In January 2024, Krutrim became India's fastest AI unicorn on the promise of a sovereign stack: an Indic foundation model, custom silicon, the whole vertically integrated dream. In mid-2026 the company made its pivot explicit — foundation model work and the custom-chip programme paused, the business refocused on AI cloud.

That is the single most informative sentence in Indian AI this year, and almost nobody framed it as good news.

The retreat that isn't one

The instinct is to read Krutrim's narrowing as failure, and to a degree it is: the original pitch was bigger than the delivered product. But look at what the company kept. It kept the part that has customers, contracts and a price per hour. It dropped the two parts that require competing head-on with firms spending more on a single training run than most Indian AI startups have raised in total.

Building a frontier model is a capital contest India is structurally not set up to win right now, and pretending otherwise burns years. Renting the compute that everyone else's models run on is a business with visible unit economics from day one. One of those is a moonshot. The other is infrastructure — and infrastructure is what India has historically been very good at monetising.

The Sarvam AI round tells the same story from the other end. In June, Sarvam closed a $234 million Series B at a $1.5 billion valuation. The notable line item wasn't the number, it was the buyer: HCLTech took a stake of roughly 10.46 per cent for about $150 million.

A listed IT services company putting nine figures into a model startup is not a venture bet. It is a distribution deal wearing a cap table. HCLTech does not need another research lab; it needs something Indic, sovereign-compliant and differentiated to put in front of enterprise and government buyers who are being asked to keep data onshore. Sarvam gets a customer with a sales force. HCLTech gets an answer to a question its clients keep asking.

Three things changed at once

  • Public markets got a price. Fractal Analytics completed its listing this year, giving India's AI-adjacent sector something it has never had: a quoted comparable. Private valuations argue; a share price settles. Every subsequent round in the space now gets marked against a number the market sets daily rather than one a term sheet asserts.
  • The buyer became the enterprise, not the consumer. IBM has put active AI deployment at 59 per cent of enterprise-scale organisations in India in 2026, and around 83 per cent of global capability centres in the country as engaging with generative AI. GCCs are the quiet giant here — they are cost centres with mandates to prove AI savings, and they buy tooling and capacity, not chatbots.
  • The services trade found its footing. Infosys raising its FY26 revenue forecast on the back of AI-led partnerships was the first hard evidence that the $283 billion IT industry's AI exposure might be a tailwind rather than the disruption thesis everyone assumed. If AI collapses the billable hour, Indian IT is in trouble. If AI is a migration project of the kind Indian IT has run twenty times, it is the best decade in a while. The market is currently voting for migration.

What the forecasts actually claim

Two numbers get quoted constantly and they measure different things. BCG's projection puts India's AI market above $17 billion by 2027. IDC's puts Indian AI and GenAI spending at about $6 billion by 2027, growing at a 33.7 per cent compound rate. Both can be right — one counts a market including services and downstream value, the other counts direct outlay. Anyone using them interchangeably is not reading either.

The government layer sits underneath all of it. The IndiaAI Mission has been the demand signal that made sovereign-compute pitches financeable in the first place, and deals like RailTel's five-year tie-up with Microsoft for cloud and AI work across the railways show what the actual pipeline looks like: large, slow, public-sector, infrastructure-shaped.

The uncomfortable version

Here is the part the ecosystem's own marketing avoids. A country that supplies compute, integration and services to models built elsewhere captures real revenue and very little of the durable advantage. The margin on GPU hours compresses the moment supply catches demand. The margin on an integration contract has a well-known ceiling — India has spent thirty years discovering exactly where it sits.

The counter-argument is that this is how it always starts, and that a services and infrastructure base is what funds the eventual attempt at the harder thing. That was broadly the semiconductor story in Taiwan, and it took decades rather than funding rounds.

Both readings are defensible. What is no longer defensible is the 2024 framing — that India would produce a competitive general-purpose foundation model on venture timelines. The two companies that pitched hardest on that premise have now, in different ways, told you what they learned.

Company financials, valuations and stake sizes as reported. Adoption and spending figures are third-party research estimates, not audited data.