A local desktop app that recommends the right GPU or cloud infrastructure for your AI workload. Works in two modes — instant dataset-based scoring with no account required, or AI-powered analysis using your existing Claude Pro or ChatGPT Plus subscription.
Mac / Linux
bash launch.shWindows
Double-click launch.bat
The launcher checks for Node.js v18+, installs it automatically if missing (with your permission), starts the server on http://localhost:3131, and opens your browser.
No npm install. No dependencies. Pure Node.js built-ins only.
Click "Skip sign-in — get instant dataset recommendations" on the auth screen.
The scoring engine evaluates all 17 platforms against your requirements entirely in the browser — no server call, no AI, instant results. You get:
- Top 4 ranked platforms with a match score out of 100
- 8 metrics per platform: token speed, monthly cost, annual cost, max model size, power draw, concurrent users, CUDA support, cooling
- Pros, cons, and sovereignty tags on every card
- Up to 4 context-aware optimisation tips on the top pick (INT4 quant, Dynamo 1.0, IndiaAI subsidy, etc.)
Sign in with your existing subscription (no new API key purchase). The server sends your requirements to Claude Sonnet 4.6 or GPT-4o with full India 2026 pricing context and returns a structured recommendation with rationale, cost breakdown, alternatives, and gotchas.
If the AI call fails for any reason, the app automatically falls back to dataset mode.
- Open your terminal and run
claude setup-token - Copy the token it prints — it looks like
sk-ant-st01-xxxxxxxxx... - Paste only that token value (not the
exportline) into the app
Requires the claude CLI: npm install -g @anthropic-ai/claude-code
Tokens are valid for ~1 hour. Run claude setup-token again when it expires.
Click "Open ChatGPT Sign-in". A browser tab opens at auth.openai.com. You sign in normally. OpenAI redirects back to localhost:3131, the server exchanges the code for a token, and you're in. Nothing is stored to disk.
Paste any Anthropic or OpenAI API key directly. Stored in memory for the session only.
- Claude keys:
sk-ant-api03-... - OpenAI keys:
sk-...
| Field | Options |
|---|---|
| Expected run time | ⚡ Ad-hoc / 🕐 Fixed hours/day (slider 1–24h) / ♾️ 24×7 Always-on |
| Concurrent users | 1 · 2–5 · 10–25 · 25+ |
| Primary use case | Inference · Fine-tuning · RAG · Edge AI · Research · Multimodal · Agentic · Education |
| Target model size | ≤3B · 7B · 13–30B · 70B · 100B+ |
| Data sovereignty | 🇮🇳 India-only (DPDP/BFSI) · Prefer India · No restriction |
| Budget | No limit · Monthly cap (₹) · One-time CapEx (₹) |
| CUDA required | Yes · No · Either |
| Deployment | On-premise · Cloud · Hybrid |
| Device | VRAM | Price (INR) | Max Model | CUDA |
|---|---|---|---|---|
| Raspberry Pi 5 + AI Kit | 8 GB | ₹7,200 | 3B | ✗ |
| Jetson Orin Nano Super 8GB | 8 GB | ₹21,600 | 8B | ✓ |
| Mac Mini M4 (16 GB) | 16 GB unified | ₹57,900 | 7B INT4 | ✗ |
| Mac Mini M4 Pro (64 GB) | 64 GB unified | ₹1,69,900 | 30B INT4 | ✗ |
| AMD Strix Halo Mini PCs (128 GB) | 128 GB unified | ~₹1,75,000 | 70B INT4 | ✗ |
| NVIDIA Jetson AGX Thor | 128 GB unified | ₹2,16,250 | 70B FP8 | ✓ |
| HP ZGX Nano G1n (GB10) | 128 GB LPDDR5X | ~₹2,59,500 | 200B FP4 | ✓ |
| MSI EdgeXpert MS-C931 (GB10) | 128 GB LPDDR5X | ~₹2,59,500 | 200B FP4 | ✓ |
| ASUS Ascent GX10 (GB10) | 128 GB LPDDR5X | ~₹3,02,500 | 200B FP4 | ✓ |
| Dell Pro Max GB10 | 128 GB LPDDR5X | ~₹3,46,000 | 200B FP4 | ✓ |
| NVIDIA DGX Spark (GB10) | 128 GB LPDDR5X | ₹4,06,454 | 200B FP4 | ✓ |
| Mac Studio M4 Max (128 GB) | 128 GB unified | ~₹4,19,000 | 70B INT4 | ✗ |
| TAALAS 10× RTX 4090 | 240 GB GDDR6X | ₹38–43 lakhs | 130B INT4 | ✓ |
GB10 family note: ASUS GX10, MSI EdgeXpert, Dell Pro Max, HP ZGX Nano, and NVIDIA DGX Spark all run the identical GB10 Grace Blackwell chip (1 PFLOP FP4, 128 GB, 273 GB/s). They differ only in thermal design, storage options, software bundle, and price. MSI is the cheapest; Dell has the best thermals; ASUS has the best cooling design; HP has the best software stack.
