Indian artificial intelligence startup Sarvam AI is facing criticism after adding Chinese-developed AI models to its newly launched Token Factory platform, reopening debate over India’s ambition to build sovereign artificial intelligence capabilities. The Bengaluru-based company is offering models from Chinese developers including DeepSeek, Moonshot AI, Z.ai and Alibaba alongside its own models and Google’s Gemma through computing infrastructure located in India.
The controversy is significant because Sarvam has been positioned as one of India’s leading attempts to develop homegrown foundation models and reduce dependence on foreign AI technology. The company was selected under the government’s IndiaAI Mission to develop indigenous AI models with public support. However, Sarvam says its Token Factory is a separate commercial inference business that allows developers and enterprises to access multiple AI models using computing infrastructure privately procured by the company. The debate now centres on whether India’s AI sovereignty should prioritise domestically developed models, Indian-controlled computing infrastructure or both.
Key takeaways
- Sarvam AI’s Token Factory now offers Chinese-developed models from DeepSeek, Moonshot AI, Z.ai and Alibaba.
- The platform also provides access to Sarvam’s own models and Google’s Gemma.
- The service allows developers and enterprises to run AI models through computing infrastructure located in India.
- Sarvam says Token Factory operates on privately procured GPUs and does not use government-subsidised computing capacity under the IndiaAI Mission.
- Critics argue that hosting Chinese models conflicts with expectations surrounding India’s efforts to develop indigenous AI technology.
- Supporters say offering international models is a commercially sensible way to build an AI inference business and generate revenue.
- Sarvam says its platform is already processing more than 450 billion tokens daily through DeepSeek V4.1 Flash.
- The controversy highlights the difference between developing AI models domestically and hosting foreign-developed models on Indian infrastructure.
Why Sarvam is facing backlash over Chinese AI models
The controversy began after Sarvam introduced Token Factory on November 16, 2026.
The platform gives developers and businesses access to several AI models through a common infrastructure and application programming interface (API).
Its catalogue includes Chinese-developed models such as DeepSeek V4.1 Flash, Kimi K3.1 Pro, GLM 5.3 and Alibaba’s Qwen models. It also includes Sarvam’s own Sarvam 105B model and Google’s Gemma.
The decision attracted criticism because Sarvam has been closely associated with India’s efforts to build an independent AI ecosystem.
The company was selected under the IndiaAI Mission’s indigenous foundation model programme, which aims to develop AI models suited to India’s languages, businesses and public needs.
Some technology observers questioned why a company receiving government support to develop Indian AI capabilities was simultaneously building a commercial platform around Chinese-developed models.
The criticism became more visible because Sarvam co-founder Vivek Raghavan had spoken about the risks of India’s dependence on foreign AI technology just a day before the Token Factory controversy gained attention.
At the Bengaluru Tech Summit on November 17, Raghavan warned that relying on foreign AI models could make India dependent on technology developed elsewhere.
He argued that when Indians use overseas AI systems, their data can leave the country and contribute to the development of models outside India.
His remarks intensified the debate around the company’s decision to offer Chinese models through its own platform.
However, the two activities are not necessarily contradictory from a business perspective. Developing an original AI model and providing infrastructure to run other companies’ models are separate commercial activities.
What is Sarvam’s Token Factory?
Token Factory is an AI inference platform designed for developers and enterprises.
Inference is the process of running an already-trained AI model to generate an output.
For example, when a user asks a chatbot a question, requests a summary or generates code, the AI model performs inference to produce the response.
Running AI models at scale requires expensive GPUs, specialised software, data-centre infrastructure and engineering expertise.
Token Factory allows businesses to access different models without purchasing and maintaining their own GPU clusters.
Instead of managing separate infrastructure for every model, developers can use a common API to connect their applications to models hosted by Sarvam.
The platform is designed to provide access to different AI models through computing infrastructure located in India.
This can be useful for businesses that want to experiment with different models, compare performance or select the most suitable option for particular tasks.
For example, one model may perform better at coding, another at reasoning, while a third may offer lower inference costs for routine applications.
By offering multiple models, Sarvam can serve customers with different requirements rather than relying exclusively on its own foundation models.
Which AI models are available?
The reported Token Factory catalogue includes:
| Model developer | Model or family | Country or origin |
|---|---|---|
| Sarvam AI | Sarvam 105B | India |
| DeepSeek | DeepSeek V4.1 Flash | China |
| Moonshot AI | Kimi K3.1 Pro | China |
| Z.ai | GLM 5.3 | China |
| Alibaba | Qwen 3.8 27B | China |
| Gemma 4 31B | United States |
Availability and deployment configurations can differ by model. Some models are available on demand, while others require dedicated computing capacity.
