Turn your idle gaming GPU into a private AI & rendering server for friends and family — they only pay for electricity used.
- 🔌 Fair pricing: Users only pay actual electricity costs — no cloud markups
- 🔒 Privacy first: Your data never leaves the trusted group
- ⚡ Zero waste: That RTX 4090 mining dust? Now it's serving AI models
- 🎮 Keep gaming: Pause jobs instantly when you need your GPU back
- 💳 No vendor lock-in: Works with Ollama, OpenRouter, standard Blender
- Indie teams sharing a workstation GPU for development
- Friends splitting the cost of AI experiments
- Families with one powerful PC serving everyone's needs
- Students pooling resources for coursework
It runs two services off your GPU:
| Service | How | Billed by |
|---|---|---|
| AI Inference | Ollama (local) or OpenRouter (cloud) | Per million tokens |
| 3D Rendering | Blender job queue | Per render-minute |
Costs are derived from your actual electricity rate and GPU wattage. You set ELECTRICITY_RATE_KWH and GPU_INFERENCE_WATTS in a .env file — all prices flow from that. No margins, no markup, no hardcoded rates.
Users get a credit balance. They top up via Stripe (or you manually adjust their balance), use inference/rendering, and get a monthly invoice for what they owe. You're not making a business out of this — you're just not subsidising other people's GPU time.
Hardware: Any GPU that can run Ollama models. The bigger the VRAM, the bigger the models you can load. A 12–16GB card handles 14B parameter models comfortably.
Software:
Optional (can add later):
- Stripe — for card payments and invoicing
- Cloudflare R2 — for render output storage (free tier is fine)
- Resend — for email notifications
- A custom domain — or use the free
trycloudflare.comURL
Vercel (always on) Your PC (via Cloudflare Tunnel)
────────────────── ────────────────────────────────
React frontend ────────► FastAPI proxy ──► Ollama
Job worker ──► Blender CLI
Neon / Supabase Cloudflare R2
users, ledger, .blend files,
jobs, invoices rendered output
The frontend and database stay reachable even when your PC is off. Users can check balances and view invoices any time. Only inference and render submission need your machine up.
Your machine never needs an open port. Traffic enters through a Cloudflare Tunnel — you run cloudflared on your machine and it maintains an outbound connection to Cloudflare's edge.
macOS / Linux — paste this into your terminal:
curl -fsSL https://raw.githubusercontent.com/Slaymish/GPUShare/main/install.sh | bashWindows — double-click setup.bat, or in PowerShell:
.\setup.batFully silent install (no prompts, all defaults):
./setup.sh --quick # macOS / Linux
.\setup.ps1 -Quick # Windows
.\setup.bat --quick # Windows (via .bat)The setup script auto-detects your GPU and OS, then:
- Checks prerequisites — Docker, Git, port availability
- Installs Ollama — via Homebrew (macOS), winget (Windows), or official script (Linux)
- Recommends a model — based on your GPU VRAM (4B for 4GB, 8B for 8GB, 14B for 16GB, 32B for 24GB+)
- Asks for your database URL — with instructions for free Supabase/Neon setup
- Configures your node — electricity rate, GPU wattage (auto-estimated), name
- Optional integrations — Stripe billing, OpenRouter, 3D rendering, email
- Builds and starts Docker services — with health checks
- Optionally starts a Cloudflare tunnel — free public URL, no account needed
Total time: ~10 minutes (interactive) or ~5 minutes (--quick).
| Flag | Description |
|---|---|
--quick / -Quick |
Non-interactive, use all defaults |
--skip-ollama |
Don't install Ollama or pull models |
--skip-tunnel |
Don't start Cloudflare tunnel |
--dry-run |
Show what would happen without executing |
1. Install Ollama
# macOS
brew install ollama && brew services start ollama
### Manual setup
**1. Install Ollama**
```bash
# macOS
brew install ollama && brew services start ollama
# Linux
curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen2.5:14b # or whatever model you want to serve2. Start the server
docker compose up -dThis starts fastapi on localhost:8000 and the render worker. Ollama runs natively on the host and is reached via host.docker.internal:11434.
3. Expose your server
# No account needed, URL changes on restart:
cloudflared tunnel --url http://localhost:8000
# Or with your own domain (persistent URL):
cloudflared tunnel login
cloudflared tunnel create gpushare
cloudflared tunnel route dns gpushare gpu.yourdomain.com
cloudflared tunnel token gpushare # add TUNNEL_TOKEN to .env
cloudflared tunnel run gpushare4. Deploy the frontend
Connect the repo to Vercel and set the root directory to packages/frontend. No environment variables needed — vercel.json proxies /v1/* and /health requests to your server automatically.
5. Point Stripe webhooks (if using Stripe)
Add a webhook endpoint in the Stripe dashboard:
https://your-vercel-app.vercel.app/api/webhooks/stripe
Events needed: invoice.paid, invoice.payment_failed, checkout.session.completed
All config lives in .env. Key values:
# Your electricity rate drives all cost calculations
ELECTRICITY_RATE_KWH=0.346 # e.g. 34.6c/kWh in Wellington
# Measured GPU wattage (use a smart plug for accuracy)
GPU_INFERENCE_WATTS=150
GPU_RENDER_WATTS=300
SYSTEM_WATTS=80
# Access control — recommended when sharing with a small group
INVITE_ONLY=true
REQUIRE_APPROVAL=true
ADMIN_EMAIL=you@example.comChange ELECTRICITY_RATE_KWH and all future usage reprices automatically. Past ledger entries are not touched.
The inference endpoint is OpenAI-compatible — any client that works with OpenAI works here without modification:
curl https://gpu.yourdomain.com/v1/chat/completions \
-H "Authorization: Bearer gpus_sk_YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "qwen2.5:14b", "messages": [{"role": "user", "content": "Hello"}]}'Both /v1/chat/completions and /v1/inference/chat/completions are supported. Models with a / in the ID (e.g. anthropic/claude-3.5-sonnet) are automatically routed through OpenRouter.
Users run a postpaid balance. Soft warnings at $0 and -$5, hard block at -$20 (configurable per user). Monthly invoices are generated automatically. If you're not using Stripe, you can manage balances manually through the admin panel.
.blendfiles are sanitised before queuing — embedded Python scripts are stripped via headless Blender before the file is stored- API keys are stored as bcrypt hashes only, shown once on creation
- Ollama is never exposed through the tunnel (localhost only)
- All inbound traffic enters via Cloudflare Tunnel — no open ports required on your machine
pnpm dev # frontend + backend together
pnpm docker:up # start all Docker services
pnpm docker:logs # follow logs
pnpm docker:rebuild # rebuild and restart
pnpm db:migrate # run migrations
pnpm db:revision # create new migrationI would love help! Check out:
Current priorities:
- NVIDIA GPU monitoring dashboard
- AMD GPU support
- Automatic job scheduling
- Mobile app
- One-click installer (Windows, macOS, Linux)
MIT