GPU servers, billed by the hour

NVIDIA GPU servers you pay for in crypto, with no identity check. Sizes are listed live and open for ordering when the provider has capacity.

In short

Neckrez lists NVIDIA L40S GPU servers with 48 GB, 96 GB and 192 GB of video memory, billed by the hour from a crypto balance with no identity check. The sizes in the table below come straight from the live catalogue, including whether each one can be ordered right now. At the moment the provider is not allocating GPU capacity to our account, so these sizes cannot be ordered yet. They open for ordering by themselves when capacity returns, and you can check the status on this page or on the Deploy page at any time.

Everything else works like the rest of the platform: you sign up with an email, top up with crypto, pick a size and pay only for the hours the server exists.

Who a GPU server is for

A GPU is worth renting when the work is large and parallel: matrix maths for machine learning, video encoding, 3D rendering, scientific simulation. For a website, a bot, a VPN or a database it only adds cost, and a regular VPS is the honest answer.

• Running a model. Inference for language, image or speech models, where video memory decides whether the model fits at all.
• Fine-tuning and experiments. Short runs where owning a card for a month makes no sense but renting one for an afternoon does.
• Rendering and transcoding. Batch jobs that finish and release the machine.
• CUDA development. Testing code on real hardware before you commit to buying any.

What is listed

The listed cards are NVIDIA L40S, a data-centre GPU with 48 GB of video memory. The 96 GB and 192 GB sizes are the same family with more video memory and with more vCPU, system RAM and NVMe storage around it; the exact figures are in the table. Prices are the live prices from the catalogue in US dollars and change with it.

Availability, stated plainly

A GPU appears in a provider's price list long before it appears in the provider's stock. Vultr publishes these sizes for its locations but allocates the physical cards region by region and account by account, and today none are allocated to us. The table shows exactly that: the size, the price and a clear mark that it cannot be ordered.

We do not take payment for a size that cannot be created. A size marked unavailable is refused by the Deploy page, and when the provider allocates capacity the mark disappears with no change on your side.

How to choose a size: video memory first

For machine-learning work the first question is whether the model fits in video memory. As a rough guide, the weights take about two bytes per parameter in 16-bit precision and about half a byte per parameter in 4-bit quantisation, plus room for the context cache and activations. A 7-billion-parameter model therefore needs around 14 GB in 16-bit and roughly 4 GB quantised, a 13-billion-parameter model about 26 GB in 16-bit, and a 70-billion-parameter model fits in 48 GB only when heavily quantised and runs comfortably on the larger sizes.

Rule of thumb: take the smallest size whose video memory leaves about a fifth free after the weights, then check system RAM and disk for your data. Memory matters far more than clock speed.

Billing and payment

GPU servers are billed hourly from the same prepaid crypto balance as every other server, with no reservation and no minimum term. Destroying the server stops the charge. Because the hourly price of a GPU is higher than that of a small VPS, keep an eye on the balance page: it shows how long your balance lasts at the current burn rate.

How to start with crypto

• 1. Sign up with an email. No KYC, no ID, no phone and no card.
• 2. Top up the balance in crypto. Bitcoin, USDT, ETH or any of 300+ coins. The minimum top-up is $15 and the payment page shows the exact amount for your coin and network.
• 3. Pick a GPU size. On the Deploy page choose the GPU line. Sizes that can be ordered are enabled, the others say why not.
• 4. Connect and work. Root access over SSH with a dedicated IPv4, then set up your framework: the driver and CUDA versions depend on the image you choose.
• 5. Destroy it when the job is done. Billing stops.

FAQ

How much does a GPU server cost per hour? The table above shows the live price per hour and per month for every listed size. Billing is hourly with no minimum term.

Can I order one today? Not at the moment: the provider is not allocating GPU capacity to our account. The sizes stay listed with their live status and become orderable by themselves when capacity returns.

Do I need to verify my identity? No. An email is enough to sign up, and payment is in crypto, so there is no card or ID check.

Can I pay with Bitcoin or Monero? Yes. Bitcoin, Ethereum, USDT on several networks, Monero and hundreds of other coins are accepted through NOWPayments. The payment page shows the exact amount and the minimum for the coin you pick.

What happens if my balance runs out? The same balance rules as for every other server apply, described in the Refunds & Billing policy. Keep enough balance for the whole run and copy results off the machine as you go.

When is a dedicated CPU server enough instead? If your job does not need a graphics card, Dedicated CPU gives you reserved cores without the GPU price, and a regular VPS is cheaper still for light work.