India Is Building a Trillion-Parameter AI Model. It Won’t Cost What You Think.

Sarvam AI is pricing frontier-scale intelligence at a fifth of OpenAI’s rate — the subsidy and the architecture that make it possible, and the timeline I don’t buy yet.

If you’ve priced out a frontier AI model for a product roadmap, you know the number that stops the conversation.

It’s rarely the capability. It’s the invoice.

I assumed a trillion-parameter model was a rich country’s game — the kind of compute bill only OpenAI, Google, or Anthropic could absorb.

Then I found the price tag on the one being built in Bengaluru: $0.80 per million tokens, versus $4.50 for GPT-5.4 Mini.

That’s not a rounding error. That’s a different category of business model.

In this article, I’ll walk through Sarvam AI’s path from a 2-billion-parameter model to a trillion-parameter one, show you exactly what makes that price possible, and tell you the one part of the story I’m not ready to believe yet.

What Sarvam AI Actually Is

Sarvam AI was founded in Bengaluru in August 2023 by two people who’d spent a decade each on opposite halves of the same problem.

Vivek Raghavan spent nearly two decades building digital infrastructure at national scale — the kind of systems that have to work for a billion people at once, not a million.

Pratyush Kumar spent years at IBM Research and Microsoft Research measuring something specific: exactly how badly Western AI models fail when you ask them to work in Hindi, Tamil, or the 20 other languages India recognizes officially.

Both came out of AI4Bharat, a research group at IIT Madras. That’s not a branding detail — it’s the reason Sarvam started with languages instead of chasing a general-purpose chatbot like everyone else in 2023.

Raghavan holds a PhD in Electrical and Computer Engineering from Carnegie Mellon and built two Electronic Design Automation companies before this. Kumar earned his PhD at ETH Zurich, then split time between IBM Research, Microsoft Research, and a faculty post at IIT Madras.

Neither founder came from a typical “AI startup” background. Both came from infrastructure and research — the kind of experience that shows up later in how methodically the company built up from tokenization to trillion-parameter ambitions, rather than skipping straight to a flashy demo.

Example: A model trained mostly on English text doesn’t just translate badly into Tamil — it often can’t even tokenize the script efficiently, meaning every sentence costs more compute and gives worse answers before the “intelligence” part even starts.

That inefficiency compounds at scale. A poorly tokenized language can cost 3–4x more tokens to express the same sentence an English speaker would type in a fraction of the space — which means every user in that language was quietly subsidizing a worse, more expensive experience with every query.

Fixing tokenization for 22 languages before touching model scale was a deliberate bet: solve the unglamorous infrastructure problem first, then build up.

The Timeline: From 2 Billion to Trillion-Plus Parameters

The jump from Sarvam’s first model to its trillion-parameter announcement happened in under two years. Here’s the actual sequence.

Notice what’s missing from that table: a multi-billion-dollar compute budget. Sarvam didn’t buy its way to trillion-parameter ambitions the way a hyperscaler would. It got there through two specific mechanisms — one architectural, one governmental.

Also notice the pace. Twenty-three months separate the founding date from a trillion-parameter announcement. For comparison, it took OpenAI roughly five years to go from GPT-2 to GPT-4-scale ambitions, with a war chest measured in tens of billions of dollars along the way.

Sarvam compressed a similar scale-up into a fraction of that time, on a fraction of the funding. That compression is either the most interesting part of the story, or the most suspicious part — possibly both.

The Real Lever: Mixture-of-Experts, Not Magic

Sarvam 105B — the model that exists today and sets the pricing baseline — uses something called a Mixture-of-Experts architecture, or “MoE.”

Here’s what that means in plain terms. A traditional model activates its entire brain for every single question, however simple.

An MoE model is built from many smaller specialist sub-networks, called “experts.” For any given question, it only wakes up the handful of experts relevant to that question — not all of them.

Example: Ask it a question about Tamil grammar, and only the experts trained heavily on Tamil language patterns activate. The experts trained on, say, Python syntax stay dormant for that request.

Bad: A dense 105-billion-parameter model that activates all 105 billion parameters for every query — expensive, slow, wasteful.

Good: An MoE 105-billion-parameter model that only activates a fraction of those parameters per query — most of the “size” exists on paper, not on your compute bill.

Scale that logic up to a trillion parameters and the gap becomes enormous. A dense trillion-parameter model would be nearly impossible to serve affordably to real users — the compute cost per answer would be staggering.

