The money is melting. That’s the only way to describe the current vibe in Bengaluru’s HSR Layout and the glass towers of Mumbai. If you listen to the press releases, India is on the verge of an AI renaissance. If you look at the balance sheets, it’s a controlled demolition.

Indian tech firms are currently sprinting into a furnace. They’re ditching the quaint notion of "profit" or "sustainable margins" to play a high-stakes game of user hoarding. It’s a familiar script. We saw it with ride-hailing. We saw it with food delivery. Now, we’re seeing it with Large Language Models (LLMs) that speak Marathi and Kannada. The strategy is simple: give the tech away, buy the users, and pray that someone figures out how to charge them before the venture capital evaporates.

It’s a massive gamble.

Building an LLM isn't like building a grocery delivery app. You don’t just hire a few thousand guys on scooters and call it a "platform." You need compute. Specifically, you need NVIDIA H100s. These chips cost about $30,000 a pop on the open market, assuming you can even find one. For a startup like Krutrim or Sarvam, the hardware bill alone is enough to induce heart palpitations. Yet, they’re offering their services for pennies—or for nothing at all.

Why? Because in India, the "user" is the ultimate currency, even if that user has no intention of ever paying a monthly subscription fee.

Take a look at the localized models hitting the market. Projects like "Hanooman" or the various Bharat-centric wrappers are being pushed as a way to "bridge the digital divide." That’s the marketing fluff. The reality is a desperate land grab. These companies are trading near-term revenue for a slice of a population that’s just getting used to talking to a bot. They’re subsidizing the massive electricity bills and API costs of millions of queries just to ensure that when an Indian farmer asks about crop yields in Hindi, he uses their interface and not ChatGPT.

But there’s a friction point no one wants to talk about: the "Llama" problem. Most of these "sovereign" Indian AIs aren't built from scratch. They’re fine-tuned versions of Meta’s open-source models. It’s like putting a fancy new grill on a Ford and calling it a local invention. The real cost isn't the innovation; it's the inference. Every time a user asks a question, the company loses money. It’s a reverse ATM.

Investors don’t seem to mind. Not yet. They’re still drunk on the "next billion users" narrative. They see a country with 800 million smartphone users and think, surely, we can squeeze a dollar a month out of 10% of them. They ignore the fact that India’s average revenue per user (ARPU) is notoriously bottom-heavy. In the telecom sector, it took a decade of brutal price wars to get people to pay more than the price of a vada pav for their data. AI won't be any different.

The trade-off is getting ugly. To keep the lights on while chasing these free users, firms are cutting corners elsewhere. Marketing budgets are being cannibalized. Engineering talent is being shifted away from boring, revenue-generating B2B tools to focus on flashy consumer-facing chatbots that hallucinate in twelve different scripts. It’s a pivot toward the shiny, away from the solvent.

We’ve been here before. We remember the "blitzscaling" era where startups burned billions to own a market that didn't want to be owned. The difference this time is the sheer scale of the burn. A food delivery startup loses money on the petrol and the chicken. An AI startup loses money on the fundamental laws of physics and the global scarcity of silicon. It is a significantly more expensive way to go bankrupt.

The pitch to the public is that this is about national pride. It’s about not letting Silicon Valley dictate the future of Indian intelligence. That’s a noble sentiment. It’s also a convenient shield for a business model that currently resembles a Ponzi scheme built on top of a GPU cluster. If you can’t show a path to revenue, you show a path to "adoption." If the adoption doesn't lead to revenue, you just call it "ecosystem building" and hope for a bailout or an acquisition by a conglomerate like Reliance or Tata.

At some point, the bill comes due. The H100s will get old. The investors will want their 10x return. The users, currently enjoying their free AI-generated poetry and coding help, will be asked to open their wallets.

Will they? Or will they just move to the next startup willing to set its cash on fire for the privilege of their attention?

Silicon Valley’s cast-off business models are being recycled in Mumbai and Bengaluru with a fresh coat of "sovereign AI" paint. It’s a bold move. It’s also a reminder that in the tech world, the easiest thing to grow is a deficit.

The real question isn't whether India can build a world-class AI. It’s whether it can afford to keep it running once the "free" sign is taken down.