# Sarvam Token Factory: Chinese AI Models, Sovereignty Concerns, And Developer Pricing In 2026

> Source: <https://www.swapupdate.in/sarvam-token-factory-chinese-ai-models-sovereignty-concerns-and-developer-pricing-in-2026/>
> Published: 2026-11-19 12:26:00+00:00

**Key takeaway:** Teams on the OpenAI toolkit can migrate to Token Factory by pointing base_url at Sarvam's v2 endpoint and changing only the model name, with serverless models available immediately and Provisioned Throughput requiring organisation-level enablement.

India’s AI community reacted sharply today to Sarvam’s Token Factory model list. The platform added Chinese-developed models, triggering questions around sovereignty and public subsidies. Sarvam clarified that inference runs on privately procured GPUs in India. It said subsidised IndiaAI compute ended in March 2026, addressing conflict concerns.

Launched on November 16, Token Factory serves open-weight models through an OpenAI-compatible API. Developers point their clients at the Sarvam endpoint, switch the model field, and go. Workloads run on Sarvam GPUs in India with optional zero data retention for enterprises.

## How Token Factory serves developers in India

Migration is quick for teams already on the OpenAI toolkit. Point base_url to Sarvam’s v2 endpoint and change only the model name. Serverless models work instantly; Provisioned Throughput needs enabling for your organisation. Language‑model requests are rate limited per account, with retries on 429 or 503.

## Why Chinese models on Indian inference raise sovereign-AI flags

Criticism focused on Chinese models listed alongside Sarvam’s indigenous stack. “The token factory is running on GPUs privately procured by Sarvam,” cofounder Pratyush Kumar said, adding subsidised access ended in March 2026. Sarvam also states prompts are processed in India and never forwarded to foreign providers’ APIs.

## Token Factory pricing and latency: India vs OpenRouter-style routing

Rupee pricing and local latency are the pitch. DeepSeek V4.1 Flash lists ₹14 per million input tokens and ₹55 for output. GLM‑5.3 Flash lists ₹13.5 and ₹45 respectively. Sarvam’s benchmarks from Mumbai show faster first‑token times than several global providers on long prompts.

## Compliance watch: SGI deepfake rules and MeitY consultation timeline

Text‑only chat outputs are not classified as Synthetically Generated Information under the 2026 amendments. Yet intermediaries must prevent unlawful content and handle synthetic audio‑visual media under strengthened rules. MeitY plans a consultation paper on AI safety, expected next month. Enterprises should review usage policies now.

## Career outlook: hiring pulse in MLOps, inference and platform engineering

Students targeting AI roles should track inference engineering, platform reliability and data governance. Useful skills include vLLM or Dynamo familiarity, context caching, structured outputs, observability and cost modelling. Knowing SGI labelling duties and incident response will strengthen interviews for enterprise teams.

Developers can trial serverless models now, and request Provisioned Throughput for gated ones. Set base_url, confirm rate limits, and enable zero‑retention before piloting with sensitive data. Track MeitY’s consultation window in December 2026, and document SGI safeguards early.
