU when the model (or the diffusion/speech graph) must run on your server and you care about interactive latency or throughput. Typical cases:
- Self-hosting an open-weight LLM for a team chat, an internal API, or a product that cannot send prompts to a third party.
- Serving many requests at once from a local model. This is where engines such as those discussed in vLLM vs Ollama start to matter, and where VRAM and batching dominate CPU core count.
- Image or video generation graphs, or local speech-to-text / speech synthesis at a speed people will actually use.
- Fine-tuning or LoRA work. Even small training jobs are usually impractical on a CPU VPS.
VRAM is the first constraint, not vCPU count. Size the card with the method in how much VRAM you need for an LLM before you compare clock speeds. A 7B or 8B quantized model may fit in 8 GB. A 70B 4-bit model usually does not. A CPU with 64 GB of RAM can sometimes load a model that a small GPU cannot, and still be the slower choice for anyone waiting on a reply.
How to compare them in practice
Use the same prompts, the same context length and the same concurrency on both machines, then record tokens per second, time to first token, error rate and memory use. A single “it loaded” test is not a comparison. The hardware checks in the VPS benchmark guide still help for disk and network; add nvidia-smi and the runtime’s own metrics on the GPU side.
Also compare the operational cost of the GPU path: driver updates, CUDA compatibility with your containers, and what happens if the provider oversubscribes the card. CPU steal time, covered in the CPU steal time guide, has a cousin on busy GPU hosts when the card is shared. Ask whether the GPU is dedicated or time-sliced, and confirm you can see the full VRAM you paid for.
Disk and bandwidth still matter. Model pulls are large; a narrow port or a small disk will stall a GPU VPS just as they stall a CPU VPS. See bandwidth vs port speed if the plan’s transfer cap is unclear.
A simple decision path
- If the model is called over the network and never loaded locally, start with a CPU VPS.
- If you need a local model only for occasional, non-interactive jobs, try a quantized small model on CPU first and measure.
- If people will sit on a chat or an API and wait for tokens from a local mid-size or large model, or you need images, speech, or many concurrent local requests, use a GPU VPS and size VRAM first.
- If you need both a public app and a local model, it is valid to run the app on a CPU VPS and the model on a separate GPU VPS, and to keep the model API private. That split is often easier to secure and to scale than one overloaded box.
Containers do not change the decision; they only package it. A Dockerized model on CPU is still a CPU workload. The general container notes in VPS hosting for Docker apply to both types of plan.
Frequently asked questions
Can I add a GPU to an existing CPU VPS later?
Only if the provider supports it on that host, which many do not. GPU plans often live on different nodes. Budget for a migration: snapshot, move data, and re-point DNS or the internal API, using the same care as in a normal VPS migration.
Will more vCPUs make a local LLM fast enough without a GPU?
More cores can help a bit, especially for batch or embedding work, but they do not replace GPU memory bandwidth for interactive decoding of larger models. Measure tokens per second. If the result is already too slow on a large CPU plan, adding cores is the wrong lever.
Is a shared GPU slice enough?
It can be, for light or bursty use, if you get the VRAM and the isolation you were promised. It is a poor fit when you need predictable latency or the full card. Read the allocation terms and verify with nvidia-smi under load.
Do I need a GPU to run n8n, Open WebUI or an agent?
Not for the control plane. n8n, Open WebUI and most agent processes are ordinary services. You need a GPU only if the model they call is local and large enough that CPU inference is too slow.
Does KVM vs another hypervisor change this?
You need clean GPU passthrough or a supported virtual GPU. The hypervisor comparison in KVM vs Hyper-V is background; the practical test is still whether the guest sees a stable device and the expected VRAM.