The power died at 10:14 AM. It wasn’t a total blackout, just enough of a flicker to send the high-definition displays into a digital seizure. For three minutes, the Bharat Mandapam—a venue designed to scream "global superpower"—was just a very expensive, very hot room full of men in linen suits sweating over silent MacBooks.

It was a fitting start. India’s big global AI summit was supposed to be the moment the country stopped being Big Tech’s back office and started being its boardroom. Instead, day one felt like a high-stakes fire drill.

The pitch is simple. Prime Minister Modi’s government wants "Sovereign AI." They don't want to rent intelligence from Mountain View or Redmond. They want to own the chips, the data, and the weights. To prove it, they’ve earmarked roughly $1.2 billion for a national AI mission. Most of that cash is destined for one place: Santa Clara. Specifically, Nvidia’s pockets.

Jensen Huang was there, of course. He’s the closest thing the tech world has to a rock star now that Elon has gone full bunker-mode. Huang sat on stage looking like he’d rather be literally anywhere else, nodding as officials explained how India would build a 10,000-GPU cluster.

Ten thousand. It sounds like a lot. It’s not.

Microsoft probably buys that many H100s on a random Tuesday before lunch. But in New Delhi, this is the grand strategy. The problem is that while the government is ready to cut checks, the local startup ecosystem is currently engaged in a civil war.

On one side, you have the "Optimists." These are the founders claiming they’ve built Indian models that understand the nuance of 22 official languages. On the other side, you have the "Realists." They’re the ones pointing out that most of these "indigenous" models are just Meta’s Llama 3 wearing a digital saree. They’re wrappers. Expensive, state-subsidized wrappers.

The friction turned ugly during a panel on data sovereignty. A founder of a mid-sized Bangalore startup stood up and asked why the government was prioritizing GPU subsidies for conglomerates when local players can’t even get reliable cloud credits. He was politely, then firmly, told to sit down. The mic was cut. Very democratic. Very efficient.

Then came the Bhashini demo. Bhashini is the government’s pride and joy—an AI-led translation tool meant to bridge the country’s massive linguistic divide. It’s a noble goal. It’s also, apparently, very hard to do in real-time when the Wi-Fi is being choked by five thousand journalists trying to upload 4K video. The system lagged. Then it hallucinated. Then it started translating English into a dialect that one linguist in the third row described as "pure gibberish."

Meanwhile, the big American firms are playing a delicate game. They need India’s data. 1.4 billion people generate a lot of training material. But they don’t want the regulation that comes with it. The Ministry of Electronics and IT has been oscillating between "come build here" and "don't you dare release a model without our permission." It’s a regulatory whiplash that has VCs reaching for the Xanax.

The trade-off is glaring. India wants to leapfrog the middle-income trap by betting the house on silicon. But you can’t run a revolution on a power grid that hiccups when the humidity hits 80 percent. You can’t build "Sovereign AI" while you're still begging Jensen for a spot in the delivery queue.

By lunch, the chaos wasn't just technical. It was existential. The high-level delegates retreated to the VIP lounge to eat cold paneer and discuss "digital public infrastructure." Outside, the smaller developers—the ones actually writing the code—were complaining about the lack of open-source datasets. They’re the ones being told to innovate while the government treats AI like a procurement project, like buying fighter jets or paving highways.

But you can’t pave your way to AGI. You can’t just throw 10,000 crore rupees at a problem and expect a GPT-4 rival to emerge from the dust.

As the sun set over the smog-choked horizon of New Delhi, the screens finally stayed on. A flashy video played, showing a futuristic India powered by clean code and perfect algorithms. It looked great. It looked expensive. It looked nothing like the room we were sitting in.

If this is the global pitch, the world might just wait for the second draft. After all, the demo usually looks better than the product. But in this case, the demo couldn't even find the Wi-Fi password.

One official told me, off the record, that the hiccups were just "teething problems." Maybe. But when you're asking the world to trust you with the next century of compute, you should probably make sure the lights stay on for the whole keynote.

Jensen Huang left early. He had a private plane to catch and more chips to sell to people who don't have to ask for permission to turn them on.

The bill for the day’s catering alone likely exceeded the annual budget of the three smartest research labs in the country.