| Provider | Rate (INR/hr) | GPU | IndiaAI Subsidy |
|---|---|---|---|
| Yotta Shakti | ₹115–400 | H100 / GH200 / B200 | ✓ ₹115/hr eligible |
| E2E Networks | ₹150–250 | H100 Hopper | ✓ empanelled |
| Jio Cloud | ₹200–350 | GH200 | ✓ empanelled |
| Sify CloudInfinit+AI | ₹200–380 | H100 / H200 | ✓ empanelled |
| Tata Communications | ₹220–400 | Hopper + Blackwell | ✓ empanelled |
| CDAC / PARAM Rudra | ₹115–150 | H100 / A100 / MI300X | Core provider — research only |
IndiaAI Mission subsidised compute: eligible Indian startups building foundation AI models get H100/GH200 at ₹115–150/hr (vs ₹250–400 full rate). Apply at indiaai.gov.in.
| Provider | Rate (INR/hr/GPU) | Break-even vs DGX Spark |
|---|---|---|
| Oracle Cloud | ₹108 | 5.2 months |
| RunPod | ₹172 | 3.3 months |
| GCP A3-High | ₹260 | 2.2 months |
| AWS p5.48xlarge | ₹340 | 1.7 months |
| Azure NC H100 v5 | ₹604 | 0.9 months |
The engine applies hard disqualifiers first, then scores positively:
Hard disqualifiers — any of these returns a score of -1 (platform excluded):
- Platform's max model size is smaller than the requested model
- India-only sovereignty requested but provider is not India-sovereign
- CUDA required but platform has no CUDA
- On-premise only but platform is cloud (or vice versa)
- Platform's max concurrent users is less than requested
- Budget exceeded (monthly or CapEx)
Positive scoring — points added for:
- India sovereignty match when requested (+30)
- Prefer-India match (+15)
- CUDA available when preferred (+8)
- Users headroom (2× or more capacity) (+12)
- Budget under 60% of cap (+8)
- Correct usage pattern (ad-hoc → cloud +20, 24×7 → on-prem +18)
- Project type in platform's
bestForlist (+25) - Model size headroom at 120B+ (+20)
- Token speed bonus (up to +18)
- Specific boosts: Jetson Thor for Edge AI (+30), Jetson Nano for budget edge (+25), RPi for education (+30), CUDA platforms for fine-tuning (+12)
Optimisation tips are matched to the form inputs — e.g. IndiaAI subsidy tip only appears when India sovereignty is selected, INT4 quantization tip appears for 30B+ models, Dynamo 1.0 tip appears for multi-user or agent workloads.
gpu-advisor/
├── launch.sh ← Mac/Linux launcher
├── launch.bat ← Windows launcher
├── server.js ← Node.js backend (zero npm dependencies)
├── package.json ← Metadata only
├── README.md ← This file
├── OAUTH.md ← Detailed OAuth flow documentation
└── public/
└── index.html ← Full app UI (auth + form + dataset engine + results)
| Platform | Install methods tried (in order) |
|---|---|
| macOS | Homebrew → nvm → official .pkg |
| Linux | apt/NodeSource → dnf → yum → pacman → nvm |
| Windows | winget → PowerShell MSI download |
- Tokens stored in RAM only — nothing written to disk
- No request logging — errors print to terminal only
- All tokens cleared on app exit (Ctrl+C or closing the terminal)
- Browser stores only the session ID in
sessionStorage(tab-scoped, cleared on tab close) - Server binds to
127.0.0.1only — not reachable from your network - Zero npm packages — no supply chain risk
"Failed to fetch" on any button
The server isn't running. Re-run bash launch.sh or node server.js in the project folder.
Claude: "Token format invalid" Copy only the token value — not the full export line.
✓ sk-ant-st01-xxxxxxxxxxxxxxxx...
✗ export CLAUDE_CODE_OAUTH_TOKEN=sk-ant-st01-xxxxxxx...
Claude: session expired
Setup-tokens last ~1 hour. Run claude setup-token again and paste the new token. Or switch to an API key which has no expiry.
ChatGPT: OAuth fails or blank screen Use the API key option instead — create a key at platform.openai.com.
Port 3131 already in use
# Mac/Linux
lsof -ti:3131 | xargs kill
# Windows
netstat -ano | findstr :3131
taskkill /PID <pid> /FWant recommendations without signing in? Click "Skip sign-in — get instant dataset recommendations" on the auth screen.
- Node.js v18+ (auto-installed by launcher if missing)
- Internet connection (for OAuth and AI API calls — dataset mode works offline)
- For AI mode: Claude Pro +
claudeCLI, or ChatGPT Plus, or any API key
Prices as of March 2026. Cloud GPU rates are dynamic — verify before procurement.
Full research report with all benchmarks and comparison tables: GPU_LLM_Infrastructure_India_2026_v3.docx