Teams that run or fine-tune language models now have a familiar choice: call a hosted model and pay per token, or rent a graphics card and run the work themselves. The second path only makes sense if the card is the right size, the billing model matches the duty cycle, and the processing location is acceptable. This piece is about that second path, using published catalog figures from a European operator as a worked example.
Start with memory, not with the brand name
The weights of a model have to fit in graphics memory, with room left for the key-value cache of the requests that are in flight. A simple first estimate is enough to throw out the wrong card:
- About 2 GB of graphics memory per billion parameters for weights in FP16 or BF16.
- About 1 GB per billion parameters if the weights are stored at 8 bits.
- About 16 bytes per parameter for full training, which is why training and inference often need different machines.
An 8-billion-parameter instruction model in FP16 therefore wants around 16 GB for weights alone, which is why a 24 GB card is a common inference starting point. A 70-billion-parameter model at 8 bits wants around 70 GB for weights, which is already a multi-card problem once the cache is included.
On the GPU page of ENTRONYX CLOUD, visitors can enter a workload, a model size, and a numeric precision. The page names the smallest configuration whose memory can hold the model and calculates the hour, a run, and the monthly cap, using the same code as the configurator. That is the right order: size first, then price.
How the bill is actually calculated
The ENTRONYX CLOUD catalog lists six card types in ten configurations, in Frankfurt am Main, Düsseldorf, and Helsinki. The entry configuration, NG-RTX4000, is one NVIDIA RTX PRO 4000 Blackwell with 24 GB of graphics memory, billed at EUR 0.2932 net per hour. Usage is metered by the second at that hourly rate. No more than 730 hours are charged in a calendar month, so the monthly cap for that configuration is EUR 214.00 net. There is no minimum term. A 12-month or 24-month term reduces the monthly price by 5 percent or 8 percent.
Two consequences follow from that design.
First, a weekend experiment is not a month. Seventy-two hours on NG-RTX4000 is 72 × 0.2932, or about EUR 21 net, not the monthly cap. The cap only matters once a node is up for most of the month.
Second, the hourly rate is the monthly cap divided by 730. A team that plans to run the same card twenty-four hours a day should compare the cap, and the term discount, with buying or reserving capacity, not with a few hours of exploratory use.
The top of the catalog is an eight-GPU NVIDIA H200 node with 1,128 GB of graphics memory in total, at a monthly cap of EUR 30,660.00 net. Reserved capacity for longer training runs is handled through sales, because H200 nodes are not listed as freely available at all times. For clusters of four or more nodes and terms over 12 months, the company says it prices individually.
Helsinki is billed at a price factor of 0.94 relative to Frankfurt am Main; Düsseldorf uses the same list prices as Frankfurt (factor 1.00). If the workload has no reason to be in Germany, that 6 percent is a real difference, not a rounding error on a long training run.
What you do not get, on purpose
Some catalogs offer preemptible or spot instances at a lower price, with the right to reclaim the card. ENTRONYX CLOUD does not. A running card keeps running until the customer stops it. Each graphics card is assigned exclusively to one node with a full PCIe connection. There is no vGPU split between customers. CPU cores, system memory, and local NVMe drives are in the same chassis. Drivers, the CUDA toolchain, and a container runtime are preinstalled.
That package is easier to reason about than a shared slice, and it is more expensive than a slice. If the job can be checkpointed and restarted, a spot market elsewhere may still be cheaper. If the job cannot be interrupted, or if a shared card is unacceptable for isolation reasons, exclusive assignment is the feature, not a footnote.
Availability is not uniform. According to the company, RTX PRO 4000, L4, and L40S cards are listed as ready to start without a prior capacity check. RTX PRO 6000, H100, and H200 deployments depend on free capacity at the chosen location, with reservations available from 30 days. The product page shows availability by configuration and location before checkout. That matrix is worth reading before a deadline is promised to someone else.
Prices on the public list are net amounts in euros. VAT is added where it applies; the right rate for a US or UK company depends on the tax treatment of the order, not on a single 19 percent figure copied from a German invoice. The GPU configurator shows the net amount before the order.
When a card is the wrong unit
If the work is inference on a model with open weights, and the load is uneven, paying per million tokens can be cheaper than keeping a card warm. ENTRONYX CLOUD also offers that layer on the same hardware: six open-weight models billed per million tokens, with no base fee, and customer endpoints that scale to zero when they are idle. Failed requests are not charged. The AI platform page lists the models and the input and output prices.
A simple break-even sketch is enough. Estimate monthly input and output tokens from real logs, convert them at the published token prices, then compare that sum with the cost of the smallest card that can host the same model for the hours it would actually run. If the card would be idle for two-thirds of the month, the token price usually wins. If the card would be busy all day, or if the weights are yours and cannot go on a shared catalog model, the card usually wins.
All three locations are in the territorial scope of the GDPR. The company states that customer system data is processed without a transfer to a third country, and that a data processing agreement is part of the main contract. For teams that have already been asked to keep prompts and weights in the EU, that is the reason to rent the card here rather than in a cheaper region elsewhere. It is not a reason to rent a larger card than the model needs.
The practical sequence is therefore short: estimate the memory, pick the smallest card that fits, decide whether the duty cycle justifies a card at all, then choose the location and look at the availability matrix. The brand name of the GPU is the last column, not the first.
About the author: This article was written for ENTRONYX CLOUD, the cloud and hosting brand of ENTRONYX Deutschland GmbH, Cologne (HRB 98908, Amtsgericht Köln), founded in 2019. The platform runs in data centers in Frankfurt am Main, Düsseldorf, and Helsinki. Prices are net catalog figures as of November 18, 2026 and may change. Website: https://entronyx.cloud/en
Disclosure: The author works in communications for ENTRONYX Deutschland GmbH. The figures in this article come from the company’s public catalog. They are not a benchmark against other GPU clouds and not a promise that a particular model will fit a particular card in production.

