The World Built AI in English First. India Just Spent $1.25 Billion to Build One in Its Own Languages

Ask most large language models a question in Bhojpuri, or Dogri, or Santali, and you will get one of two answers. Either the model fails outright, or it quietly answers as if you had asked in Hindi, guessing at a language it was never seriously trained on.

That gap is not a rounding error. India has 22 languages written into the constitution and, by the last census count, well over a hundred more spoken by at least ten thousand people each. Almost none of that shows up in the training data behind ChatGPT, Gemini or Claude, which learn overwhelmingly from English and a handful of other global languages.

India’s government decided that gap was worth closing itself, and has now put real money and real hardware behind three separate bets on how to do it. One is working better than anyone expected. One is unproven. One has already had to admit it overpromised.

The compute India is giving away

In March 2024 the Union Cabinet approved the IndiaAI Mission with a total outlay of ₹10,371.92 crore, close to $1.25 billion, spread over five years. Roughly 44 per cent of that, about ₹4,563 crore, went to a single problem: getting Indian researchers and startups access to GPUs, the specialised chips that train and run AI models, without having to pay Silicon Valley prices for them.

By late 2025 more than 38,000 GPUs had been brought onto the government’s AI Compute Portal, rented out to approved applicants at heavily subsidised rates. At the India AI Impact Summit in New Delhi in February 2026, IT Minister Ashwini Vaishnaw announced 20,000 more were being added on top of that, with a stated target of 100,000 public GPUs by the end of 2026.

The rest of the mission’s ₹10,372 crore funds six more pieces: a shared dataset platform called AIKosh, a startup financing arm, a skilling program aimed at getting AI literacy into the public sector, an application development push, and a safety and trust pillar. It is, in effect, an attempt to build the entire supporting infrastructure a country needs to train its own AI, not just hand out chips.

The part that is already working: translation, at scale

The oldest piece of this stack predates the compute mission by two years. Bhashini, India’s National Language Translation Mission, was launched by the prime minister in July 2022, run out of the Ministry of Electronics and Information Technology. It offers speech-to-text, text-to-speech and translation across the 22 constitutionally scheduled languages, exposed as free APIs that any developer or government office can plug into.

One industry tracker reported that by early 2026, Bhashini had crossed six billion cumulative requests since launch, handling around fifteen million a day with sub-second response times. State Bank partnerships have used it to offer multilingual banking services, and it has been deployed for live translation at large public gatherings including the Maha Kumbh.

This is the unglamorous, plumbing-layer version of the sovereign AI push, closer in spirit to UPI than to a chatbot. Nobody is claiming Bhashini writes poetry. It moves text and speech between Indian languages at a scale that matters, and it has kept doing that quietly while the more headline-grabbing parts of the strategy played out in public.

The sovereign model: a real shot, not yet proven

The showpiece of the mission is the attempt to build a foundational large language model trained specifically for Indian languages, rather than fine-tuned onto an American one after the fact. In 2025 the government selected Sarvam AI, a startup founded in 2023 by Vivek Raghavan and Pratyush Kumar, out of 67 applicants to lead this effort, backing it with access to 4,096 Nvidia H100 GPUs.

Sarvam had already raised about $41 million from Lightspeed, Khosla Ventures and Peak XV Partners in December 2023, one of the largest early bets on an India-focused AI company at the time. In February 2026 it released two open-weight models, Sarvam-30B and the larger Sarvam-105B, alongside speech models for Indian-language transcription and text-to-speech. Nvidia has featured the partnership as a case study in what it calls sovereign AI, framing it as a model built to serve 1.4 billion people in their own languages.

Here is the honest complication. Sarvam has published strong benchmark numbers against models like DeepSeek, but as Forbes reported in March 2026, those models do not appear on independent leaderboards like Hugging Face’s Open LLM Leaderboard, and much of the evaluation data comes from benchmarks Sarvam designed and scored itself, using judging criteria set by the company whose founders also built some of the standard India-specific benchmarks the field already relies on. That is not evidence of fraud. It is evidence that the claims have not yet been checked by anyone with no stake in the outcome, which matters more, not less, given that Sarvam’s models are already running inside real financial services products.

The cautionary tale: Krutrim

The fourth name in this story is Krutrim, founded by Ola’s Bhavish Aggarwal and launched at the start of 2024 with $50 million in funding at a $1 billion valuation, becoming India’s first AI unicorn within weeks of launch. It promised a large language model trained on Indian languages, a conversational assistant called Kruti, and eventually its own AI chips.

The early reception was rough. Users flagged translation errors and factual mistakes in the beta chatbot almost immediately, and reviewers found the company’s cloud offering fell short of what developers expected. By May 2026, TechCrunch reported that Krutrim had cut more than 200 roles across several rounds of layoffs, pulled the Kruti app from app stores, paused its in-house chip design work, and pivoted from building foundation models toward selling AI cloud infrastructure to enterprise customers instead. The company says it is now profitable on that new business, though earlier reporting found that roughly 90 per cent of its prior year’s revenue had come from its own parent company, Ola, rather than outside customers. One industry analyst summed up the shift bluntly: the standard of proof has to rise with the size of the claim.

What this actually adds up to

Put the four pieces next to each other and a pattern shows up that is more interesting than either the triumphant version or the cynical one. The unglamorous infrastructure bet, cheap GPUs and a translation API nobody argues about, is the part that is demonstrably working at scale. The ambitious bet, a homegrown foundation model to rival the ones out of the United States and China, has real government backing and real hardware behind it, but its own claims are still waiting on outside verification. And the bet that moved fastest and talked the loudest had to walk itself back within two years.

None of that makes the underlying goal wrong. A language spoken by tens of millions of people and ignored by every major AI lab is a real gap, and closing it is worth doing whether or not any single Indian company ends up owning the model that does it. What India has actually built so far is not a rival to GPT-4. It is the compute, the data pipes and the language layer that any future model, homegrown or otherwise, will have to run on. That part, at least, is not a claim. It is already carrying fifteen million requests a day.