Launching August 12, 2031
gpu.nz

gpu.nz

Cloud GPUs by the minute, driven by API, CLI or your AI agent

SubscriptionAICodeHosting7 impressionslaunches Aug 121 comment

Comments

>log in to comment
  • Egor[maker]· 3d ago

    Hi, I'm Gosha, founder of gpu.nz. Servers are billed by the minute, and a stopped server costs only its disk. The part I'd most like feedback on is the MCP server: add https://gpu.nz/mcp to Claude Code or Cursor and ask it to find the cheapest 24 GB card and start Ollama. Browsing offers and the VRAM estimate work without a key. Happy to answer questions about the setup.

gpu.nz rents GPU servers by the minute: from RTX 3060 to H200 and B300, single cards or 2-8 GPU boxes, with live prices and stock. Start ComfyUI, Jupyter, Ollama or vLLM from a template, or any Docker image, and stop paying the moment you stop the server. Developers can do all of it from a REST API (OpenAPI 3.1), the open-source gpunz CLI (Homebrew, Scoop), or the remote MCP server at gpu.nz/mcp, which lets Claude, Cursor or Codex find a GPU and launch it. There is also a free VRAM calculator built on real GGUF file sizes for 35 LLMs.

gpu.nz provides on-demand GPU server rentals billed per minute via API, CLI or UI.

for
Developers and AI researchers needing short-term GPU compute.
pricing
paid
license
MIT
gpunz/gpunz-cli 1 0Goupdated 3 days ago
works withDockerSSHJupyterPyTorchOllamavLLMStable DiffusionComfyUI

Key features

6 features of gpu.nz
  • Per-minute billing — Pay only for the exact minutes the GPU runs, with prices fixed at launch.
  • API and CLI access — Control servers programmatically via an OpenAPI 3.1 REST API or the open-source gpunz CLI.
  • Ready-to-run templates — Launch pre-configured environments like ComfyUI, Stable Diffusion, Ollama, vLLM, Jupyter, or your own Docker image.
  • Global GPU pool — Access 3,886 GPUs across 62 countries with live price and stock updates.
  • SSH and web UI access — Connect to running servers over SSH or through web interfaces for each environment.
  • Interruptible servers — Cheaper bid-based instances that pause when outbid, ideal for batch jobs.

Use cases

  • Run a single-GPU Stable Diffusion job and stop when finished
  • Train a LoRA model with ComfyUI and pay only for active minutes
  • Serve an LLM via vLLM with an OpenAI-compatible endpoint
  • Execute Jupyter notebooks for PyTorch experiments without owning hardware
  • Render video frames using RTX 4090 and shut down instantly after rendering

gpu.nz vs alternatives

gpu.nzModalRenderFirebaseFly.io
Best forPer-minute GPU rentals via APIServerless GPU functionsGPU rendering farmsRealtime app backendEdge app deployment
PricingPaidSubscriptionSubscriptionSubscriptionSubscription
DevHunt upvotes10000
LaunchedAug 2031————
  • gpu.nz vs Modal: Modal offers serverless GPU workloads in Python but focuses on managed functions rather than per-minute VM rentals.
  • gpu.nz vs Render: Render provides cloud GPU rendering services, not general-purpose GPU servers with API control.
  • gpu.nz vs Firebase: Firebase is a full-stack app platform without on-demand GPU instances.
  • gpu.nz vs Fly.io: Fly.io runs full-stack apps close to users, lacking minute-based GPU rentals.

gpu.nz FAQ

How is per-minute billing calculated?+

A minute costs the hourly rate divided by 60; the rate is fixed when the server starts.

Do I pay for a stopped server?+

GPU billing stops when the server is stopped, but disk storage continues to be billed.

How can I top up my balance?+

Balance can be topped up via SBP from Russian banks or with cryptocurrencies (USDT, TON, BTC, etc.) starting from $2.

What happens if my balance runs out?+

You receive warnings; if the balance cannot cover even one minute, the server and its disks are deleted.

Can I choose the server location?+

Yes, you select a country and region from the catalog; servers are available in 62 countries.

What are interruptible servers?+

They run on a bid system, are cheaper, and can be paused if outbid, suitable for non-critical batch jobs.

Summarized by DevHunt from gpu.nz · Oct 19, 2026. Details may change; check the official site.