ZettaGPU

[docs]developer guide

ZettaGPU documentation

OpenAI-compatible API, contributor agent guide, and the concepts behind the network. Pricing is up to ₹350/M tokens, or ₹12,500/month for a 50-seat classroom with 50M tokens. Pricing varies. Available pan-India, seamlessly.

Quickstart

Swap base_url and api_key. Keep your existing OpenAI SDK code.

import openai

client = openai.OpenAI(
    api_key="YOUR_ZETTA_API_KEY",
    base_url="https://api.zettagpu.in/v1"
)

response = client.chat.completions.create(
    model="gemma-4-26b-moe",
    messages=[{"role": "user", "content": "Hello from IIT Bombay"}],
)

print(response.choices[0].message.content)

Base URL: https://api.zettagpu.in/v1. Request a key from the AI Compute signup.

Authentication

Send your API key as a Bearer token. Never commit keys to source control.

Authorization: Bearer YOUR_ZETTA_API_KEY

Rotate keys from your dashboard once accounts go live. If a key leaks, revoke it immediately and contact contact@zettagpu.com.

API reference

POST/v1/chat/completions

Chat completions

OpenAI-compatible chat endpoint for Gemma 4 26B MoE and other supported open models.

GET/v1/models

Models

List available models on the ZettaGPU network.

GET/v1/usage

Usage

Token usage and billing summary for your workspace.

GET/v1/health

Health

Cluster health and GPU Contributor pool availability.

Full OpenAPI reference publishes at launch (April 2026). Get my AI Compute for early access credentials.

Concepts

How scheduling, metering, and the sandbox keep the network reliable.

Distributed routing

When you request inference, the scheduler matches your job to healthy Contributor nodes, with automatic failover if a node drops offline.

ZettaUnit

One ZettaUnit is a 50 tok/s target throughput reservation, subject to fleet availability (not a hard SLA). Price per ZettaUnit-hour is ₹63.

Metering

AI Users are billed on delivered tokens (or seat plans). GPU Contributors are paid for every token they deliver, measured by the scheduler.

Sandbox model

The Contributor agent runs workloads in a constrained environment. Payloads are processed in GPU memory. The agent pauses when local GPU load spikes, so it does not affect day-to-day work.

Contributor guide

Run a node, understand metering, and get paid for idle compute.

1. Check requirements

NVIDIA RTX 3060 or better, Windows 10/11, 16 GB RAM, stable broadband (10 Mbps+).

2. Join the waitlist

Request the Windows agent from the GPU Contributors page. We email the EV-signed installer before launch.

3. Set caps and go online

Install the agent, set availability hours, and leave it running when the machine is idle. It does not affect your day-to-day work.

4. Get paid

Earn up to ₹5,300/mo. The up to ₹10/hr availability bonus applies only to premium RTX 5090-class nodes. Payouts via UPI/NEFT (₹500 minimum). Earnings vary with your GPU, hours online, and network demand. Estimates only, not a guarantee.

Pricing snapshot

Per Token

Up to ₹350/ 1M tokens

Up to · pan-India · pricing varies

Per Classroom

₹12,500/ month

50 seats · 50M tokens included

Per ZettaUnit-hour

₹63/ hour

50 tok/s guaranteed

We do not compete with subsidized IndiaAI bare GPU rental on per-token price. ZettaGPU includes managed, India-hosted inference available pan-India seamlessly, automatic failover, INR billing, and zero MLOps. See AI Compute pricing.