When frontier AI can be switched off: India's sovereignty challenge

The world can't base its AI strategy on the assumption that the world's most advanced models will remain globally available, commercially accessible and continuously usable. This is the signal sent by the US government's restrictions on Anthropic's Fable 5 and Mythos 5, which make clear that any country can be denied access to frontier intelligence at any given moment.

The US uses export controls to slow the diffusion of strategic technologies, such as encryption source code, advanced chips, chip design software and semiconductor equipment. The Anthropic restrictions offer an early glimpse of a potential future in which the next layer, controlled by export directives, could become frontier intelligence. So, India and others cannot rely on partnerships with foundational AI tech companies alone, since they are not sovereign players.

India had earlier requested Anthropic for access to Mythos under Project Glasswing, to understand capabilities of frontier AI models for cybersecurity across several sectors like banking and telecom. That access has now been disrupted.


US export control directives are not unpredictable moves. In fact, other countries should have seen this coming. Historically, technology restrictions have shaped global information flows. Anthropic operates inside the US state. So, if White House decides a frontier model must be restricted, even the largest AI lab will have to comply. It doesn't matter that India is Anthropic's second-largest consumer base, since being a customer is not the same as having control.

History shows India having faced such denial before. But it led to a pivot in its own pathway, not a change in its destination. India's nuclear programme was built over decades, and its strategic capabilities developed in a world where external access could not be assumed. But it can't be compared with AI. Nuclear capabilities were anchored in physical capabilities and reached strategic thresholds, while frontier AI development is evolving.

But governing AI is like tackling nuclear technology. AI is politically sensitive, strategically consequential and difficult to govern through certainty alone. With AI models improving rapidly, chips evolving, and inference costs and applications changing across the economy, frontier intelligence will proliferate with or without India's quest for AI sovereignty.

Some argue that India can respond by diagnosing it as a competitiveness problem. The issue is not only that it has a compute dependency and hasn't developed foundational models, labs or research ecosystems. The deeper problem is that India lacks an AI continuity doctrine to preserve its agency and continuity in the world, as frontier intelligence is becoming a more gated and controlled strategic layer.

That's why India cannot answer this challenge through a 'checklist for AI sovereignty' - buying GPUs, building data centres, funding a few startups and announcing some foundational models are being built. It should not expect legacy IT to build frontier intelligence labs. These require research depth, not thin R&D spending, a willingness to fail, and commitment to building for India and the world. Indian IT services model has never operated on these aspects, having instead built its legacy on labour arbitrage.

IndiaAI Mission's compute capacity and foundational model push are important beginnings. Sarvam's progress in sovereign AI, compute models and deployment is encouraging. But a country of India's scale needs many competing labs, not 1-2 symbolic winners.

Anthropic emerged from OpenAI. India needs conditions for 10 Sarvams. It needs to start building institutions and conditions that can support an AI continuity doctrine now.

One answer lies in Singapore's biomedical journey. The entrepot economy was transformed into a global biotech hub over several decades. It did for biotech what India needs to do for frontier AI: build institutions before the ecosystem is ready. Through Agency for Science, Technology and Research (A*STAR) and Biopolis, Singapore not only recruited global talent, built research autonomy and gave institutional backing to MIT's Jackie Ying to lead its Institute of Bioengineering and Nanotechnology, but also saw its success compound over the years.

The US may ease controls on frontier models in the future to protect interests of its AI ecosystem again. But can we afford to wait for that, given that we cannot stop relying on current frontier intelligence, while those institutions can't wait in perpetuity to be developed if India doesn't want to be the one billed for importing frontier intelligence?

(Kapoor is chair, and Zutshi is research manager, Institute for Competitiveness)