Sarvam AI Faces Sovereignty Debate Over Chinese Models on Token Factory
Indian AI startup Sarvam is facing questions about its sovereign-AI positioning after adding Chinese-developed models to its Token Factory inference platform. The company says the service runs on privately procured GPUs in India and is intended to give local developers access to a wider choice of models.

Indian artificial intelligence startup Sarvam AI is facing fresh scrutiny over the meaning of sovereign AI after making Chinese-developed models available through its Token Factory platform. The service offers developers access to several open-weight systems through a single interface, alongside Sarvam's own models.
The catalogue includes models from DeepSeek, Moonshot AI and Zhipu AI, as well as Google's Gemma. The move has prompted debate because Sarvam has been presented as a leading Indian effort to build domestic AI capabilities, with support from the government's IndiaAI Mission.
What Token Factory offers
Token Factory is an inference platform: it runs AI models and makes them available to developers through an API, rather than requiring each customer to operate the models on their own infrastructure. Sarvam says the service hosts models on GPU clusters located in India and offers a common interface for accessing its own and third-party open-weight systems.
The platform is designed to give Indian businesses and developers a local option for running a range of models. That can simplify integration and allow teams to choose a model based on the task, cost and performance requirements.
Why the Chinese models prompted questions
The presence of DeepSeek, Kimi and GLM has drawn criticism from people who see a tension between offering foreign-developed models and the goal of building sovereign Indian AI. The debate is not simply about where a model is hosted. It also concerns who developed and controls the underlying model, how it is updated, what licence applies and what assurances customers can obtain about security and data handling.
Sarvam's co-founder Pratyush Kumar told Moneycontrol that Token Factory operates on GPUs privately procured by the company. Sarvam said its subsidised GPU access under the IndiaAI Mission had been used to train its own models and ended in March 2026. The company has said its central mission remains building its own full-stack AI technology.
Sarvam's case for local inference
The company argues that making open-weight models available from infrastructure in India gives local organisations another option for accessing advanced AI. Sarvam says customer requests are processed on GPU infrastructure located in the country, and has framed local hosting as a way to support data locality while keeping access to a broad model selection.
Local inference can address some data-residency concerns, but it does not by itself settle every question about sovereignty. Organisations may also need to assess model provenance, software supply chains, update processes, licensing, security testing and the terms under which model developers distribute their weights.
A wider question for India's AI strategy
The disagreement highlights a practical challenge for India's AI ecosystem. Building competitive domestic foundation models requires significant investment in computing, talent and research. At the same time, businesses want access to capable models now, and open-weight systems from around the world can offer a commercially useful way to meet that demand.
For a platform provider, serving third-party models can create infrastructure revenue and bring customers into its ecosystem. For users, a locally hosted service may offer benefits in latency, billing and data governance. But the origin and control of the underlying model remain separate considerations.
What businesses should examine
Companies considering a hosted model should look beyond the location of the servers. They can ask where model weights originated, who can modify or replace them, how updates are verified, whether prompts and outputs are retained, and what independent security assessments are available. Regulated organisations may also need to document how each model is approved for its intended use.
Sarvam's Token Factory has therefore become part of a larger discussion about what sovereign AI should mean in practice. The question is whether sovereignty is defined mainly by domestic infrastructure and control of data, or whether it also requires domestic ownership of the models themselves. India's AI industry is likely to keep debating that distinction as more local platforms combine home-grown and international systems.
Sources
Related stories

US Tightens AI Chip Restrictions, Raising Concerns for Alibaba and Tencent
New US export rules target remote access to advanced American computing chips through overseas data centres, closing a route Chinese technology firms used to obtain AI computing power. Alibaba says the measures could affect some global data-centre expansion plans, while Tencent warns that tighter access may slow future model development.

Google Brings Gemini Computer Use to Mac, Letting Its AI Handle Multi-Step Desktop Tasks
Google has added computer-use capabilities to its Gemini app for Mac, allowing the assistant to navigate selected desktop apps, local files and websites in a separate window. The English-language rollout is initially limited to the United States.

A leaner 5G connection moves closer to everyday sensors and wearables
Ericsson and Altair Semiconductor say they have completed the first eRedCap network connection, a step towards lower-cost, lower-power 5G devices for sensors, trackers and connected wearables.

Small Businesses Get an AI Campaign Assistant as InMobi Unveils Accelerate
InMobi has introduced a self-serve advertising platform that uses seven specialised AI agents to help small businesses and agencies plan, build and optimise campaigns across major ad networks from one dashboard. Advertisers retain approval control over campaign plans, creatives and budgets.
