India Built Two AI Unicorns on the Same Bet. A Year Later, They Took Opposite Sides.

Two years ago, "sovereign AI" meant roughly the same thing to two different Indian startups: build a large language model trained on Indian languages and Indian data, so the country isn't permanently renting its AI infrastructure from OpenAI or Google. Krutrim, backed by Ola founder Bhavish Aggarwal, became India's first GenAI unicorn on exactly that premise in January 2024. Sarvam AI, founded around the same story, has spent the two years since raising money on it too. This month, both companies are still standing, still described as India's sovereign-AI champions — and pursuing strategies that no longer resemble each other at all.

Krutrim has quietly stopped trying to build a frontier model. In a pivot initiated in late 2025, the company paused its in-house AI model and chip development and reallocated its talent and capital toward cloud infrastructure — selling compute and enterprise AI services rather than competing on model quality. The bet paid off in the way that matters to a company's survival: Krutrim reported roughly ₹3 billion (about $31.5 million) in FY2026 revenue, triple the year before, along with its first-ever annual net profit and margins above 10%. It now serves more than 25 enterprise customers across telecom, finance, healthcare, logistics and manufacturing. What it does not have anymore is a serious claim to owning the underlying model layer.

Sarvam went the other way. In June, it closed a $234 million Series B led by HCLTech, with Bessemer Venture Partners, Khosla Ventures and Peak XV Partners joining, at a $1.5 billion valuation — its largest raise yet, months after Krutrim had already begun retreating from the same territory. Sarvam used part of that momentum to open-source two models in February, Sarvam 30B and a 105-billion-parameter mixture-of-experts model called Sarvam 105B, both trained from scratch on data weighted toward Indian languages and built to handle reasoning, math and code, not just conversational text. It also launched a programme in March giving early-stage startups six to twelve months of free API credits — a distribution play only a company still betting on its model being the product would bother making.

Same government mandate, different customer

The divergence isn't just strategic taste; it maps onto who each company is actually selling to. Krutrim's 25-plus enterprise customers span telecom, finance, healthcare, logistics and consumer internet — the ordinary commercial market for AI infrastructure, where the product is compute and reliability, not a specific model's benchmark scores. Sarvam's deployments run through banking, insurance and, more distinctively, government services: it participates in the IndiaAI Mission's sovereign large language model programme, has worked with UIDAI on Aadhaar-linked services, and lists defence among its sectors. That's the constituency that actually asked for a sovereign model in the first place, and it's the one Sarvam kept building for after Krutrim moved on.

Who's actually carrying the risk

The other difference is who is exposed if either bet fails. Krutrim's founder, Bhavish Aggarwal, has been personally underwriting the company's runway by pledging Ola Electric shares — 5.88 crore shares pledged by March 2025, worth roughly ₹603 crore at the time — which means Krutrim's downside is partly collateralised against the fortunes of a separate, already-struggling listed company. Sarvam's downside sits with institutional venture investors and, now, a strategic backer in HCLTech, a much larger and more diversified balance sheet to absorb a model-building bet that hasn't yet turned a public profit.

Neither company's approach is obviously the wrong one. Krutrim's pivot produced something Indian AI startups rarely have: an actual profit, built on unglamorous enterprise infrastructure sales rather than headline model benchmarks. Sarvam's approach produced something else rare in the same market: a model good enough, and open enough, that the institutions that matter most for a "sovereign" AI claim — the government's own IndiaAI Mission, UIDAI, the banking sector — kept choosing it after a well-funded peer decided the model layer wasn't worth the capital.

What the divergence actually tests is the premise both companies were funded on: that "sovereign AI" is one strategy rather than two incompatible ones wearing the same label. Building a frontier model that stands up to Western labs' compute budgets is a different, and much more expensive, undertaking than building the cloud and enterprise rails an economy needs regardless of whose model runs on them. India funded both as if they were the same bet. A year on, exactly one of the two companies that took that money is still making it.

The question neither company answers

What's missing from both stories is any public accounting of how much runway each path actually needs before it either fails or becomes self-sustaining on its own terms. Krutrim can point to a profit-and-loss statement now, which settles the question for enterprise cloud economics but says nothing about whether India still has a company capable of building a frontier model if Sarvam's bet doesn't pan out — the two firms were never really competing with each other so much as hedging the same national question from opposite ends. Sarvam, for its part, has not disclosed revenue alongside its funding and valuation figures, which means the $1.5 billion price tag is a bet on the model itself, not yet evidence that anyone is paying enough for it to matter the way Krutrim's ₹3 billion already does.

That asymmetry is worth sitting with the next time either company's name comes up as proof that India has "solved" sovereign AI. Two years in, India has one profitable AI infrastructure company that no longer builds frontier models, and one well-funded frontier-model builder that hasn't yet had to prove the business works. Calling either one the answer skips the part where nobody has actually answered anything yet.