The platform’s commercial proposition is to provide customers with a range of models through one infrastructure provider.
Sarvam says its Token Factory does not use subsidised government GPUs
One of the biggest questions raised by the controversy concerns the relationship between Sarvam’s commercial platform and the public support it received under the IndiaAI Mission.
Sarvam has responded by clarifying that Token Factory runs on GPUs privately procured by the company.
The company said the platform does not use computing capacity subsidised under the IndiaAI Mission.
This distinction matters because the government programme was established to support the development of India’s domestic AI capabilities.
Sarvam was selected under the programme to develop indigenous foundation models and received financial and computing support for that work.
The government has disclosed assistance amounting to approximately ₹246.72 crore for Sarvam’s model-development efforts, according to reporting based on official disclosures.
Sarvam’s position is that its Token Factory is a separate commercial operation using privately funded infrastructure.
In other words, the company is distinguishing between its publicly supported domestic model-development programme and its privately financed inference business.
The clarification does not end the wider debate about what a company supported under a sovereign AI programme should prioritise. However, it addresses the specific concern that subsidised government GPU capacity is being used to host Chinese models.
Why Sarvam wants to host Chinese AI models
There is a commercial reason behind the decision.
Chinese AI developers have become significant competitors in the global open-weight model ecosystem.
Open-weight models allow developers to access model parameters and deploy the technology on their own infrastructure, subject to the relevant licence conditions.
This makes them different from many proprietary AI services that can only be accessed through the original developer’s hosted platform.
Chinese companies including DeepSeek, Alibaba, Moonshot AI and Z.ai have developed models that attract international developers because of their capabilities, flexibility and potential cost advantages.
For an inference provider such as Sarvam, hosting these models creates an opportunity to attract customers who want access to advanced AI capabilities without managing their own infrastructure.
The company does not need to develop a competing foundation model for every possible use case.
Instead, it can provide the computing resources required to operate models developed by different companies.
This is an established business model in the AI infrastructure market.
Companies such as Together AI, Fireworks AI and DeepInfra offer similar services by hosting and serving models developed by other organisations.
Amazon’s AWS Bedrock platform also gives enterprises access to models from multiple developers through a common cloud service.
Sarvam is attempting to build a comparable offering with infrastructure located in India.
Sarvam claims more than 450 billion daily tokens on DeepSeek
The scale of usage is another important part of Sarvam’s commercial argument.
The company has claimed that it is already serving more than 450 billion tokens daily through DeepSeek V4.1 Flash.
A token is a small unit of text processed by an AI model. It may represent a word, part of a word or a punctuation mark.
AI inference providers commonly charge customers based on the number of input and output tokens processed.
A platform serving hundreds of billions of tokens daily could therefore operate at significant scale, although token volume alone does not establish revenue or profitability.
The economics of inference depend on several factors:
- GPU acquisition and rental costs.
- Electricity and cooling expenses.
- Hardware utilisation.
- Model efficiency.
- Customer pricing.
- Network and storage costs.
- The balance between input and output tokens.
For Sarvam, the commercial opportunity is to attract enough recurring demand to justify the cost of operating large GPU clusters.
Hosting popular international models can help generate customer traffic while the company continues developing its own models.
The contradiction in India’s sovereign AI ambitions
The controversy has exposed two different interpretations of AI sovereignty.
The first interpretation focuses on developing domestic foundation models, owning the underlying intellectual property and building advanced AI research capabilities within India.
Under this approach, India needs its own competitive models rather than relying on technology developed in the United States or China.
The second interpretation focuses on controlling the infrastructure through which AI is operated.
Under this approach, a country can improve its technological independence by ensuring that computing resources, data storage, deployment infrastructure and customer access remain under domestic control, even when the underlying models are developed abroad.
Sarvam’s Token Factory follows the second approach.
The platform hosts foreign-developed models using infrastructure located in India. This can give Indian enterprises access to AI capabilities without necessarily sending their requests to servers operated by the original model developers.
However, local hosting does not automatically eliminate every security concern. Software dependencies, model licences, security vulnerabilities, access controls and the way models process information still require evaluation.
The two approaches also do not have to be mutually exclusive.
An Indian company can develop its own foundation models while earning revenue by hosting international models.
The larger question is whether the commercial activity helps finance domestic research and builds infrastructure that can eventually support more competitive Indian models.
Technology industry divided over Sarvam’s decision
The response to Token Factory has been mixed.
Some developers and technology commentators argue that Sarvam should focus more heavily on building frontier AI models in India, particularly because of its association with the government’s sovereign AI programme.
Sanchit Monga, chief executive of AI infrastructure company Runanywhere AI, questioned whether hosting foreign models reflected the ambitions associated with Sarvam’s selection under the IndiaAI Mission.
