Disclosure: We earn commissions from partner links. This doesn't affect our rankings. Learn more

Reviewed by VPSchart Editorial Team · Updated 3 September 2026

Ollama

Ollama VPS Hosting: CPU, RAM and GPU Choices

Choose the server around the model you will actually run. A 16 GB CPU VPS can be a sensible low-cost test box, while interactive use, long context and concurrent requests usually make a GPU instance the better fit. This guide separates those two buying decisions and shows the price conditions behind each option.

Quick decision

Personal testing with a small quantized model: start with a 16 GB CPU VPS and measure response time before paying for a GPU.
Interactive chat or repeated API calls: rent a supported GPU whose VRAM can hold the model and context allocation.
Agents, coding or long documents: budget extra memory for context. Ollama recommends at least 64K context for these workloads.
Variable or temporary GPU work: a marketplace instance can reduce idle cost, but compare reliability, storage and bandwidth before renting.

How to size an Ollama server

Parameter count alone is not a safe sizing rule. Quantization changes the model file size, and context length adds a separate memory cost. Parallel requests increase that context allocation again. Use this sequence before choosing a plan:

Open the Ollama model library and note the size of the exact model variant you intend to pull.
Choose a context target. Ollama currently defaults to 4K below 24 GiB VRAM, 32K from 24 to 48 GiB and 256K at 48 GiB or more.
Leave headroom for the operating system, model cache and any reverse proxy or web interface running on the same server.
After launch, run ollama ps. Its PROCESSOR column shows whether the model is in CPU memory, GPU memory or split across both.

See Ollama's current context-length guidance, memory and concurrency FAQ and supported GPU list.

Ollama hosting options checked 3 September 2026

Option	Compute	Memory	Storage	Listed price	Best fit	Action
Hetzner CX43	8 shared vCPU	16 GB RAM	160 GB	$18.49/mo	Low-volume CPU inferenceNo GPU. Germany and Finland price excluding VAT and IPv4.	View option →
Hostinger KVM 4	4 vCPU	16 GB RAM	200 GB NVMe	$12.99/mo promo	CPU setup with a control panelTwo-year promotional term. Listed renewal is $28.99 per month.	View option →
DigitalOcean Memory-Optimized	2 dedicated vCPU	16 GiB RAM	50 GiB SSD	$84/mo	Predictable CPU capacityNo GPU. Per-second billing with a monthly cap.	View option →
Vast.ai marketplace	GPU varies by offer	VRAM and RAM vary	Priced separately	Live market rate	GPU inference and temporary jobsCompute, storage and bandwidth prices vary by host and offer.	View option →

USD reference prices shown by each provider. Hetzner's value is the Germany and Finland price effective 15 June 2026, excluding VAT and IPv4. Hostinger's value is a promotional two-year term. Vast.ai has no static price because offers change with its marketplace.

CPU VPS or GPU server?

Choose a CPU VPS when

you are validating a small model or integration;
requests are occasional and slower generation is acceptable;
predictable monthly cost matters more than response speed;
you are hosting Open WebUI or an API gateway while inference runs elsewhere.

Choose a GPU server when

users expect interactive responses;
the model and context fit a supported GPU's VRAM;
you need long context or parallel requests;
the server will process repeated inference jobs rather than an occasional prompt.

Which provider fits which use case?

Hetzner: the CX43 is the lowest fixed monthly price in this set for 16 GB RAM. It has no GPU and uses shared vCPU, so treat it as a low-volume CPU starting point rather than a performance promise.

Hostinger: KVM 4 provides 16 GB RAM and a larger disk with a guided VPS interface. The advertised price requires a two-year promotional term, and the listed renewal is materially higher.

DigitalOcean: the 16 GiB memory-optimized Droplet uses dedicated vCPU and transparent billing. It costs more than the budget VPS options and still has no GPU, but it is easier to budget for steady CPU work.

Vast.ai: use the marketplace when GPU speed matters and you can evaluate individual offers. Filter for a GPU supported by Ollama, enough VRAM, host reliability and a rental duration that matches the job. The visible GPU rate is not the whole bill.

Secure deployment outline

Request path

Client

→

TLS and authentication

→

Ollama on 127.0.0.1:11434

→

Model in RAM or VRAM

1. Install on Linux

Follow the official Linux installation and run Ollama as a systemd service. Install the NVIDIA container toolkit only if you use the official container with NVIDIA GPU access.

2. Verify memory placement

Pull the intended model, send a test request and inspect ollama ps. If a GPU workload is partly offloaded to CPU, reduce model or context size or choose more VRAM.

3. Protect remote access

Keep port 11434 private. Put a reverse proxy with TLS and authentication in front of the service, limit inbound firewall rules and update both Ollama and the host operating system.

Frequently Asked Questions

How much RAM does an Ollama VPS need?

There is no single RAM requirement for every Ollama model. Start with the downloaded model size, then leave memory for the operating system and context cache. A larger context and parallel requests increase memory use. Verify the actual CPU and GPU split with ollama ps.

Can Ollama run on a CPU-only VPS?

Yes. Ollama can load a model into system memory and run on CPU. This can suit testing and low-volume use, but generation is normally slower than on a supported GPU. Do not choose a CPU VPS from RAM alone; CPU contention also matters.

When should I rent a GPU server for Ollama?

Choose a GPU when response speed, larger models, long context or concurrent users matter. Check that the model and context allocation fit the available VRAM. Ollama documents the active memory split in the PROCESSOR column of ollama ps.

Is Vast.ai pricing fixed?

No. Vast.ai is a marketplace where hosts set prices. The total can include GPU compute, storage and bandwidth, and storage may continue while a stopped instance exists. Check the live offer before renting.

Should port 11434 be public?

Ollama binds to 127.0.0.1 by default. Keep the service private unless remote access is required. For remote use, place authentication and TLS in front of it, restrict the firewall and avoid exposing an unauthenticated API to the internet.