4 minutes reading time
I hope I am wrong.
I genuinely hope that three years from now, some state education department or municipal corporation isn't signing a multi-crore annual procurement contract to rent API tokens from a private domestic vendor like Sarvam, just so politicians can stand in front of a banner and claim we have achieved "Sovereign AI."
Because that is the easiest, laziest path available to the bureaucracy, and our bureaucracy never misses an opportunity to take the easiest path.
For months now, "Sovereign AI" has been bouncing around tech panels and ministerial press releases. Whenever the term gets trotted out, it sounds heroic: national technological autonomy, strategic independence, cultural representation in silicon.
Then you look at how Indian procurement actually works, and the real danger becomes obvious.
The government will confuse having a domestic vendor with having national infrastructure. Instead of funding open-weight foundation models as a shared public utility, they will simply pick a favoured domestic startup and funnel public money into paying its inference bills.
That isn't technological sovereignty. That is just private rent-seeking wrapped in a tricolor flag.
And before some salaried middle-class cunt jumps into the comments to complain about how this wastes "the taxpayer's hard-earned money": shut up.
The Indian middle class has this obnoxious, persistent delusion that they single-handedly finance the Republic because they file an ITR on their salary. They completely ignore indirect taxes. A daily-wage laborer making four hundred rupees a day digging ditches pays between 5% and 28% GST on practically everything he buys: his tea, his soap, his biscuits, his mobile recharge voucher. Relative to his disposable income, that man pays a far more brutal tax rate than the IT engineer crying about income tax deductions.
It is public money. It belongs to the laborer as much as the tech worker. And handing that money over to private software monopolies to rent closed APIs is an insult to everyone who paid into the pool.
Sorry, I digressed.
There's another physical side of this too.
Anyway, If you want actual sovereignty, the stack has to be split into three distinct layers instead of being bundled into a single corporate handout:
Model Capability (The Public Commons): The foundation weights must be open. Whether built through an NPCI-style non-profit utility or academic consortia, the weights, training recipes, and tokenizers must be public goods. Universities, state governments, startups, and solo developers must be able to download the model, run it offline, fine-tune it on local dialects, and inspect it without asking anyone for permission.
Physical Infrastructure: The silicon, high-voltage substations, and cooling sheds. Even if government departments buy compute on merit, we know Adani and Ambani are sitting on the physical chokepoints of land and grid connections. That bottleneck is hard enough to break through on its own.
Commercial Inference Layer: This is where private companies should actually make money. Let Sarvam, local cloud hosts, and startups compete purely on merit: latency, cost per token, fine-tuning services, and enterprise reliability.
If a government department needs inference, it shouldn't be locked into a single vendor's closed model. It should be able to run a sovereign open-weights model on whatever compute provider offers the cheapest, most reliable throughput. If Vendor A gets greedy or shuts down, the department takes the exact same weights and points them at Vendor B. Zero lock-in.
That was the entire genius of the UPI framework. The core protocol and settlement rails were built as public infrastructure; private apps like PhonePe and Google Pay competed on interface and distribution on top.
If we don't build open-weight foundation models, we won't get an AI commons. We will just get a domestic version of the same old proprietary lock-in: paying monthly rent on a closed API, while calling it patriotism.
Image Credit: Honoré Daumier - Gargantua.jpg via Wikimedia Commons