Article complete

Sarvam’s Token Factory Sparks India’s Sovereign AI Debate, Again

Sarvam AI’s latest product launch was supposed to be a milestone for India's AI infrastructure ambitions. Instead, it has reopened a more uncomfortable question: what exactly does India mean when it talks about sovereign AI?

sarvam

Key takeaways

  • Sarvam AI's Token Factory launch has reignited discussions on India's definition of sovereign AI.
  • The platform offers access to multiple AI models, but most are developed by Chinese companies.
  • Sarvam's reliance on foreign models raises questions about its role in India's indigenous AI ambitions.
  • The initiative aims to simplify access to AI models while keeping data processing within India.
  • Government support for Sarvam comes with high expectations for contributing to local AI development.

Takeaways drafted with AI from the article text.

Sarvam AI seems to have found itself in a familiar position. Every new move by the company seems to reopen a larger debate about India's AI ambitions: should domestic AI companies focus on building indigenous models or on making the world's best models accessible through Indian infrastructure?

Its latest launch, Token Factory, has once again exposed the conflicting expectations surrounding India's most prominent AI startup. This time, the debate is not about whether Sarvam has built a genuinely indigenous foundation model, but why a company backed by the IndiaAI Mission is hosting models developed by Chinese AI companies.

On November 16, Sarvam launched Token Factory, an inference platform that gives developers and enterprises access to eight AI models through a single OpenAI-compatible API. All the models are hosted on GPU infrastructure in India, allowing customers to run inference without routing requests through overseas providers.

The catalogue includes Sarvam’s own 105B model, alongside models from DeepSeek, Kimi, GLM and Google. Pricing starts at ₹14 per million input tokens for DeepSeek V4.1 Flash, with Sarvam also offering zero-data-retention options for enterprises.

The proposition is commercially straightforward. Instead of building and maintaining expensive GPU infrastructure, developers can access several open-weight models through one provider, with their data remaining in India.

But six of the eight models listed are Chinese, while only one is developed by Sarvam. For a startup selected under the IndiaAI Mission to build an indigenous foundation model, the optics are difficult to ignore.

Sarvam received a ₹98.68 crore GPU subsidy for 4,096 NVIDIA H100 GPUs and raised $234 million at a $1.5 billion valuation in June. The government-backed support came with expectations that the company would contribute to India's own foundation model capabilities.

However, government support for developing indigenous AI capabilities does not necessarily preclude Sarvam from pursuing other commercial opportunities. The larger question is whether its infrastructure business complements its foundation model ambitions or eventually takes precedence over them.

The criticism also comes with a history.

In May 2025, Sarvam faced considerable backlash after releasing Sarvam-M, a 24-billion-parameter open-weight hybrid language model built on top of Mistral Small. The model supported 10 Indian languages, including Hindi, Bengali, Gujarati, Kannada and Malayalam.

Despite its positioning as an important step for Indic AI research, Sarvam-M received just 334 downloads on Hugging Face in its first two days. Investor Deedy Das called the numbers "embarrassing", arguing that there was little audience for incremental work. He compared it with an open-source model developed by two Korean college students that had attracted around 2,00,000 downloads.

The criticism extended beyond Sarvam. BharatGen's Param-1, another government-backed foundation model built from scratch, had received just 12 downloads on AIKosh at the time.

The episode exposed a difficult problem for India's AI ambitions. Domestic foundation models were expected to compete with global offerings, but the resources, research capabilities and developer ecosystems needed to do so were still developing. At the same time, even incremental progress was being judged against the likes of OpenAI and DeepSeek.

Now, with Token Factory, the criticism has taken a different direction. Instead of questioning why Sarvam built on top of a foreign model, people are questioning why it is hosting foreign models at all.

Hemant Mohapatra, an investor in Sarvam from Lightspeed Ventures, pushed back against the criticism on X.

"Really don't understand why launching a token-factory product is getting so much criticism thrown" at Sarvam, he wrote, acknowledging that he was biased as an investor. He argued that people were underestimating how much global inference providers were making by hosting global, including Chinese, models in India.

His argument points to a part of the AI business that often gets overshadowed by foundation model development. Training a model is only one part of the equation. Running it efficiently, reliably and at a competitive price is an entirely different business.

Every time a developer or enterprise sends a request to a model, somebody has to provide the computing power to process it. As AI agents and enterprise applications scale, inference could become an important revenue opportunity for companies that can provide the infrastructure at competitive prices.

Really don't understand why launching a token-factory product is getting so much criticism thrown

Hemant Mohapatra

Sarvam is attempting to capture that market by bringing multiple models under one platform while retaining control over where the computation happens.

However, it is not the only Indian company pursuing this opportunity. JarvisLabs launched its own Token Factory around two months before Sarvam's announcement and has been working on improving its inference capabilities.

The company has since begun comparing its performance against Sarvam's.

Atharva Ingle, associated with JarvisLabs, said it benchmarked three open-weight models across both platforms and claimed JarvisLabs was on par with or faster than Sarvam in speed and latency.

These are, however, claims made by a competing provider, and performance benchmarks alone do not establish which platform offers a stronger overall proposition. Pricing, throughput, reliability and enterprise support are equally important considerations.

This introduces another uncomfortable question for Sarvam. If smaller Indian infrastructure providers can deliver comparable or better inference performance, how much of the advantage comes from its technology, and how much comes from the capital and government support available to it?

There is, however, a distinction that much of the criticism appears to overlook. Hosting a Chinese model on Indian GPUs does not make the model Indian. But it allows Indian enterprises to access that model through domestic infrastructure, potentially keeping sensitive data within the country's borders and reducing dependence on overseas inference providers.

Sovereign AI does not necessarily mean developing every model, chip and infrastructure component domestically. A country can also build sovereignty by controlling where its data is processed, how its computing resources are deployed and who provides the services.

But that raises a bigger question about what India is actually trying to achieve with its AI investments.

If Sarvam can build a profitable inference business by hosting the world’s best models, it could create a sustainable domestic AI infrastructure company. That infrastructure could eventually support its own models and those developed by other Indian researchers.

Yet if the company becomes a platform that primarily monetises access to foreign-developed models, the original ambition of building India's own foundation models risks becoming secondary.

India has already struggled to generate meaningful adoption for some of its early domestic models. At the same time, its AI infrastructure companies are competing in a market where access to the best models, cheap computing and reliable inference matters more immediately to customers than the nationality of the model developer.

Ultimately, Sarvam’s Token Factory may be a sensible business decision. But it also raises a question that extends beyond the company: should a company receiving public support to build India’s AI capabilities be judged primarily by its commercial success, or by how much original AI technology it contributes to the country?

Published

Spotted an error? Tell our editors. Read our editorial standards and corrections.

More in this story

AI in India · 5 stories
  1. Xebia Partners with NVIDIA, Anthropic to Scale Enterprise AI
  2. Cypher 2026 Day 3: Enterprises Build Their Own Models as Agent Control Takes Centre Stage
  3. The 20 Cypher 2026 Talks That Got People Talking
  4. SoftBank's AI Empire Rests on a Single Asset—And It's Over-Leveraged
  5. You are here: Sarvam’s Token Factory Sparks India’s Sovereign AI Debate, Again
About the author

Journalist

Mohit writes about AI in simple, explainable, and often funny words. He's especially passionate about chatting with those building AI for Bharat, with the occasional detour into AGI.

All stories
Belamy · Daily

The day’s AI news, before your first meeting.