SARVAM AI AND INDIA'S SOVEREIGN AI AMBITIONS

20 Nov 2026

Tags: Science & Technology   Emerging Tech   Information technology

Source: The Indian Express

Context: Sarvam AI, an Indian AI startup supported under the government's IndiaAI Mission, has launched its Token Factory platform, offering access to AI models developed by Chinese companies alongside its own models. This raises questions about the meaning of AI sovereignty, India's dependence on foreign technologies and the commercial viability of indigenous AI development.

Sarvam AI's Token Factory: Key Features

  • Launch: Sarvam AI introduced Token Factory as a platform for developers and enterprises to access multiple AI models through its computing infrastructure located in India.
  • AI inference platform: Enables businesses to run AI models without purchasing expensive computing hardware or independently managing infrastructure.
  • Models offered: Includes open-weight models developed by Chinese companies such as DeepSeek, Moonshot AI (Kimi), Z.ai (GLM) and Alibaba (Qwen), Google's Gemma models and Sarvam's proprietary models.
  • Sarvam's models: The platform also offers its indigenous Sarvam 105B model.
  • Deployment options: Some models are available on demand, while others require dedicated computing capacity.
  • Global business model: Similar inference services are offered by companies such as Together AI, Fireworks AI, DeepInfra and OpenRouter, which provide access to models developed by multiple organisations.

Why is Sarvam Offering Chinese AI Models?

  • Commercial opportunity: Chinese AI companies have developed competitive open-weight models that can be downloaded and deployed by other organisations, providing alternatives to proprietary models developed by OpenAI and Anthropic.
  • Revenue diversification: Hosting third-party models creates an additional revenue stream for Sarvam beyond developing and commercialising its own foundation models.
  • Indian data hosting: Running Chinese-developed models on computing infrastructure located in India allows businesses to access these models without necessarily sending their data to servers operated by Chinese companies.
  • GPU constraints: Demand for high-performance Graphics Processing Units (GPUs) has created computing-capacity constraints globally. Hosting foreign models on Indian infrastructure can attract international customers seeking access to available computing resources.
  • Cost considerations: Despite its commercial potential, Sarvam's platform may not necessarily provide the cheapest AI inference services available in the market.

Implications for India's Sovereign AI Mission

  • India's Indigenous AI Development Push
    • IndiaAI Mission: Approved with an outlay of ₹10,371.92 crore to strengthen India's AI ecosystem, including indigenous foundation models, computing infrastructure, datasets, skills and innovation.
    • Government support: Sarvam AI was selected under the mission to develop indigenous foundation models, receiving financial and computing support amounting to ₹246.72 crore, according to government disclosures cited in the article.
    • Core objective: Reduce dependence on foreign AI technologies and develop domestic capabilities across the AI technology stack.
    • Emerging contradiction: Sarvam's decision to offer Chinese AI models demonstrates that even companies developing indigenous models see commercial value in deploying foreign-developed technologies.
  • Debate over the Meaning of AI Sovereignty
    • Infrastructure sovereignty: Ensuring that AI models operate on computing infrastructure, data centres and servers located within India and subject to Indian jurisdiction.
    • Technological sovereignty: Developing domestic capabilities across the AI value chain, including foundation models, semiconductors, computing hardware, datasets and AI research.
    • Data sovereignty: Ensuring that sensitive Indian data is governed by Indian laws and appropriate data protection and security safeguards.
    • Key distinction: Hosting foreign-developed models in India improves control over infrastructure and data handling but does not eliminate dependence on foreign-developed model architectures, software or technological ecosystems.

