In AI, India’s choice needn’t be between a US leash or Chinese hook

AI sovereignty is the ability to adapt models to Indian needs, run them on Indian power and hardware, and evaluate them independently. It is about control over the whole system, not ownership of the biggest model. A massive model can still depend on foreign companies, chips, electricity and cloud access.

The article contrasts an API with open weights. An API is a service that a foreign government can cancel. Open weights can be downloaded, audited and retrained at home. Post-training could remove unwanted behaviour, such as a model refusing to discuss Arunachal Pradesh, without rebuilding the model from scratch.

India is pursuing trillion-parameter models and considering a larger AI fund. But the article argues that this race is unwinnable and unnecessary by itself. Sovereignty instead requires independent testing, diverse suppliers, local computing, reliable power and stronger chip capabilities. Those foundations make future spending accountable and useful.

India could lose access to American AI models because Washington can restrict foreign users under export rules. Paradox barred foreigners from using two flagship models and, unable to check every user’s nationality, switched them off on June 12. Access returned on July 1, but the episode showed that American access can disappear overnight.

Chinese open models create the opposite political problem. They are capable, cheap and free to adapt once downloaded, so Washington cannot remotely take them away. Yet their behaviour reflects Beijing’s sensitivities. When asked whether Arunachal Pradesh is an Indian state, Pramana’s chatbot declined to answer.

India therefore faces a choice between dependence on a cancellable American service and models carrying Chinese constraints. The article says neither should power the banks, courts and public services of a country of 1.4 billion. It recommends diverse open weights from American, Chinese, European and Indian sources.

Computing power is the machinery needed to train, adapt and run AI models. India generates nearly a fifth of the world’s data, yet possesses less than 5 per cent of global AI computing power. The gap matters because data alone does not create useful or sovereign AI. Models also need enough accelerators, data centres and electricity to process it.

The IndiaAI Mission has subsidised computing, but every accelerator in that pool is foreign-designed, overwhelmingly by Nvidia. A US rule in January 2025 also capped how many advanced chips India could import, before the restriction was rescinded in May. These facts show how outside decisions and suppliers can constrain India’s capacity.

India is responding with large model plans and public investment, including 103.7 billion rupees approved through the IndiaAI Mission. But the article argues that chasing scale before building foundations will not close the gap. Domestic hardware, reliable power and broad access to computing would support many Indian teams instead.

An API lets users access a model remotely rather than possessing the model itself. That makes the service convenient, but it also creates dependence on the supplier and its government. If access rules change, a country may suddenly lose important AI capabilities, even when its users and institutions have built services around them.

The article gives Paradox’s shutdown as the warning. On June 12, foreign users lost access to two flagship models after a US export directive. Paradox could not check every user’s nationality, so it switched the models off. By contrast, open weights downloaded at home cannot be remotely switched off by Washington.

Local open weights are not automatically safe or neutral. Chinese models may contain Beijing’s sensitivities, such as Pramana’s refusal to answer about Arunachal Pradesh. But open weights can be audited and retrained. India could then adapt them for its own laws, languages and public services while avoiding dependence on one cancellable API.

Pre-training creates a broad model, but post-training changes how that model performs on particular tasks. Teams can use Indian-language data, legal material and public-service examples to improve an open model for local needs. This is cheaper and more focused than building a frontier model from the beginning.

The article says post-training costs a small fraction of pre-training. It can also fix behaviour that is unsuitable for India. A model that will not discuss Arunachal Pradesh, for example, can be adjusted through post-training. In the author’s published research, an agent automating this process lifted a small model’s pass rate on one task from 84.9 to 99.3 per cent.

This approach lets many Indian teams work on agriculture, health, law and public services. Singapore used further training on open base models such as Google’s Veritas for SEA-LION. The article argues that adaptation delivers useful capability now while India develops stronger base models over time.

Independent testing matters because governments need evidence before committing more public money. The evaluator should not be controlled by the funders or model builders. Private tests in Indian languages could measure real performance while published results would let citizens and institutions compare systems.

The article says there is currently no independent institution checking whether funded models perform as claimed. Most of the 12 companies selected two years earlier to build public-welfare AI reportedly moved away from that work because the government had no way to buy what they built. India is considering an anchor investment of up to 200 billion rupees without a public scorecard for the earlier spending.

The proposed answer is simple: no milestones, no money. India should fund adaptation, data and computing for hundreds of teams, and diversify across American, Chinese, European and Indian open weights. Competition would make failure visible, reward progress and prevent one supplier from becoming indispensable.

AI chips are specialised hardware that supplies the computing power models need. Data centres provide the buildings and equipment where those systems run. Electricity powers both the chips and the wider facilities. Together, they determine whether AI can be trained, adapted and delivered reliably at national scale.

India’s current foundations are weak. Every accelerator in the IndiaAI Mission’s subsidised pool is foreign-designed, mostly by Nvidia. Nearly five years after approving a 760-billion-rupee semiconductor programme, India has no fab in commercial production. Tata’s Dholera plant will reportedly begin at 90 nanometres rather than the promised 28nm. AI data centres may add 26.3 gigawatts to electricity demand by 2031-32.

Water and power are already concerns, including at Google’s 15-billion-dollar data-centre project. The article calls for clean, firm power and a chip roadmap built around packaging and testing plants India already has. Those investments determine whether AI is genuinely controllable at home.