An MoE trillion-parameter model, by contrast, can still only wake up a small slice of its total size for any single question. That’s the entire premise behind Sarvam pricing an unreleased trillion-parameter model at a similar discount to its current 105B one — the architecture, not the parameter count, sets the price.

This is the actual mechanism behind the low price. It’s not that Sarvam found cheaper GPUs. It’s that the model architecture itself asks for less work per answer.

MoE isn’t a Sarvam invention — DeepSeek, Mistral, and several frontier labs use versions of the same idea. What makes Sarvam’s version notable is combining it with a corpus built specifically for Indian languages from the start, rather than bolting multilingual support onto an English-first model after the fact.

The distinction matters more than it sounds. A model retrofitted for a new language is translating internally before it reasons. A model built with that language in its training data from day one reasons directly in it — fewer wasted computation steps per answer, which is its own quiet contributor to the lower price.

Compute You Don’t Pay Market Rate For

Architecture explains part of the price gap. The other part is where the compute itself comes from.

In April 2025, India’s government selected Sarvam from a pool of 67 applicants to build the country’s sovereign large language model, under the ₹10,370 crore ($1.2B) IndiaAI Mission.

The prize wasn’t just prestige. It came with access to 4,096 Nvidia H100 GPUs, delivered through Yotta Data Services — roughly ₹99 crore, about $12 million, in compute subsidies.

That’s the detail most coverage skips past. Sarvam isn’t renting compute at the same market rate a Silicon Valley startup pays a cloud provider.

It’s running on infrastructure a government decided was worth subsidizing as national policy — the same logic countries use to subsidize highways or power grids, applied to AI compute instead.

That doesn’t make the achievement fake. It makes it a different kind of achievement — one built on policy leverage, not just engineering cleverness.

The IndiaAI Mission itself is worth understanding as more than a grant program. It’s a ₹10,370 crore government commitment — roughly $1.2 billion — designed explicitly to prevent India from renting its AI capability from American or Chinese labs forever.

Sarvam’s 4,096 GPUs are one allocation inside that larger bet. Tata Communications and E2E Networks were separately brought in to build parallel data center capacity, meaning the government isn’t betting the entire program on one company succeeding.

Think of it as a national infrastructure project wearing the clothes of a startup funding round. Highways and power grids get built the same way — public money absorbing the upfront cost that private capital alone wouldn’t justify at this stage of the market.

Compare 4,096 H100s to what a frontier lab typically trains on. OpenAI and Anthropic have discussed training runs using tens of thousands of GPUs, sometimes over 100,000 for the largest efforts.

Sarvam’s allocation is a fraction of that fleet size. Which means the trillion-parameter ambition isn’t just about having compute — it’s a bet that smarter architecture and a tightly focused training corpus can close a gap that raw GPU count alone would otherwise decide.

The Price Tag That Changes the Math

Put the architecture and the subsidized compute together, and you get the number that actually matters to a developer deciding what to build on.

Sarvam 105B is priced at $0.80 per million blended tokens. GPT-5.4 Mini costs $4.50 for the same volume — 5.5 times more. Gemini 3.5 Flash costs $9–11 times more.

The trillion-parameter model, still unreleased, is being pitched at a similar 5x discount to frontier rivals — the same playbook, at a much bigger scale.

For a developer building a product with real usage volume, that’s not a rounding difference. At meaningful scale, it’s the difference between a viable product and a compute bill that kills the business model.

Example: A customer service platform processing 500 million tokens a month pays roughly $400 on Sarvam 105B. The same volume on GPT-5.4 Mini costs $2,250, and on Gemini 3.5 Flash it’s $4,500.

Over a year, that gap alone can fund another engineer’s salary — money that would otherwise disappear into a token bill.

This is the same pattern I’ve written about with DeepSeek’s pricing moves — architecture plus infrastructure economics beating raw scale. Sarvam is running that playbook with a government’s compute budget behind it instead of a hedge fund’s.

There’s a second-order effect worth naming too. Cheap frontier-adjacent intelligence doesn’t just help Indian developers — it resets the floor for what every regional AI player globally has to compete against.

If a $0.80-per-million-token model performs even reasonably close to a $4.50 one on real-world tasks, the pricing power of the expensive incumbents erodes a little every time a developer runs that comparison for themselves.

