Does your app run a model, or call one?
Ten seconds to answer, and it decides everything. If it calls a hosted API, this is the right machine and a GPU would be money set on fire — the work is orchestration and waiting, not matrix maths. If it runs a model locally, we do not sell GPU instances and you should buy elsewhere. We would rather say that on the first screen than take the order.
- Agents and API backends
- Vector databases and RAG
- Stable IP for webhooks
- No GPU instances
$4.95/mo
CPU-based — which is the correct machine for most AI applications
- Python, Node, Docker — anything
- NVMe available for vector indexes
- Dedicated IPv4 for callbacks
- No GPU — go elsewhere for local inference
Why Most AI Applications Do Not Want a GPU
Training needs GPUs. Production usually does not. The AI workloads that actually run day to day — an agent calling a hosted model, a RAG pipeline querying a vector store, a webhook that classifies incoming tickets, a nightly embedding job — spend the overwhelming majority of their wall-clock time waiting on someone else's API. They are I/O-bound. What they need locally is enough CPU to handle concurrent requests, enough memory to hold a vector index, storage fast enough for small random reads, and a stable IP address for callbacks. All four are things a CPU VPS does well and cheaply. What we do not sell is GPU capacity — so if your weights live on disk and inference happens locally, this is the wrong supplier and the page says so up front.
- Fits: agents, RAG retrieval, pipelines, webhook workers.
- Does not fit: training, fine-tuning, local model serving.
- Size by the vector index — memory and storage, not cores.
- NVMe available where retrieval latency matters.
- Dedicated IPv4 on every plan for callbacks and allowlists.
- No GPU instances — stated plainly, not buried in an FAQ.
Related: VPS for Docker · NVMe VPS · VPS for developers · Dedicated IPv4 · High performance VPS
What Fits Here, and What Does Not
Four workloads this suits, and one it does not — listed together rather than the second one being left out.
- Agents and API-calling backends An agent that reasons by calling a hosted model API spends almost all of its wall-clock time waiting on someone else's network. The local machine is orchestrating, parsing and holding state — work a single core does comfortably. This is by far the most common production AI workload and it does not want a GPU. Fits
- Vector databases and RAG retrieval Qdrant, pgvector, Chroma and similar are memory and disk workloads. Retrieval is index lookups and small random reads, which is why storage type and RAM matter far more here than core count. This is the one AI workload where the plan choice genuinely changes the experience. Why storage type matters → Fits
- Pipelines, workers and scheduled jobs Embedding batches, document ingestion, classification queues, nightly summarisation. Bursty, concurrent, and bounded by how fast the upstream API answers rather than by local compute. Fits
- Webhook endpoints that must stay reachable A dedicated IPv4 on every plan gives you a stable address for callbacks, allowlists and outbound integrations — the boring requirement that quietly rules out running this from a laptop or a dynamic connection. Dedicated IPv4 → Fits
- Anything that runs a model locally Training, fine-tuning and self-hosted inference of a large model need a GPU, and we do not sell GPU instances. A large model on CPU is not slow-but-workable; it is unusable for anything interactive. If that is your project, buy from a GPU provider — we would rather say so than take the order. Does not fit
Size It By the Index, Not the Traffic
The plan that matters is the one that holds your vector index in memory with fast storage under it. Each row shows the cheapest plan in the catalogue that meets the requirement.
| If this is you | Cheapest plan that fits | What it gives you | Price |
|---|---|---|---|
| An agent or bot calling a hosted model API Almost pure I/O wait. This needs far less machine than people expect. | Business Classic VPS Classic VPS range | 1 Core · 2 GB RAM 20 GB SAS | $7.95/mo + $2.00 one-time setup |
| RAG with a modest vector index Retrieval is small random reads. Storage type matters more than cores here. | Professional NVMe VPS NVMe VPS range | 2 Cores · 4 GB RAM 40 GB NVMe | $17.95/mo + $0.00 one-time setup |
| A larger index, or concurrent ingestion Embedding batches and queries at the same time is where memory pressure appears. | Enterprise NVMe VPS NVMe VPS range | 3 Cores · 8 GB RAM 50 GB NVMe | $23.95/mo + $0.00 one-time setup |
| Production workload with real concurrency Multiple workers, a database, a queue and headroom to not be the bottleneck. | Enterprise Cloud VPS Cloud VPS range | 4 Cores · 8 GB RAM 60 GB NVMe (triple-replicated) | $28.95/mo + $0.00 one-time setup |
Three Questions Before You Order
The first one decides whether you should be buying from us at all.
Does your app run a model, or call one?
This is the whole question, and it takes ten seconds to answer. Calls one — an API key in your environment, a request going out to a hosted model — then a CPU VPS is the correct machine and a GPU would be money set on fire. Runs one — weights on disk, inference happening locally — then you need a GPU host and this is the wrong page.
How big is the vector index?
The resource that actually decides your plan. A few thousand documents is unremarkable. Hundreds of thousands of embeddings is a memory and disk question, and the answer is more RAM and faster storage rather than more cores. Size the index first, then pick the plan.
How much concurrency, really?
Most agent backends serve far fewer simultaneous requests than their authors expect, because each request spends most of its life waiting on an upstream API. Measure before buying cores — concurrency in this workload is usually a queueing problem, not a CPU one.
AI Hosting Questions
The right machine for the AI you are actually running.
CPU-based plans from $4.95/month with NVMe storage and a dedicated IPv4 for callbacks — and an honest no if what you need is a GPU.