The buffet was lukewarm, but the rhetoric was boiling.

New Delhi likes a good spectacle, and the latest tech summit didn’t disappoint. We got the usual parade of dark suits, the mandatory mention of "Digital India," and a fresh pile of promises about making the country the world's "AI back office." Or front office. Honestly, the metaphors get a bit blurry after the third panel on "Sovereign AI."

The big play here is a two-pronged offensive: building domestic AI models and sticking high-end labs into every campus that can find a spare room. It’s ambitious. It’s expensive. It’s also incredibly messy.

Let’s talk about the labs first. The government wants to plant AI centers across universities like they’re Starbucks franchises. The idea is to hand students the keys to expensive compute power before they have a chance to finish their visas and head for Mountain View. It’s a nice sentiment. But anyone who’s spent ten minutes in a standard Indian engineering college knows the friction isn't just about the hardware. It’s about the power grid. It’s about the fact that you’re trying to run Nvidia H100s—chips that suck electricity like a small town—in regions where the lights still flicker when it rains too hard.

Then there’s the price tag. The "IndiaAI" mission has a budget of roughly $1.25 billion. In the world of global venture capital, that’s a decent Series B. In the world of building a sovereign AI stack from the dirt up, it’s pocket change. For context, Microsoft is dropping $10 billion just to keep Sam Altman’s servers humming. India is trying to build an entire ecosystem, subsidize GPUs for startups, and retrain a workforce of millions for the cost of a few mid-sized San Francisco skyscrapers. The math doesn't just feel tight; it feels impossible.

And don’t get me started on the "Sovereign LLMs." There’s this desperate urge to build models that speak the country’s 22 official languages fluently. It’s a noble goal, sure. But these models aren't being built in a vacuum. They’re being trained on datasets that are, at best, incomplete and, at worst, an absolute dumpster fire of internet scrapings. You can’t just "policy" your way into a clean dataset of Marathi or Kannada colloquialisms. You need clean, labeled data. You need humans to sit in rooms for thousands of hours for miserable wages. It’s a grind, not a ribbon-cutting ceremony.

The friction is real. While the ministers were on stage talking about democratizing intelligence, the reality is a supply chain bottleneck that would make a logistics manager weep. India wants 10,000 GPUs. It wants them now. But so does everyone else. Nvidia has a waiting list longer than a Delhi traffic jam, and they aren't exactly handing out "emerging market" discounts. Every chip that lands in a campus lab in Chennai is a chip that isn't earning six figures for a hedge fund in New York. That’s a hard trade-off to navigate when you’re playing with taxpayer money.

The startups are feeling the squeeze, too. They’re being told to "innovate" while being starved of the very compute they need to compete. It’s like asking someone to win a Formula 1 race while giving them a voucher for a bicycle. They’re pivoting to "wrappers"—thin layers of software built on top of OpenAI or Claude—because actually training a foundation model in India is a quick way to burn through your seed round in a weekend.

So, we have the summit. We have the campus labs. We have the bold proclamations about "AI for all." It looks great on a LinkedIn carousel. It makes for a fantastic press release that domestic outlets will parrot without asking about the kilowatt-hour requirements or the actual availability of the silicon.

But behind the high-gloss presentations, there’s a nagging sense of déjà vu. We’ve seen this play before with semiconductors. We’ve seen it with "smart cities." The government builds the shell, invites the cameras, and then realizes that the plumbing is harder than the PR.

It’s easy to announce a lab. It’s much harder to make sure the lab has air conditioning that works 24/7 so the $30,000 chips don’t melt into expensive puddles of silicon.

By the end of the day, the delegates headed for the exits, dodging the stray cows and the luxury SUVs alike. The "tech push" is officially on. We’ll see how many of those 10,000 GPUs actually show up, or if they’ll just become the world’s most expensive paperweights in a series of very quiet university rooms.

The real question isn't whether India can build an AI model. It’s whether it can keep the air conditioning on long enough to train it.