Govt's AI strategy problem: Why sovereign models may not be the answer

India should focus on AI talent, innovation and global integration rather than spending public funds on building a sovereign large language model, argue the authors

Anthropic, an American artificial-intelligence (AI) company, was recently forced by the United States (US) government to keep access to its latest model Fable away from non-Americans. Some in India are now demanding a “sovereign” AI model. Does India need its own large-language model (LLM)? Should government funds be allocated for this cause? We are sceptical; we think such proposals are just industrial policy. 

The instinctual panic is understandable. Nobody wants to be locked out of the technology race. But a gap between Indian knowledge and the global frontier is hardly new. Most great milestones of knowledge — the transistor, the Internet or Unix — were not invented here. The Indian Tejas uses an American jet engine. With infusions of public money, some low-end hardware work on semiconductors has started in India, which we expect will induce the usual industrial policy outcome. 

Indian firms are world-beaters in information-technology (IT) services. But blaming Indian IT services for not investing in LLMs is like chastising IndiGo for not manufacturing jet engines. Indian IT-services companies have trained over a million Indians on using the technology invented in the West to serve customers worldwide. They managed this without being on the global knowledge frontier. At every step of the miracle of great Indian services exports, there was the danger of nationalist or industrial policy demands for sovereign central processing units, a sovereign operating system, or sovereign hard disks. Indian policymakers did globalisation correctly in that period: Indian IT services firms imported Western technology, exported software and services, and generated an economic miracle for India. 

Foreign export controls are not new, either. The US blocked the sale of a Cray supercomputer for weather forecasting in the late 1980s. In the 1990s, they treated strong encryption as a weapon and hauled up the author of PGP (“Pretty Good Privacy”) through a criminal investigation. In 1999, it reclassified commercial communication satellites as munitions. The same mechanism now caps the sale of advanced graphics processing units (GPUs) to China (something that we in India should be grateful for). None of this interfered in our objective in India of achieving high economic growth. 

What is new with LLMs is that access to a novel technology has been put into the hands of normal citizens. Millions have access to the latest technology, almost immediately along with their global counterparts. This feels exciting. A privately funded AI revolution seeking customers globally has led to a new generation of enthusiasts. This has helped create more noise in this field than with (say) US export controls on CNC (computer numerical controlled) lathes. 

Indian firms will not be harmed by not having access to frontier models. These models are expensive, burning thousands of dollars of tokens in a few hours. One of us is building TheProfesseer, where LLMs are used to provide litigation analytics, which requires processing millions of Indian court orders. The scale demands cost efficiency, through using older models, open-source models, etc. Whether we think about serving foreign customers, or building in India, the latest models are not the bottleneck. 

The defence argument favouring “sovereign AI” is weak. We buy most of our defence equipment. It is possible to say: We want a military drone where every single component is made in India. The cost would be prohibitive, and there is a high risk of such military drones losing battles to Chinese rivals. It makes far more sense to collaborate with our allies — Europe, Japan, South Korea, and Taiwan — who have the exact same objective (military drones that are fully safe from the Chinese backdoor or supply chain). We pursue the Indian interest better through the tools of compromise, deal-making and alliances, rather than doing sovereign AI. 

The triumph of the US in AI did not happen through even a hint of sovereign AI. The expression “sovereign AI” is used only by people who don’t do AI innovation. The greatness of the US is that Anthropic, Google and OpenAI are just private companies that innovated on their own steam, backed by the world’s greatest financial system. These three firms emerged as winners in a race where 1,000 firms tried to compete, and 997 of them failed. What worked in the US were the financial system and the innovation system, not sovereign AI. 

Industrial policy is unable to engage in such a process of discovery. It will never match the energy and risk taking of private people. It will translate coercive and financial inputs into poor outcomes owing to low state capability in India, and will be hijacked by domestic political economy. And, it is irrelevant owing to the tiny resource envelope of the Indian state. 

What then should the Indian state do in the field of AI? We suggest a National AI Leadership Policy comprising four elements: 

1. The contribution of the Indian state for the IT miracle was in human capital building. At sites like the Indian Institutes of Technology, NCST, Ernet, the Indian Institute of Science, etc, the Indian state invested in hundreds of researchers. They were the seed corn of the people who built the Indian IT miracle. We should invest in such human capital initiatives. Our thinking on innovation policy has since improved: We now know how to better translate public money into gains for the country. Ragunath Mashelkar, Ajay Shah and Susan Thomas (https://bit.ly/xkdr_MST_2024) have proposed innovation policy organised around public money sent into private universities and private firms (as opposed to state organisations only). 

2. A full review of frictions in buying IT equipment and overseas services is required, so that it becomes effortless for persons in India to plug into the world. We need full convertibility on the current account. 

3. Finance is the brain of the economy. Financial-sector reforms are required to drive risk-taking by private firms which will then find their place in the global AI supply chain. We need full convertibility on the capital account, so that the vast resources and knowledge of the global financial system reshape how Indian firms think about business strategy in the AI age. 

4. Partner our allies to get world-class defence equipment that is untainted by China. 

---

The authors are, respectively, a researcher at XKDR Forum and chief technology officer, TheProfesseer