7 minutes reading time
A few weeks ago, developers across the world woke up to find their API keys dead. Well not exactly. However, let me go on. pretend that API keys were dead, as that sounds better than what alternative I can write.
So, A few weeks ago, developers across the world woke up to find their API keys dead.
No warning email. No deprecation grace period. Just a sudden policy shift or geopolitical order in Washington, and a private American AI company severed access for thousands of foreign users. One morning you're building a product on top of someone else's model; by lunch, your entire tech stack is a collection of dead HTTP 403 responses.
It was a hilarious, brutal lesson in what happens when you mistake a rented API for infrastructure.
For the last three years, Indian tech founders and policy consultants have been running around throwing the phrase "Sovereign AI" into every panel discussion. They talk about it like it's a national defense initiative. They give speeches about how India cannot remain a passive consumer of intelligence trained elsewhere.
And then, when you ask what the plan actually is, the government points to a handful of seed-stage startups and says: "We gave them some compute credits! They'll build our foundation models!"
It's absurd.
Training a frontier foundation model isn't like building another SaaS dashboard to automate invoice parsing for US dental clinics. It requires tens of thousands of cluster GPUs, petabytes of clean multilingual text, specialized distributed systems research teams, and hundreds of millions of dollars burnt in electricity alone.
A startup's primary goal is not national technological autonomy. A startup's goal is to survive until the next funding round, show 3x month-over-month ARR growth, and eventually exit. You cannot ask a twenty-person startup backed by venture capital to spend fifty million dollars training a foundation model and then hand the model weights over to the public for free. The incentives don't line up. The moment the venture money dries up, the startup either closes the weights behind a paid API or gets swallowed by an enterprise.
I shouldn't pretend zero research is happening. Academic initiatives like AI4Bharat and Bhashini are doing legitimate, difficult work scraping Tamil, Telugu, and Marathi speech corpora and releasing open language toolkits on shoe-string budgets. I see them. But while academic researchers scramble for crumbs of government grants to map regional dialects, our multi-billion-dollar corporate behemoths sit back, build real-estate datacenters, and wait for foreign labs to hand down foundation models.
Other countries understood this basic arithmetic years ago.
Look at South Korea. They knew a country of fifty million people couldn't outspend the United States or China on raw compute. So instead of praying that some lone startup would magically invent a national model, they formed coordinated consortia. Naver, LG, SK Telecom, Seoul National University, and KAIST pooled their funding, research teams, and datasets into a shared initiative to build open-weight models. They treated foundation AI as national infrastructure - the same way governments treat power grids or telecom spectrum.
Even China, despite its tight state controls, recognized that open-weight models like Qwen and DeepSeek give their engineering ecosystem global leverage and strategic resilience against Western sanctions.
If India actually wanted a serious sovereign AI strategy, the blueprint is staring us right in the face. Take NPCI's Section 8 non-profit utility structure, blend it with South Korea's consortium approach, and apply it to foundation models.
Imagine Infosys, TCS, Reliance, Zoho, and the IITs pooling capital into a Section 8 AI Foundation. The mandate: train and periodically release open-weight models for Indian languages. The weights are released to the public for free so anyone can run them locally or host them. But how does the foundation survive? By offering high-throughput, enterprise-grade managed API inference to corporations and member firms at competitive, cost-plus rates.
NPCI proved that a Section 8 non-profit utility can generate over ₹1,500 crore in annual surplus while building massive data infrastructure, without needing to hand out stock dividends to private shareholders. If member companies consume inference from a shared non-profit model at competitive rates, the foundation stays self-sustaining, the country gets a sovereign frontier model, and the firms who pooled funds get shared, zero-lock-in intelligence rails. Even if it barely breaks even, everyone wins.
Why hasn't this happened? Because NPCI didn't spring spontaneously from the generous hearts of commercial bank CEOs. It happened because the Reserve Bank of India stood over them with regulatory authority and forced them into a room until they built a shared payment rail.
I'm not asking for state commissars to dictate private balance sheets or force companies to burn capital at gunpoint. God knows we have enough bureaucratic overreach as it is. But the depressing reality of Indian corporate capital is that without regulatory coercion or immediate client billing hours, it has zero capacity for enlightened self-interest. In Silicon Valley, Meta releases open weights to commoditize their competitors' moats. In China, internet giants build and release Qwen and DeepSeek to secure developer ecosystems. In India, corporate capital doesn't move out of technical vision or strategic self-interest; it only moves when a central bank whip cracks over its back, or when there's subsidized land to buy for server farms.
Now look at what Indian capital is doing instead.
Are Reliance or Adani or Tata funding national open-weight foundation models? Are they putting fifty thousand GPUs into a non-profit consortium with IISc and the IITs to release open model weights for Indian languages?
Of course not.
They're buying up acres of farmland, putting up steel sheds, hooking them to the high-voltage grid, and building GPU datacenters.
Now, to be fair, you can't train a model on fresh air. Compute capacity is a strict prerequisite for AI sovereignty; you need the silicon and power infrastructure before you can run a single training epoch. Building datacenters isn't inherently evil or useless.
The grift isn't in putting up the servers. The grift is stopping there: treating server racks as the final destination, securing subsidized land and power, and then acting shocked when all that compute just gets rented out as cloud hours to run proprietary foreign APIs instead of training open weights. Adani gets to sell power and real estate; Reliance gets to pack GPU servers into cloud bays; and Indian startups get to burn VC cash renting domestic compute to call closed foreign APIs.
It’s the real estate grift rebranded as artificial intelligence.
Let's be clear about what "open weights" actually mean. They aren't pure FOSS. A true open-source model would mean having the full training dataset, the exact filtering code, the hyperparameter recipes, and a reproducible pipeline that anyone can run on a laptop for five bucks. We don't have that yet - reproducing a 70-billion parameter model costs millions in compute alone.
But open weights give you something crucial: resilience. If you have the weights downloaded on a server inside your own cluster, no API vendor in San Francisco or policy bureaucrat in Washington can revoke your key tomorrow. You can quantize it, fine-tune it on local Tamil or Kannada datasets, host it on your own hardware, and run it until the silicon melts.
Without open weights, "Sovereign AI" is just a marketing slogan for buying American hardware and hosting it in a Bania datacenter.
We had the same illusion twenty years ago with telecommunications, ten years ago with cloud hosting, and today with foundation models. We build the physical sheds, lay the fiber, light up the server racks, and then bow down to whatever proprietary foreign software layer runs inside them.
The datacenters will get built. The press releases will announce thousands of H100s sitting in air-conditioned halls. And when the switch gets flipped from above, we'll still be sitting there with dead terminal prompts, paying rent on the land we used to own.