AiXVPScost & terms for AI / automation

Contabo Decision guide Updated 2026-11-12

Contabo GPU VPS: When It Beats a CPU VPS for a 7B or 13B Model

Contabo's GPU VPS is one large class, not a menu of small GPUs: a dedicated NVIDIA RTX PRO 6000 with 96 GB of VRAM, 18 vCPU, 96 GB of RAM, and 900 GB of NVMe, listed at €999/month list, excluding VAT, or €799.20/month on a 24-month term (20% off), excluding VAT. For a single 7B or 13B model, that is usually more GPU than the model needs. The interesting question is not "is a GPU faster?" It is "does the workload earn back the monthly difference?" For occasional 7B or 13B inference, I would start with a high-RAM Core VPS and measure. For sustained, low-latency, concurrent serving, or models that do not fit a CPU box, the GPU VPS becomes the relevant Contabo option, but at that price I would compare it with other GPU providers before committing.

Contabo review Sources
Contabo decision guide — one purchase question, one decision.
Scope. Decision draft. Contabo product and price facts were checked on the official Contabo GPU page on 2026-11-12. This is a planning analysis, not a GPU benchmark, a tokens-per-second test, a paid order, or a provider ranking. Model memory and speed depend on quantization, context, runtime, and batching; measure your own workload.

The model math

A rough memory estimate is the first filter. Model weights are approximately the parameter count times the bytes per parameter:

  • A 7B model at 4-bit quantization is roughly 3.5 GB of weights before overhead.
  • A 13B model at 4-bit is roughly 6.5 GB before overhead.
  • An 8-bit or 16-bit load multiplies those figures.
  • Context length, KV cache, runtime overhead, and concurrent requests add more.

The exact number depends on the model format and runtime. These are weight estimates, not a full deployment budget: add KV cache, runtime overhead, and concurrent requests before comparing with a plan's RAM or VRAM. The point is that a 7B or 13B model can fit in a modest amount of memory; 96 GB of VRAM is not required for one small model. It is useful for larger models, long context, multiple concurrent requests, image/video models, or fine-tuning.

A CPU VPS runs the same model in system RAM. It is slower per token, but it can be adequate for a nightly batch, a private prototype, or a small number of users who can wait. The local AI guide covers the RAM and disk side.

What Contabo's GPU VPS includes

The current GPU page lists one configuration:

  • Dedicated NVIDIA RTX PRO 6000, 96 GB VRAM, CUDA.
  • 18 vCPU, 96 GB RAM, 900 GB NVMe.
  • Unlimited traffic, with the port and fair-use terms to verify.
  • €999/month list, excluding VAT; €799.20/month on a 24-month term (20% off), excluding VAT.

That is a data-center GPU class, not a consumer card. The price reflects it. A single small model will not use most of the VRAM, and the 18 vCPU/96 GB RAM are there to feed the GPU, not because the model needs them.

CPU vs GPU for a small model

QuestionCPU VPSContabo GPU VPS
Fits a 7B/13B quantized model?Usually yes, with enough RAMYes, with far more VRAM than one model needs
Speed per tokenSlower; depends on CPU and runtimeFaster; depends on GPU, batching, runtime
ConcurrencyLimited by CPU/RAMMuch higher headroom
Monthly costLower Core/Performance plan€999 list, excluding VAT; €799.20 on a 24-month term, excluding VAT
Best useOccasional batch, prototype, small private useSustained serving, larger models, media/GPU work

The CPU column is not a consolation prize. If the job is a nightly document summary or a private chat for one or two people, a slower answer may be acceptable. The GPU column is not automatically the right buy; it is a large fixed cost that needs utilization.

When the GPU VPS wins

I would consider the Contabo GPU VPS when:

  • The model is larger than a 13B, or the context is long enough that KV cache dominates.
  • Several users or workflows need concurrent answers, and queueing on a CPU box misses the deadline.
  • The workload includes image, video, or other GPU-native processing.
  • Fine-tuning or a GPU-only runtime is part of the job.
  • The monthly cost is justified by saved time or revenue, not by curiosity.

The last point matters. At €999/month list, excluding VAT, or €799.20 on a 24-month term, the GPU VPS needs a workload that would otherwise cost more in delay, failed deadlines, or additional servers. The 20% term discount does not change the order of magnitude. A prototype that runs twice a day does not justify it.

When the CPU VPS wins

I would keep a CPU VPS when:

  • The model is a 7B or 13B quantized model used occasionally.
  • A slower response is acceptable, or the work runs overnight.
  • The budget is better spent on RAM, disk, and backups.
  • The workload is API orchestration, not model inference.
  • The team has not measured tokens per second or concurrency yet.

A high-RAM Core plan may be enough to load the model and serve a small number of requests. It will not match a GPU on latency, but it may meet the actual requirement.

What I would measure before buying

I would run the real model, prompt length, and concurrency on a CPU box and record:

  • Time to first token and tokens per second at the expected concurrency.
  • Peak RAM and VRAM (if any), plus model load time.
  • Queue wait and failure rate at the deadline.
  • The monthly cost of the CPU box versus the GPU VPS.

If the CPU box meets the deadline with margin, the GPU is a want, not a need. If it misses consistently, the GPU or a different provider is the next test. I would also compare the Contabo GPU VPS with other GPU providers at the same VRAM and term; one large class is not a competitive range.

What this does not decide

This article does not benchmark the RTX PRO 6000, does not report tokens per second, and does not test a 7B or 13B model on Contabo. It does not claim that a CPU VPS is fast enough for production. It gives the memory filter and the utilization test.

Sources