Concerns and Criticism

  • Dependence on foreign technology: Reliance on Chinese and other foreign models may limit the development and adoption of indigenous foundation models over the long term.
  • Dilution of indigenous ambitions: Sarvam was viewed as India's potential answer to leading AI developers in the U.S. and China. Offering foreign models has raised concerns about whether indigenous AI development will receive sufficient commercial and institutional priority.
  • Security concerns: Local hosting does not automatically eliminate risks arising from model vulnerabilities, software dependencies, training data, model updates and supply-chain security.
  • Commercial vs. strategic objectives: While offering foreign models may improve commercial viability, it also raises questions about balancing short-term revenue generation with long-term technological self-reliance.
  • Industry perspective: Some investors and industry participants view the decision as a commercially rational response to global demand for AI inference capacity and shortages of GPUs. Others believe it may weaken the symbolic significance of Sarvam as an indigenous AI champion.

Additional Relevant Information for UPSC

Important AI Terminology

  • Foundation Models: Large AI models trained on extensive datasets that can be adapted for multiple downstream applications, such as text generation, translation and image recognition.
  • Open-weight Models: Models whose trained parameters (weights) are publicly accessible, allowing users to download, deploy and sometimes modify them, subject to their respective licences. Open weights do not necessarily mean that training data and source code are open.
  • AI Inference: The process through which a trained AI model generates predictions or responses using new inputs.
  • AI Training: The process of developing an AI model by exposing it to large datasets and adjusting its parameters to learn patterns.
  • GPU (Graphics Processing Unit): A specialised processor capable of performing parallel computations, widely used for training and running AI models.
  • AI Compute Infrastructure: The combination of GPUs, servers, data centres, networking and storage required to develop and operate AI systems.

IndiaAI Mission

  • Nodal Ministry: Ministry of Electronics and Information Technology (MeitY).
  • Approved outlay: ₹10,371.92 crore over five years.
  • Objectives: Develop indigenous AI capabilities, expand computing capacity, improve access to datasets, promote AI startups and develop skilled human resources.
  • Key components: IndiaAI Compute Capacity, IndiaAI Innovation Centre, IndiaAI Datasets Platform, IndiaAI Application Development Initiative, IndiaAI FutureSkills, IndiaAI Startup Financing and Safe and Trusted AI.

Way Forward

  • Strengthen indigenous foundation models: Increase investment in domestic AI research, training datasets and model development.
  • Develop domestic computing capacity: Expand access to high-performance GPUs, semiconductor technology and energy-efficient data centres.
  • Adopt a balanced approach: Allow Indian companies to commercially deploy global models while maintaining sustained investment in indigenous alternatives.
  • Ensure data and cybersecurity: Establish safeguards for data localisation, model security, software dependencies and supply-chain transparency.
  • Promote open innovation: Encourage collaboration among startups, universities, research institutions and industry.
  • Measure sovereignty across the AI value chain: Assess domestic capabilities in computing hardware, foundation models, datasets, talent and deployment infrastructure rather than relying only on the geographical location of servers.

Conclusion: AI sovereignty does not necessarily require complete technological isolation. India needs to balance commercial access to global AI technologies with the development of indigenous capabilities. The long-term objective should be to ensure that India is not merely a market or hosting destination for foreign AI models but also a competitive developer and innovator across the AI value chain.

Prelims Question

Q1. With reference to AI models and their deployment, consider the following statements:

  1. Open-weight AI models make their trained parameters publicly accessible, but this does not necessarily imply that their training datasets are also publicly available.
  2. AI inference refers to the process of training an AI model by adjusting its parameters using large datasets.
  3. A company can deploy a foreign-developed AI model on computing infrastructure located within India without having developed the model itself.
  4. GPUs are widely used in AI because they support parallel computations.

Which of the statements given above are correct?

(a) 1 and 3 only
(b) 1, 3 and 4 only
(c) 2 and 4 only
(d) 1, 2, 3 and 4

Answer: (b) 

Explanation:

  • Statement 1 is Correct: Open-weight models provide access to trained parameters (weights), but training data and source code need not be open.
  • Statement 2 is Incorrect: This describes AI training. Inference involves generating predictions or responses using a trained model.
  • Statement 3 is Correct: Hosting and developing an AI model are distinct activities. Foreign-developed models can be deployed on Indian servers.
  • Statement 4 is Correct: GPUs are suited to parallel processing, making them important for AI training and inference.