India Cannot Rent Its Way to AI Autonomy: Capability at Home Is the Only Seat at the Table

India's AI Future Will Be Built, Not Declared

India's place in the new AI order will be decided by what it can build and test at home, not by what it announces. Declarations of technological sovereignty are cheap. Compute, chips, models, datasets, talent and the ability to independently test frontier systems are not and they are the only things that will count.

Two powers, two logics, one problem for India The Washington summit of September 23 showed the gap clearly.

The two sides agreed on a thin package: a dialogue on risks, a channel for incident notification and another meeting by November. They could not even agree on a name for the technology.

• Washington treats AI as a race. It will defend its narrow lead through chip controls, action against distillation and tightly held supply chains and prefers to police misuse after the fact rather than slow the frontier.
• Beijing treats AI as a question of control. Its security establishment sees AI as the main arena of great-power rivalry and its first concern is regime security. It exports influence through open-weight models: Alibaba's Qwen has spawned more than 150,000 derivatives, with countries like Singapore, Malaysia and Brazil building on Chinese models or partners.

Why dependence is the real risk Access can vanish. In June, foreign access to top American models was briefly cut off.

Any Indian product, ministry or startup built entirely on one foreign API learned that its continuity was someone else's decision. Rules can be written without India. When two powers hold most of the world's frontier compute, their bilateral risk arrangements tend to become everyone's rules.

The nuclear order shows how risk reduction between leading powers can harden into discriminatory controls on others. Cheap can be costly. Chinese open-weight models are competitive and adaptable, which is exactly why they spread.

But inside government systems, critical infrastructure or sensitive data, an opaque model from a strategic rival is a vulnerability, not a bargain.

TAN's position India should practise differentiated de-risking:

1. Keep Chinese models out of government systems, critical infrastructure and sensitive data.
2. Allow private firms to use open-weight models for low-risk commercial uses only with security testing and local hosting as the minimum condition.
3. Avoid single-vendor dependence on the American stack by keeping applications portable.
4. Invest in a floor of home capability, above all the independent capacity to test frontier systems, because that is what allows India to judge any model rather than trust it. This is not a call to out-spend Washington or Beijing. It is a call to own enough of the stack that India is never merely a customer when the rules are written or the switch is flipped. In AI, as in nuclear policy before it, the countries that shape norms are the ones that bring capability to the table.