About OverGPU

OverGPU rents dedicated GPU servers by the month. Each GPU you rent is a whole physical card that belongs to you alone for as long as you pay for it, and the price is one flat monthly figure you know before you order. This page covers what you should know about OverGPU before you order: what we offer, why we chose monthly terms and whole GPUs, the principles we hold ourselves to, and how an order works.

About OverGPU: what we do and who it is for

We rent two kinds of servers. Single-GPU servers are a dedicated virtual machine with the full GPU passed through, from a GeForce RTX 5090 with 32 GB of memory to an NVIDIA B300 with 288 GB. 8-GPU servers are bare-metal machines: NVIDIA HGX B300, B200, H200 and H100, AMD Instinct MI355X and RTX PRO 6000 Blackwell. NVIDIA GB300 NVL72 racks are available on request, and reservations are open for Vera Rubin NVL72. Every model is described on our GPU overview.

We build for teams whose GPUs stay busy for weeks or months: ML engineers training and fine-tuning models, AI startups and inference teams serving them around the clock, research labs working to fixed budgets, VFX studios rendering through crunch periods, and Web3 teams running zero-knowledge provers. They share a need for the same hardware every day, a bill they can plan, and root access to a machine they control.

Why we rent by the month

Hourly billing suits short, bursty jobs. Much serious GPU work is not like that. A training run, a fine-tuning campaign or a model served to users keeps a GPU busy all month, and a meter running on every hour of it adds uncertainty without adding anything useful. Monthly terms give you three things instead.

• Predictable cost. One price per server per month, billed in advance and fixed for each paid month. Finance gets a number, not an estimate. A single NVIDIA H100 is $2,190 a month, about $3.00 per GPU-hour if it runs around the clock.
• An environment that stays put. Datasets, checkpoints, containers and installed packages stay on local NVMe between jobs. Nothing is torn down when a session ends, and there is no queue to rejoin.
• No meter. Because the price is flat, an idle server costs nothing extra. You can pause to debug or leave a job queued overnight without watching the clock.

Monthly is not always the right answer. If you need a GPU for a few hours a week, hourly billing elsewhere will usually cost less, and we would rather say so up front. Our monthly vs hourly pricing comparison shows where the break-even point sits.

Why whole GPUs, never slices

Shared GPU platforms divide one card between users with virtual GPU slices, MIG partitions or time-sharing. We do none of that. Each GPU we rent is assigned to one customer: our 8-GPU servers are bare metal, and our single-GPU plans pass the entire card through to a dedicated VM, so all of its memory and compute are yours.

This matters in practice. Memory is usually the hard limit in AI work, and a whole H200 gives you all 141 GB of HBM3e, not a fraction of it. Performance stays consistent because no other workload competes for the same GPU's compute or memory bandwidth, which also makes a slow run easier to diagnose. Our article on dedicated vs shared GPUs covers the trade-offs in more detail.

Our principles

Everything published

Every configuration we offer is listed on our GPU server pricing page with its GPU, CPU, RAM, NVMe storage, network cards, price and stock status: in stock, limited stock, on request or reservations open. You should not need a sales call to learn what a server costs or whether we have one.

Commitments in writing

Our promises live in our terms, not only in our marketing. We commit to 99.9% monthly availability for power, network and hardware, with service credits of up to 100% of a server's monthly fee, as set out in our service level agreement. Single-GPU and 8-GPU servers are delivered in under 10 minutes after payment. There is no setup fee, and no egress fees within the included bandwidth of 20 TB per month on single-GPU plans and 50 TB on 8-GPU servers, under fair use.

Dollar prices, paid from a prepaid balance

Prices are set in US dollars and paid from a prepaid balance in your account, which you top up in crypto: Bitcoin, Ethereum, Monero, USDT on TRON (TRC-20) or Solana. Card payments are coming soon. You see the price before anything is charged, and every server payment and renewal comes with an invoice. Support comes from engineers, 24 hours a day, every day of the year.

How we work

Ordering is deliberately simple:

1. On the deployment page, you choose the configuration, OS image, optional software and login (root password or SSH key) you need, and create a free account or log in, all in the same form.
2. The configuration and price are shown before you pay.
3. You pay the first month from your balance, which you top up in crypto. If it is short, you top up in the same form, and the order is paid as soon as the top-up is credited.
4. We deliver the server in under 10 minutes after payment, with root access over SSH. Its IP address, SSH details and data-center location appear in your dashboard, and your first month starts on the day of delivery.

From then on, the server renews monthly from your balance and we email you before each renewal. You can cancel any month with no fee from your dashboard, and the server keeps running until the end of the paid month. When you need help, you open a ticket from your dashboard and an engineer answers, at any hour. Before you order, answers are in our help center, docs, GPU pages and calculators; see how support works.

We keep verification proportionate. Most orders need only an account and a funded balance. We ask for more only in specific cases, such as large orders, fraud or sanctions screening, or when the law requires it, and we always tell you what we need and why. Our verification policy has the details.