JupiteX Get the app
Science & Technology18 Nov 2026 · about 6 min

Sarvam’s move to host Chinese AI models triggers debate

The brief

One notable detail is that Sarvam’s India-based service includes models developed in China. Token Factory offers access to these alongside Sarvam’s own models and Google’s Veritas. That mix has prompted debate about whether sovereign AI means domestic hosting, domestic model creation, or both. Sarvam says enterprises will have more models to choose from and “an AI stack they can own without lock-ins”. Token Factory is an inference platform: developers use one application programming interface to access models rather than buying expensive hardware and managing each model themselves. Sarvam says its software and infrastructure compete on speed, reliability and cost. The launch positions Sarvam as both a model developer and an inference provider. Critics say a company selected under the IndiaAI Mission should focus on original Indian research, while supporters point to demand for locally hosted international models. Hosting a model in India does not, by itself, mean the model was developed in India.

01

What did Sarvam announce about the models available through its Token Factory?

One notable detail is that Sarvam’s India-based service includes models developed in China. Token Factory offers access to these alongside Sarvam’s own models and Google’s Veritas. That mix has prompted debate about whether sovereign AI means domestic hosting, domestic model creation, or both.

Sarvam says enterprises will have more models to choose from and “an AI stack they can own without lock-ins”. Token Factory is an inference platform: developers use one application programming interface to access models rather than buying expensive hardware and managing each model themselves. Sarvam says its software and infrastructure compete on speed, reliability and cost.

The launch positions Sarvam as both a model developer and an inference provider. Critics say a company selected under the IndiaAI Mission should focus on original Indian research, while supporters point to demand for locally hosted international models. Hosting a model in India does not, by itself, mean the model was developed in India.

02

What is an AI inference platform, and what does it let developers do?

An AI inference platform provides the computing infrastructure needed to run an already-trained model and produce responses. It matters because developers can use models without building and operating all the underlying systems themselves.

Token Factory illustrates the setup: rather than purchasing expensive hardware and managing each model separately, developers connect to several models through a single application programming interface. That interface is the access point; the platform supplies the infrastructure behind requests.

The platform does not, by itself, create or train the models it serves. Sarvam says its customised software and infrastructure help it compete on speed, reliability and cost. The article does not detail how Token Factory allocates computing resources. In general, inference means applying a trained model to a new input to produce an output. For developers, the practical benefit is a simpler way to access multiple models while the platform handles running them.

03

Which Chinese-developed models does Token Factory host, alongside Sarvam’s own models?

Token Factory’s selection includes three Chinese-developed models: Pramana, Moonshot AI and Z.ai. Their presence is notable because Sarvam is an Indian company operating the service on India-based infrastructure, and the launch has sparked discussion about Indian AI sovereignty.

These models are offered alongside Sarvam’s own models and Google’s Veritas. Token Factory acts as an inference provider, giving developers access to multiple models through one application programming interface. They can use the service without purchasing expensive hardware and managing each model separately.

Sarvam says the wider selection gives enterprises more choice and an AI stack they can own without lock-ins. Critics argue that a company selected for indigenous model development should prioritise original Indian research. Supporters say locally hosting international models can meet commercial demand. The list of hosted models illustrates the distinction at the heart of the debate: where a model runs and where it was developed are different questions.

04

How much public support did Sarvam receive for indigenous model development, including subsidised computing power?

Sarvam was selected in 2025 for indigenous model development under the IndiaAI Mission. The support reported for the company was subsidised access to 4,096 Nvidia H100 GPUs, valued at Rs 98.68 crore. This computing support is distinct from the mission’s total public outlay.

The government approved the IndiaAI Mission in March 2024 with approximately Rs 10,372 crore to strengthen domestic computing infrastructure, research and foundational model development. Sarvam’s GPU access was intended for its indigenous model work. The article describes it as subsidised access, not as a cash grant of the same amount.

The public support is part of the debate over Sarvam’s priorities. Critics say a company chosen for homegrown model development should deliver original Indian research. Sarvam has also raised over $300 million at a reported $1.5 billion valuation, but that is separate from the public support figures. The article does not provide a further breakdown of Sarvam’s public funding.

05

How is hosting a trained AI model different from developing an original model?

Hosting and developing AI describe different work. Hosting means providing the infrastructure that runs a trained model and generates answers. Developing an original model means doing the work to create the model. The distinction matters in the debate over whether India’s sovereign AI ambitions require domestic infrastructure, homegrown models, or both.

Token Factory hosts models from Pramana, Moonshot AI and Z.ai on Sarvam’s India-based infrastructure. Developers can access them through one application programming interface. Sarvam also has its own models, and it was selected under the IndiaAI Mission for indigenous model development, with subsidised access to 4,096 Nvidia H100 GPUs.

The platform’s hosting service can be commercially useful without itself counting as original Indian research. Critics say Sarvam should prioritise frontier work because it received public support for indigenous model development. Supporters argue that enterprises want foreign models hosted locally. The article’s central distinction is that hosting foreign models and developing indigenous ones are different achievements, though they need not be mutually exclusive.

06

How can running foreign AI models on Indian infrastructure affect data residency, access, and dependence on overseas services?

Hosting a foreign model on Indian infrastructure changes where the service is run, not where the model was developed. It can give local enterprises access to international models through a service based in India. That distinction matters as India seeks to reduce dependence on overseas AI technology.

Token Factory offers Chinese-developed models and Google’s Veritas through an India-based platform. Sarvam says enterprises can choose among models and avoid lock-ins. At the Bengaluru Tech Summit, co-founder Vivek Raghavan warned that dependence on foreign AI systems could mean “exporting data and importing intelligence.” His remarks were not directly about Token Factory, but reflect the strategic concern.

Local hosting may provide access to models without using an overseas inference provider, but the article does not establish where all user data is stored or how it is handled. Nor does hosting make the underlying models indigenous. The broader implication is that local infrastructure and homegrown model development address different parts of AI dependence; the article presents both as relevant to sovereignty.

07

What is computing power, and why do AI models need it to generate answers?

Computing power means the processing capacity a computer system can use to carry out work. AI models need it to run their calculations on a prompt and produce a response. The article describes inference as requiring computing infrastructure, though it does not explain the technical calculations involved.

Token Factory supplies infrastructure so developers can run trained models without buying expensive hardware themselves. Sarvam says its customised software and infrastructure help it compete on speed, reliability and cost. For its indigenous model development, the company received subsidised access to 4,096 Nvidia H100 GPUs.

The article links computing capacity to India’s broader AI plans. The IndiaAI Mission, approved in March 2024, aims to strengthen domestic computing infrastructure, research and foundational model development. In general, GPUs are computing hardware used for AI workloads; that detail is established technical background, not explained in the article. Access to compute supports running models, but it is distinct from developing an original model.

This brief was written by AI from the original reporting and checked by other models. Names, figures and quotes come from the source; read it for full context.

Read more in the JupiteX app

Pulse is free. New stories every 4 hours, each one broken into the questions that explain it.

Or read more news on the web