CloudGPU: Per-Minute GPU Rental and an OpenAI-Compatible Model API
CloudGPU rents on-demand GPUs — the RTX 5090 and RTX 4090 it names in its listing, with a wider catalogue behind them — billed per minute, and serves ready-made models through an OpenAI-compatible API billed per token. Both draw on one prepaid balance. It is built for AI developers rather than enterprises: one-click templates, root SSH, and even an integration layer that lets an AI agent rent compute on your behalf. It is also, in its own words, a public beta with no SLA. This profile describes what it publishes and quotes no prices.
How this listing got here. CloudGPU submitted this listing through our provider listing form. Everything below is drawn from CloudGPU's own website (homepage and pricing page, checked on November 9, 2026) and its listing application. On prices: billing is usage-based — per minute for GPUs, per token for the model API — and CloudGPU's own application points to its official pricing page for current rates, so this profile publishes no rate figures. The pricing page on cloudgpu.app is the only source of truth.
On this page
CloudGPU at a glance
| Item | What CloudGPU publishes |
|---|---|
| Service | On-demand GPU rental plus an OpenAI-compatible model API, aimed at AI developers |
| GPUs named in its listing | RTX 4090 and RTX 5090 |
| Full GPU catalogue | Wider than the listing suggests: consumer cards from RTX 2080 up to RTX 5090, plus A100, H800, H20 and other data-center cards on its pricing page |
| GPU billing | Per minute, rounded up, 1-minute minimum; unused minutes return to the balance when an instance is destroyed; no subscription and no minimum term |
| Model API | OpenAI-compatible (base URL https://cloudgpu.app/v1), billed per token; its homepage puts the range at 50 models across 5 modalities |
| Access | Root SSH and JupyterLab, a public HTTPS endpoint per deployment, and one-click templates (its homepage features ComfyUI / FLUX image workflows and MiniMax H3 video workflows) |
| Agent integration | An "Agent Compute Layer": an AI agent with HTTPS tools or an MCP client can discover compute, quote a task and send the user an approval link, with spending and time limits |
| Payment | One prepaid balance: card (including Alipay, Google Pay and Apple Pay at checkout), USDT on TRC-20, or bank transfer for businesses |
| Service status | Public beta — CloudGPU states there is no SLA: no uptime guarantee and no service credits |
| Prices | Not quoted in this profile — usage-based; see the official pricing page |
| Contact | Via its website — an enterprise "Talk to us" route; the listing application reached us from an address on the cloudgpu.app domain itself |
| Official site | cloudgpu.app |
Four things to know before you buy
Each point below is stated in CloudGPU's own pages or follows directly from them. They matter more than any rate card when you compare CloudGPU with the providers in our directory:
- It is a public beta with no SLA. CloudGPU says so itself: no uptime guarantee and no service credits. That is a fair trade for cheap, flexible capacity, but do not put a workload on it that needs a contractual uptime commitment.
- Destroying an instance erases its local disk. Its FAQ and its listing application both say the same thing: export your results first. An optional Workspace feature can save project files for eligible funded accounts, but CloudGPU is explicit that Workspace does not automatically back up the disk, and saved files sit under storage quotas and retention rules.
- You pay before you compute, and the balance is locked to the account. Everything runs off a prepaid balance topped up by card, USDT or bank transfer. CloudGPU states the balance does not expire — but it also cannot be transferred between users, so top up in amounts you expect to use.
- The operator behind the service is thinly documented. The pages we checked (homepage and pricing, November 9, 2026) publish the product, its rates and its payment routes, but no operating company name or headquarters that we could verify. The service is reachable and the application came from its own domain; still, do your own diligence before holding a large balance.
Two ways to use it: rental and API
CloudGPU sells the same capacity two ways. Rent the GPU: pick a template or a raw GPU, deploy it, and pay per minute for full control — its homepage positions this for running open models such as DeepSeek, Qwen, Llama and Mistral on dedicated GPUs, with training, inference and ComfyUI workflows named in the application. Or call the model API: use its hosted models through an OpenAI-compatible endpoint and pay per token from the same balance. One detail worth knowing if you use the API for brand-name models: CloudGPU states plainly that it is an independent service, not the official API of the model providers it resells.
The unusual third door is the Agent Compute Layer. Give an AI agent the integration instructions and — per CloudGPU's homepage — it can discover compute, quote a task and send you an approval link; MCP is one integration option, and connection approval is explicitly not unlimited spending permission: execution is bounded by spending and time limits you approve. If you let an agent spend money for you, those approval links are the control point to watch.
How billing and payment work
GPU time is billed per minute, rounded up, with a 1-minute minimum; one hour is pre-authorised while an instance runs, and unused minutes go back to the balance when you destroy it. There is no subscription, no minimum spend and, CloudGPU says, no egress charges. Model API calls are billed per token (some media models per image, character or second). Payment is a single prepaid balance: cards (Visa and Mastercard, with Alipay, Google Pay and Apple Pay in the same checkout), USDT on TRC-20 credited 1:1, or bank transfer (several currencies, usually 1–2 business days). An enterprise tier with reserved capacity and volume rates is quoted individually. Because every rate is live and usage-based, check the official pricing page at the moment you budget — and sanity-check any quote in our GPU cost calculator.
Who runs CloudGPU
CloudGPU's own pages describe the service in depth but, on the pages we checked, do not name an operating company or headquarters. What we could verify: the site is live, the catalogue and pricing are published, and the listing application was submitted from an email address on the cloudgpu.app domain by the person named in it as the applicant. If you represent CloudGPU and want the operator details added, contact us through our listing page and we will update this profile.
Comparing providers? Return to the GPU cloud provider directory for the official pricing pages of all listed clouds, read how GPU cloud pricing works, and sanity-check any quote in our GPU cost calculator.