Renting the whole GPU server: what GPU Mart does differently

Between the per-second marketplaces and the hyperscalers sits a third category that gets far less attention — and it's where an A4000 at $0.25/hour and a dedicated H100 live in the same catalogue.

Most GPU cloud coverage splits into two camps. On one side, marketplaces like Vast.ai and RunPod — cheap, elastic, per-second billing, and a machine you never really own for more than a few hours. On the other, hyperscalers, where an A100 costs more per hour than a dedicated server costs per week.

GPU Mart sits in a third category: GPU hosting. You rent a server and you keep it. Dedicated GPU servers provide single-tenant bare-metal hardware, while GPU VPS plans use PCIe passthrough to give each instance direct access to an isolated physical GPU. That distinction explains everything else about how the pricing is built and who it suits.

The company is a brand of Database Mart LLC, a US provider that has run dedicated servers and VPS infrastructure since 2005. The GPU line launched in 2021 and now spans 25+ NVIDIA models across roughly 3,000 active GPUs, all in SOC-certified US data centres.

Six GPUs you can rent by the hour

Hourly billing is the part of the catalogue most likely to surprise people, because the entry pricing runs below most on-demand clouds and includes cards that are genuinely hard to rent hourly anywhere.

Two things stand out. The RTX A4000 at $0.25/hour is among the cheapest ways to get a CUDA device with ECC memory at any provider — cheap enough that leaving a CI runner or a small inference endpoint online stops being a budget conversation. And the RTX Pro Blackwell cards (Pro 4000 at $0.27/hour, Pro 5000 at $0.48/hour) are rare on the hourly market: most GPU clouds stock only data-centre parts and skip the professional workstation line entirely.

The A100 40GB at $1.11/hour deserves a second look from anyone who has been paying 80GB prices for a workload that fits in 40GB. If your model fits, you pay less than most providers charge for the cheapest A100 tier they offer.

The monthly line, and when it actually wins

Most higher-end models are sold monthly, while hourly billing is available for selected GPUs. Month-to-month price first, then the 24-month prepaid rate:

All figures in USD, verified against gpu-mart.com on 10 September 2026.

Here is the honest arithmetic, which matters more than the headline. A monthly server only beats hourly billing when you actually keep it running most of the time. GPU Mart's H100 works out to roughly $3.56/hour equivalent across a 730-hour month, or $2.88/hour if you prepay two years — both above what RunPod ($1.99/hour), DataCrunch ($2.39/hour) and Lambda Labs ($2.49/hour) charge on demand. If your H100 workload is bursty, rent it by the hour elsewhere.

Where the monthly model earns its keep is everything the hourly clouds meter separately. The enterprise dedicated tier ships with 256GB of RAM, a 36-core dual Xeon, and 240GB SSD + 2TB NVMe + 8TB SATA — included, not billed by the gigabyte-month. For a dataset-heavy pipeline, that storage alone can decide the bill. Add single-tenant hardware with no noisy-neighbour contention, and the comparison stops being purely about GPU-hours.

Four situations where this shape is the right one

Where it isn't the right fit

To be direct about the limits:

• US only. There are no EU, UK or APAC regions, so latency-sensitive or data-residency workloads outside the US should look at other providers in our directory.
• No per-second billing, no serverless, no spot market.
• The H100 and RTX Pro 6000 have no hourly option at all. They are monthly plans, full stop.
• Broadwell-generation host CPUs on the enterprise dedicated line — fine for feeding a GPU, but not a modern platform.

One more point worth reading carefully: GPU Mart advertises figures like "RTX Pro 6000 from $0.66/hour" and "GPU instances from $0.02/hour". Those are the 24-month prepaid monthly price divided by 720 hours — the effective cost of a long commitment, not a rate you can book. The genuinely hourly entry point is the RTX A4000 at $0.25/hour.

Support and SLA

Database Mart publishes a 99.9% network and platform uptime SLA covering its GPU server line, effective August 2026, with service credits worth three times the excess downtime, capped at the monthly recurring charge and claimable within eight business days. Support is 24/7 live chat with an advertised sub-five-minute average response.

Provisioning takes 20 to 40 minutes on Linux and roughly 30 minutes to 2 hours on Windows — slower than the seconds of on-demand clouds, and worth planning around if you spin resources up and down frequently.

The short version

If you need a GPU for two hours next Tuesday, use a marketplace. If you need a whole machine with real storage and memory, Windows support and a US address — running for weeks or months — GPU Mart is priced for that, and the entry hourly tiers give you a cheap way to test the hardware before committing to anything.

Read next: GPU Mart's full profile on GPUCloudList · Cheapest GPU Cloud Providers in 2026