Who Else Is Racing for the Same Prize

Sarvam isn’t India’s only sovereign AI bet, and the differences between the contenders say a lot about their strategies.

Krutrim, founded by Ola’s Bhavish Aggarwal in April 2023, bundles AI with cloud infrastructure and is aiming squarely at consumers — the smartphone, not the API.

BharatGPT, built by Reliance alongside nine IITs, leans on Reliance’s telecom-scale compute for multimodal healthcare and finance applications.

Sarvam’s lane is narrower and, I’d argue, sharper: developers and enterprises who need an API, not a consumer app. Both Sarvam and Krutrim reached unicorn status in 2026 — proof the market believes at least two of these bets will pay off.

What’s striking is that none of these three companies are competing head-on for the same customer.

Krutrim wants the consumer’s phone. BharatGPT wants Reliance’s existing telecom and finance relationships. Sarvam wants the developer writing an API call.

That’s not an accident — it’s what a government-scale bet on “sovereign AI” is supposed to produce: not one winner, but several companies covering different layers of the same national stack, so no single point of failure sinks the whole strategy.

The Question I Can’t Get Past

Here’s where I have to be honest instead of impressed.

Trade press coverage of the trillion-parameter announcement has been split — real interest mixed with what one outlet called “quiet eye-rolling.”

I share the eye-rolling, specifically on one point: the timeline. Pratyush Kumar said the model would be live within roughly six months of the July 2026 announcement.

Frontier labs building trillion-parameter-class models — OpenAI, Google, Anthropic — typically take 12 to 18 months or longer, running compute budgets that dwarf 4,096 GPUs, H100s or not.

Nobody I found in my research directly disputed Sarvam’s underlying compute-cost math. The architecture argument holds up. The subsidy is real and documented.

What’s unverified is the engineering execution timeline — data curation, training stability at that scale, safety evaluation, and the inevitable delays that come with all three. Every one of those steps has humbled well-funded frontier labs before.

But cost-per-token and time-to-train are two different problems, and low cost doesn’t automatically buy you speed. I’d treat the six-month claim as a target the company is motivated to hit publicly, not a certainty.

There’s a real difference between announcing a roadmap on stage at a developer conference and shipping a working, safety-tested, trillion-parameter model that performs at the level being promised. Every frontier lab has learned that gap the hard way, usually publicly.

I’d rather be wrong for underestimating Sarvam than credulous for taking a stage announcement at face value. If the model ships on time and performs as claimed, that’s a genuinely bigger story than the announcement itself — and I’ll happily write the follow-up.

Six months from a late-July announcement puts the real test in early 2027. Mark the date, and watch what actually ships against what was promised on stage.

Take Aways

• Sarvam 105B is priced at $0.80 per million tokens — 5.5x cheaper than GPT-5.4 Mini, 11x cheaper than Gemini 3.5 Flash.
• The price comes from two real mechanisms: a Mixture-of-Experts architecture that activates only a fraction of parameters per query, and government-subsidized compute — 4,096 H100 GPUs via the IndiaAI Mission.
• This isn’t a scrappy startup out-engineering Silicon Valley alone — it’s a case study in what policy-subsidized infrastructure can do to AI economics.
• Sarvam, Krutrim, and BharatGPT are running three different strategies on the same sovereign-AI opportunity — API-first, consumer-first, and telecom-scale, respectively.
• The six-month timeline for a trillion-parameter model is the one claim in this story that hasn’t been proven yet — everything else checks out.

What’s actually being tested here isn’t whether India can build a big model. It’s whether subsidized national compute plus a smarter architecture can out-compete raw hyperscaler spending on price alone.

If that math holds at trillion-parameter scale, every country watching this experiment just got a cheaper playbook to copy.

India isn’t the only government thinking this way. The UAE, Saudi Arabia, and the EU have all floated versions of sovereign AI compute programs — Sarvam is simply the furthest along in proving whether the model actually works at scale.

Whether the trillion-parameter model ships on schedule or not, the subsidy-plus-architecture formula it’s testing will outlive this specific announcement either way. That formula is the real export, regardless of how the six-month bet plays out.

Thank you for reading.

· · ·

If you found this helpful, feel free to click the ‘Clap’ button ❤️

I write about AI infrastructure, agentic systems, and the economics behind them.

Let’s be friends! 🙂 Don’t forget to subscribe so you don’t miss the next one! Find me on LinkedIn.