Incident sheet · /infrastructure

Filed
Read
4 min · 837 words
Severity
Routine

Sarvam’s Token Factory faces scrutiny over models and subsidies

Sarvam’s November 18 Token Factory launch offers 12 named models on Indian GPUs, prompting debate over Chinese models, public subsidies and India’s AI sovereignty.

Sarvam’s token factory stirs debate on India’s sovereign AI push
FeedSarvam’s token factory stirs debate on India’s sovereign AI push — ann_zima / pixabay

Event log

  1. Token Factory launched November 18, 2026, at Rs 14 per million input tokens and Rs 55 per million output tokens.
  2. Sarvam's published catalogue lists 12 named entries; Monga described an eight-model lineup containing six Chinese models, one Google model and one Sarvam model.
  3. Monga said an April 2025 government selection brought Rs 98.68 crore for 4,096 H100s; Sarvam later raised $234 million at a $1.5 valuation in June.
  4. Sarvam says Token Factory uses privately procured GPUs and no government subsidies; its 40% discounted IndiaAI Mission access ended in March 2026.
  5. Sarvam signed a deal with HQ IDS to develop and deploy AI applications for the Indian Army, Navy and Air Force.

By Tom Whitfield

Sarvam launched Token Factory on November 18, 2026, with introductory pricing of Rs 14 per million input tokens and Rs 55 per million output tokens. The Bengaluru startup describes the platform as a commercial inference stack for startups, enterprises and developers using Sarvam's computing infrastructure in India.

Token Factory's published catalogue includes six Chinese models from developers including DeepSeek, Moonshot AI and Z.AI Co, alongside Google's Gemma and Sarvam 105B. The full list contains 12 named entries, although RunAnywhere CEO Sanchit Monga characterised the lineup as eight: six Chinese, one Google and one Sarvam.

The 12 entries are:

  • DeepSeek V4.1 Flash 1017
  • Kimi K3.1 Pro
  • Sarvam Prism
  • Qwen 3.8 27B
  • GLM 5.3
  • GLM 5.3 Flash
  • Clef
  • EmbeddingGemma 2
  • Gemma 4 31B
  • DeepSeek V4 Pro
  • Sarvam 105B
  • Sarvam 105B Conversations

Sarvam says the platform makes leading open-weight models faster and cheaper to run. Its single API lets businesses build applications, agents, workflows and copilots without buying GPUs or arranging compute, infrastructure and coding expertise; building an AI product from scratch remains expensive and time-consuming.

The backlash came days after Sarvam signed an agreement with the Headquarters Integrated Defence Staff, Ministry of Defence, to develop and deploy AI applications for the Indian Army, the Indian Navy and the Indian Air Force. The timing placed a commercial hosting launch beside a national defence contract, sharpening the sovereignty argument.

Monga, CEO of US-based AI inference company RunAnywhere, criticised Sarvam for “mostly serving Chinese models from India GPUs.” He said the government picked Sarvam in April 2025 to build India's own foundational model, then awarded a Rs 98.68 crore GPU subsidy for 4,096 H100s; Sarvam also raised $234 million at a $1.5 valuation in June.

Monga's verdict was less charitable: “So India's sovereign AI lab is mostly serving Chinese models from Indian GPUs. That's a hosting business, and plenty of companies already run one.” He wanted “the lab that ships frontier research out of India,” arguing that public capital and subsidised compute should produce more than hosting.

Srikanth Lakshmanan, founder of CashlessConsumer, warned that concentrating inference in one provider could centralise critical infrastructure. He called Sarvam a “de facto Kaku (National) inference corporation of Aadhaaristan” and urged the government to decentralise GPUs to districts, making growth “distributable, equitable.”

Following the social-media backlash, Sarvam said Token Factory runs entirely on privately procured GPUs and receives no government subsidy. Co-founder Pratyush Kumar said its earlier subsidy related to training, not the new service.

“Through the IndiaAI Mission, Sarvam had received subsidised access (40% discount) to GPUs in India to train models. The outcome was the creation of open weight models, which are published on HuggingFace and AIKosh, with permissible licences. Our access to these GPUs ended in March 2026,” Kumar said.

Kumar said Token Factory aims to provide a domestic alternative to foreign inference platforms, with rupee pricing and data sovereignty. A Sarvam FAQ states: “Inference runs on Sarvam-operated GPUs in India. Requests are not forwarded to the original model developer's overseas API.”

Nikhil Pahwa, founder and editor-in-chief of MediaNama, said he saw no issue with Sarvam launching an inference platform or hosting Chinese models. “Sarvam needs to build its monetisation and for that they have to compete in the open market,” he said.

Pahwa called coding the primary product-market fit for models and said Sarvam was targeting coders. Apart from access to Sarvam's own models, he found DeepSeek V4.1 Flash priced at market parity with OpenCode and Command Code, leaving “very little reason to switch to Token Factory.”

On public support, Pahwa said the government's contractual obligations, not public sentiment, determine Sarvam's duties. “Nationalism is the pitch that got them the GPUs, but they're not expected to deliver anything beyond their contract,” he said, adding that an inference launch does not mean Sarvam is abandoning foundational-model research.

Hemant Mohapatra, partner at Lightspeed India, which invests in Sarvam, said global AI companies, including Chinese model developers, already generate significant revenue by running their models on GPU infrastructure hosted in India. He considered Token Factory a timely commercial response to the global GPU shortage.

Mohapatra's conclusion was compact: “Just because a company launched product-x, doesn't mean that's where the launches are going to stop.” One inference product can sit alongside continued model development, he said.

What does Token Factory change for enterprises?

For enterprise buyers, Token Factory replaces a hardware-and-compute build with one domestic API and published token rates. It adds local custody and 12 model entries, but Pahwa found DeepSeek V4.1 Flash at parity with OpenCode and Command Code. The switch case is local data control, rupee billing or access to Sarvam's own models, not headline token savings.

Token Factory is commercially live and operationally Indian, on Sarvam's account. The sovereign label remains contested; arithmetic, not adjectives, should settle it.

via medianama.com (Source)

Filed under

  • sarvam
  • ai-inference
  • gpu-infrastructure
  • sovereign-ai
  • model-hosting
Share this article:

More from Tom Whitfield

Tom Whitfield

Show full bio

Market editor covering people, launches and day-to-day practice in enterprise information technology and the leaders who run at CIO Monitor.

185 articles

The newsroom digest

One email a day with the stories that matter. No noise.

By subscribing you agree to the privacy policy.

Cross-references

‹ Previous article