India AI policy: Pragmatism before sovereignty

Sovereignty has long been central to Indian AI policy discourse, and most experts contend that the country must achieve some degree of AI sovereignty. Such calls intensified after the US banned foreign access to Claude Mythos. While reliance on foreign AI systems is not ideal, the ground reality suggests that AI sovereignty has become harder to achieve than it was a year or two ago. India’s way forward thus lies in pragmatism, which would require tolerating technological dependence, at least for the foreseeable future.

The current AI landscape is largely dominated by the US and China. While the US boasts frontier models such as Claude, ChatGPT and Gemini, China is rapidly catching up with models such as DeepSeek, GLM and Kimi K3. India, on the other hand, is yet to develop a comparable model. Its most prominent model, Sarvam, is only a few hundred billion parameters, considerably smaller than frontier models with several trillion. Even China’s open-source Kimi K3 has 2.9 trillion parameters.

The reasons for India’s lag in the AI race are largely known. Limited research spending, compute access, semiconductor capacity, risk-averse capital markets, environmental constraints and foreign competition help explain India’s limited progress. While the government is investing through Semiconductor Mission 2.0, the Anusandhan National Research Foundation, expanded compute capacity and skilling, matching US and Chinese investment in these areas will take time. Given these realities, this year's Economic Survey acknowledged that India should not pursue the costly path of frontier AI. Instead, the nation should focus on developing smaller models suited to its local context.


Also read | Before the first server hums: India needs an impact test for the AI age

Although structural constraints make AI sovereignty difficult for India, a range of external factors have made it an even steeper hill to climb.

Among these external factors, the primary ones come from the US. Firstly, American models are becoming more expensive as frontier firms have expanded their user base slightly faster than their computing capacity, pushing up AI token prices. For Indian firms that build AI wrappers and applications using foreign foundation models, rising token prices can cause supply-chain shocks. While this does not mean India can no longer operate at the AI application layer, the fact that this layer is already vulnerable to upstream shocks makes it an increasingly uncertain alternative for countries that could not develop sovereign models and instead seek sovereignty at the application layer. Secondly, the US is increasingly prioritising its own interests over its conventional policy of globalisation. Its decision to deny foreign access to its most advanced model makes it more difficult for nations seeking to build AI applications on foreign models.

The other set of external factors comes from China. The country is rapidly undercutting competition with cheap models that deliver near-benchmark performance at a fraction of American models’ prices. More significantly, these models are often open source, allowing other nations to fine-tune them to their local context. This enables China to build an AI ecosystem through cheap model distribution and then charging for the services required to run these models. For instance, Alibaba uses its Qwen AI model to draw users towards its cloud computing platform. For Indian firms that already struggle to compete with American firms, the presence of Chinese models intensifies competition further.

Also read | AI could become a new revenue engine for Indian telecom companies: Survey

China’s actions are important for another reason. By undercutting competition and pursuing global penetration, China is effectively challenging USA’s claim to AI leadership. The US has responded with harsh export restrictions on semiconductors to China and has also pressed its allies to adopt similar measures. NVIDIA, under American pressure, has even prepared a whitelist of customers in Asia to prevent the diversion of chips to China through intermediary countries. China, in response, has created the World Artificial Intelligence Cooperation Organization (WAICO), potentially laying the foundations for a China-centred AI ecosystem. By drawing developing countries into the AI race, the two AI superpowers could ultimately create competing AI blocs, with nations lacking AI sovereignty aligning themselves with one or the other.

These developments matter because countries like India still need access to foreign technology to build their own ecosystems, while growing AI geopolitics could make such access conditional on political alignment. USA’s actions also send a clear message that challenging its models could invite strong reactions.

This ultimately means that India needs to pursue strategic reliance rather than meaningless dependence. This will require not only pragmatism but also careful diplomacy to ensure that the country’s interests remain paramount. India must take these steps before foreign footprints deepen and rigid AI blocs emerge.

A pragmatic approach would require the country to accept foreign dependence, but on its own terms. This would mean adopting expensive American models where firms are too small to develop or fine-tune their own models, while larger firms with the required capabilities could leverage open-source Chinese models and fine-tune them for local needs. Geopolitical tensions with China may therefore have to be set aside where doing so serves India’s broader technological interests.

Even when foreign models are used, India must maximise value creation from application-based AI when using closed-source models, while open-source models should serve as pathways towards stronger domestic fine-tuning.

Strategic alliances will be equally important. China’s WAICO is arguably the most significant non-American AI alliance, yet India is not currently part of it. Policymakers should therefore seek to position India within such emerging arrangements, enabling access to joint AI ecosystems. Foreign partnerships could strengthen domestic capabilities as well. For instance, emerging sovereign models such as Sarvam could benefit from collaborations with firms like Anthropic and OpenAI through shared compute resources, technical expertise and knowledge transfer.

At the same time, domestic capacity in compute and research must be strengthened, as pragmatic use of foreign technology cannot substitute for technological self-reliance. For now, strategic dependence may be an undesirable but necessary concession. India should therefore treat this dependence as temporary: a small compromise for the greater goal of AI sovereignty. However, the country must never forget that what is a necessity today should not become a constraint on India’s technological ambitions tomorrow.


(Amit Kapoor is chair & Mohammad Saad is researcher at Institute for Competitiveness.).