Database Mart GPU servers priced per GB of VRAM: which card runs which model, and what it costs a month

We pay for the servers and plans we use. Links marked as affiliate links earn us a commission at no cost to you; it never changes what we recommend. Details.

If you want to run a language model on your own server, one number decides almost everything: how much memory the GPU has. The model has to fit in it. Database Mart lists 37 GPU plans, from a $29 GPU VPS to a $2,599 H100 server, so we put every single-GPU plan on the same scale: dollars per gigabyte of GPU memory, next to the model sizes Ollama publishes.

The one rule: the model has to fit in VRAM

Ollama lists a download size for every model. That file is loaded into GPU memory, and the conversation (the context) needs more on top. So the download size is the floor, not the target. Leave room.

The sizes are the default (4-bit quantised) versions Ollama downloads when you give no tag. Larger quantisations are bigger.

GPU VPS plans

A GPU VPS is a virtual server with a whole GPU card attached; Database Mart's FAQ says GPU resources are not shared and each server or VPS comes with a dedicated card.

The RTX A4000 VPS also lists an hourly price of $0.25. The order page is where that option is chosen; we have not tried it.

Dedicated GPU servers, single card

A dedicated server is a whole physical machine. These come with far more system memory and disk than the VPS plans, but older CPUs and a 100 Mbps port on most plans.

Older cards in the list (K80, GTX 1650, GTX 1660, RTX 2060) are left out of this table: they are slower generations, and the K80 is two 12 GB GPUs on one board rather than one 24 GB card. Multi-GPU servers (2 to 4 cards, $599 to $2,499 a month) are on the same page; we have left them out because a model split across cards is a different setup from one card.

What stands out

• The VPS plans are the better deal per GB for most people. The RTX A4000 VPS gives 16 GB for $149; the dedicated server with the same card is $279.
• The RTX 4090 and the RTX A6000 dedicated servers cost the same, $549, and the A6000 has twice the memory. If the model you want needs more than 24 GB, only the A6000 can load it on one card.
• The H100 is the most expensive per GB here. You pay for speed, not size: the A100 80 GB has the same memory for $900 less a month.
• Several plans cost the same on every term, including the RTX Pro 2000 VPS and the P1000 server. On those, do not prepay.

The conditions that come with the price

The same rules apply as on Database Mart's ordinary VPS plans, which we covered in our reading of their pricing and terms:

• Payments are generally non-refundable once a service is paid, under their terms of service. On a $549 server, that is worth knowing before the first invoice.
• The 24-month price is paid upfront for the whole term.
• The servers are in the United States, in Dallas and Kansas City. From our server in Frankfurt we measured about 125 ms to both. For a model you call through an API that is rarely the bottleneck; for an interactive tool used from far away, it adds up.
• There is a 24-hour free trial for new clients, and Database Mart says to ask its sales team if a project needs longer. A day is short for a GPU: model downloads alone take a while. Here is how to request it and what to check.

Which one to pick

1. Trying a small model (7B to 8B) for an internal tool: Basic GPU VPS, RTX 5060, $99 a month.
2. A useful assistant on 14B to 27B models: Professional GPU VPS with the RTX A4000, $149, or the Advanced GPU VPS with the RTX Pro 4000 for 24 GB at $249.
3. 70B-class models on one card: the RTX A6000 dedicated server, $549, 48 GB.
4. The biggest open models, and you will commit for 3 months: the RTX Pro 6000 GPU VPS, 96 GB, from $749 a month.

All of these are on Database Mart's GPU server page. Use the trial before paying, and size by memory first.

What we have not tested

Speed. We have not run a model on any of these servers, so nothing here says how many tokens per second a card produces. That is the next test: Ollama on a Database Mart GPU VPS, timed, with the numbers published here.

FAQ

How much does a GPU server cost at Database Mart?

On 2 Oct 2026 Database Mart's GPU plans ran from $29 a month for a GPU VPS with a 2 GB card to $2,599 a month for a dedicated server with an NVIDIA H100, on a one-month term. Most plans are cheaper per month on 12- or 24-month terms paid upfront.

Which Database Mart GPU can run Llama 3.1 70B?

Ollama lists llama3.1:70b as a 43 GB download, so it needs a card with 48 GB or more. The cheapest single-card plans with 48 GB were the RTX A6000 and A40 dedicated servers at $549 a month, and the RTX Pro 5000 GPU VPS at $419 a month on a 3-month term.

Are Database Mart GPUs shared?

Database Mart's GPU FAQ says its GPU resources are not shared: each server and VPS comes with a dedicated GPU card.

What is the cheapest GPU per GB of VRAM at Database Mart?

On a one-month term, the RTX A4000 GPU VPS at $9.31 per GB of VRAM. Counting plans sold only on 3-month or longer terms, the RTX Pro 6000 GPU VPS at $7.80 per GB.

Can I try a Database Mart GPU server before paying?

Yes. New clients can request a 24-hour free trial with no payment. For longer testing, Database Mart says to contact its sales team.