The Fastest Unicorn in History

December 2023. Bhavish Aggarwal , founder of Ola, India's answer to Uber , made an announcement. He was going to build India's answer to OpenAI. Full stack. AI models, chips, compute, applications. Everything.

He called it Krutrim. Sanskrit for "artificial."

Within 45 days of the announcement, Krutrim became India's first AI unicorn and the fastest unicorn in the country's history , $50 million raised at a $1 billion valuation. The timing felt perfect. The money felt like a signal. India had its AI moment.

Less than two years later, Krutrim announced it was pausing all work on foundational models and chip design.

The Math That Never Worked

Training a frontier AI model costs money. According to Anthropic's CEO, the low end is around $100 million. Current frontier models run closer to $1 billion each. Anthropic spent $6.8 billion on compute alone in 2025.

Krutrim raised $50 million.

That gap is not a funding challenge. It is a structural impossibility. You cannot build a competitive frontier model on 5% of what the minimum viable budget requires. The people who invested early either did not do the math or believed something would change before the math caught up with them.

The Chatbot Disaster

Krutrim launched its chatbot with a specific pitch: built for Indian languages and culture. A billion-person market that global AI companies were underserving.

Within hours of launch, people noticed the chatbot appeared to have been built on ChatGPT 3.5 outputs. There were basic errors in maths, logic, and history. The team later cited a data leak issue, but the damage was already done.

Then OpenAI released GPT-4o with support for more than 50 languages, including all the major Indian languages. The "built for India" moat disappeared in a single product announcement. Investors who had backed the "Indian languages" angle started quietly reconsidering. No further outside investment came in.

The Chip Dream

Bhavish pivoted to chips. He acquired a Bengaluru chip design company called Bodhi Computing and outlined an ambitious plan: partner with ARM for the core architecture, partner with TSMC to manufacture them.

TSMC does not simply accept customers. Volume commitments are required. Cash investment upfront. Seriousness needs to be demonstrated in a language that semiconductor fabs understand , which is not press releases, it is orders.

That level of commitment never materialised. With no new outside investors and Bhavish's other company, Ola, struggling on public markets, there was no capital to make TSMC take the call seriously.

Late 2025: Krutrim announced the chip work was paused. The foundational model work was paused. The new focus: AI cloud infrastructure services. The product that requires the least proprietary technology and competes in one of the most commoditised markets in the industry.

What the Krutrim Story Actually Teaches

It is easy to read this as a cautionary tale about hubris. It is more accurate to read it as a lesson about capital requirements at the frontier.

The companies that are building foundation models , OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek , are all operating at a scale of capital that makes most other tech sectors look cheap. The minimum viable budget for competing at the frontier is not $50 million. It might not even be $500 million anymore.

Krutrim is now the fastest-folding AI unicorn in history. Not because Bhavish failed to execute. Because the game he said he was going to play requires resources that $50 million cannot buy, and the market knew it before he did.