Other critics said they had expected Sarvam to focus on original research and technology development rather than building an inference business around models developed elsewhere.
However, investors and other industry participants have defended the commercial logic.
Hemant Mohapatra, a partner at Lightspeed India, which is an investor in Sarvam, argued that international inference providers are already generating business by hosting global models, including Chinese models, on infrastructure located in India.
He suggested that demand for such services is significant, particularly as access to advanced computing hardware remains constrained in several markets.
Technology entrepreneur Nikhil Pahwa also argued that Sarvam is a commercial business and needs to generate revenue while serving its shareholders.
Another view, expressed by Egma AI co-founder Nischal Jain, acknowledged the commercial argument but highlighted the symbolic importance of Sarvam to India’s technology ecosystem.
He argued that the company had become an emotional rallying point for people hoping India would build a competitive alternative to US and Chinese AI companies.
The disagreement therefore goes beyond the choice of models on a commercial platform. It reflects different expectations of what a prominent Indian AI startup should deliver.
Competition and pricing could become another challenge
Apart from the sovereignty debate, Sarvam also faces competition in the AI inference market.
Hosting popular models does not automatically guarantee that a platform will attract customers.
Developers compare inference providers based on several factors:
| Competitive factor | Why it matters |
|---|---|
| Speed | Faster responses improve application performance |
| Latency | Lower delays matter for real-time AI applications |
| Pricing | Lower costs can attract developers running large workloads |
| Reliability | Businesses need consistent service availability |
| GPU efficiency | Better hardware utilisation can improve margins |
| Model availability | A wider catalogue gives developers more choices |
| Data controls | Enterprises may require specific hosting and security arrangements |
Atharva Ingle, an AI engineer at JarvisLabs, said his company had benchmarked several models available on both platforms.
He claimed that JarvisLabs delivered comparable or better speed at lower prices for the models tested, including DeepSeek V4.1 Flash and GLM 5.3 Flash.
These are competitor-reported comparisons rather than an independent comprehensive benchmark. Performance can vary with hardware, model configuration, workload, batch size and testing methodology.
Nevertheless, the comparison highlights an important challenge: Sarvam must compete on service quality and economics, not only on the location of its infrastructure.
What this means for India’s IndiaAI Mission
The IndiaAI Mission was approved in March 2024 with an outlay of approximately ₹10,372 crore.
Its objectives include strengthening domestic computing infrastructure, supporting AI research, developing indigenous foundation models and expanding AI applications across the country.
Sarvam is among the organisations selected to develop sovereign foundation models under the programme.
The government’s broader objective is to reduce dependence on overseas technology and create domestic capabilities that can serve India’s businesses, citizens and public institutions.
The programme has also supported other organisations and consortia working on indigenous AI models.
This makes the Sarvam controversy relevant beyond one startup.
It raises questions about how public support should be evaluated.
Should success be measured by the number of Indian foundation models developed? Should the government prioritise domestic ownership of AI infrastructure? Or should it also consider the commercial strength of companies that operate AI services using a combination of Indian and international models?
A commercially successful inference business could help an Indian company build computing infrastructure, attract customers and generate revenue.
At the same time, public investment in indigenous model development creates expectations that recipients will make measurable progress in original research and domestic technology capabilities.
The important distinction is between the commercial use of foreign models and the actual delivery of the domestic model-development objectives for which public support was provided.
The bigger picture
Sarvam’s Token Factory controversy highlights the complexity of building technological sovereignty in an industry where research, computing hardware, software and commercial deployment are closely connected.
India wants to develop its own advanced AI models, but it also needs commercially viable infrastructure that allows businesses to run AI systems at scale.
Hosting foreign models on Indian infrastructure can support data localisation, improve access to computing resources and create an additional revenue stream for domestic technology companies.
However, it does not by itself create an indigenous foundation model ecosystem.
The long-term value of Sarvam’s strategy will depend on whether it can build a sustainable inference business while continuing to develop original models that compete on performance, reliability and cost.
For India, the broader challenge is to ensure that commercial AI infrastructure and domestic research reinforce each other rather than becoming competing priorities.
Looking ahead
Sarvam will need to demonstrate that Token Factory can attract developers and enterprise customers while maintaining competitive pricing, performance and reliability. Its reported token volumes indicate substantial activity, but revenue, margins, customer retention and infrastructure utilisation will be more useful measures of commercial success. The company will also face continued scrutiny over how it separates its privately funded commercial operations from its government-supported model-development work.
The wider debate will continue as India attempts to build a domestic AI ecosystem while remaining connected to global open-weight models and computing technologies. The key test is not simply whether Indian companies host foreign AI models, but whether they can use commercial infrastructure and international technology access to strengthen their own research, intellectual property and competitive AI capabilities.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.



