Inside Ola's not-so-electric ride with Krutrim AI

Ola’s models have been trained over a humungous two trillion tokens of data for Indian languages. But despite this, there are concerns about the maturity and market-readiness of Krutrim’s products—particularly about the ability of its AI models to handle complex queries.
 
The concerns are not unfounded. Krutrim’s AI chatbot admitted to shortcomings (Forbes India has screenshots) during our interaction in March. When Forbes India posed two questions—related to its origins and evaluation of its performance—the responses were underwhelming. The chatbot initially struggled to provide accurate and prompt answers. And when rephrased and asked multiple times, they lacked clarity and quality. 

To the first query about its creator, the chatbot gave an overly detailed and introspective response, explaining its thought process and decision-making. Its answer to the next question about its performance and reviews was candid, but somewhat negative, highlighting inconsistencies and inaccuracies in its operation.

Krutrim refused to share the traffic numbers for its chat assistant that is free to use, but said there will be significant iterations in the weeks to come. The app will undergo a complete transformation, offering a much-enhanced experience. “We’ll be rolling out major updates. This will be the first app of its kind in India, addressing a significant need in the market,” adds Khatri.

Promises & Plan

The startup was founded in 2023 following which it announced a slew of products and developments. It claims to have launched India’s first large language model (LLM) and later revealed its ambition to build a full-stack AI solution from scratch—designing custom semiconductor chips, providing cloud infrastructure and delivering LLMs. This plan mirrors that of Google, which uses its data to train its algorithms, running on its chips, deployed on its cloud.


Questions are now being raised about the accuracy of Krutrim’s performance claims compared to other LLMs. Since its launch, the Indian AI startup has released two large LLMs, Krutrim-1 and Krutrim-2, with seven billion and 12 billion parameters, respectively. Krutrim open-sourced them in February. It is also hosting DeepSeek R1, a 671-billion-parameter AI model, on its platform.

The company is now working on Krutrim 3, a 700-billion-parameter LLM. It has partnered with Nvidia to deploy GB200 superchips and with Lenovo to develop India’s largest supercomputer, which will be operational by the end of the year.

According to Khatri, if the company focussed only on AI, it would be playing catch-up with companies like OpenAI, Google and DeepSeek, which have been at the forefront of AI research and development for decades. India is a price-sensitive market and competing over cost is a challenge. He believes that with a full-vertical integration—owning the cloud, chips and AI models—they can beat competition.  

Exit Route

It’s easier said than done though. The company has to deal with a lot of challenges, especially following the exits of around a dozen senior and mid-level managers in recent times, casting a shadow over its ambitious plans.


Vipul Shah, vice president of products, Gautam Bhargava, vice president and head of AI engineering, Samrat Saha, director of machine learning, Achal Kumar Mall, director of product management, Mikhil Raj, senior director of product management, and Ashok Jagannathan, vice president, are among those who have quit. Ravi Jain, business head of Krutrim, also put in his papers after a three-year stint at the end of last year. Shah, Bhargava and Jain were part of the senior management team, while the others held junior to mid-management roles.

Khatri tries to assuage any fears by saying Krutrim is expanding, with hiring currently underway in Palo Alto, Singapore and Bengaluru. “Although we didn't have an early start like some others, we’re making rapid progress. You’ll see innovative applications from Krutrim that set a high bar, comparable to industry leaders like OpenAI,” he says. While he confesses that Krutrim is technologically not yet on par with the leaders—Google, Microsoft, OpenAI—he points out that the competitors have had a 10-year headstart. “We’re taking a strategic approach… it will take time, but the results will be transformative,” he says.

Independent Stance

A few weeks ago, Ola founder Aggarwal announced a fresh Rs2,000-crore investment in Krutrim via equity and debt, with a commitment to invest an additional Rs10,000 crore by next year. So far, the Bengaluru-based startup has secured a total funding of Rs2,320 crore ($280 million) over four rounds from backers such as Z47 (earlier Matrix Partners), the Sarin family and Jiten Vaidya, co-founder of PlanetScale, a database firm for developers.


Aggarwal has pledged an additional 5.88 crore shares of its newly listed Ola Electric to secure funding for the Krutrim Data Centre. This comes after he had pledged 1.1 percent of Ola Electric’s shares in November 2024. With the new pledge, the Krutrim founder has committed a total of 10.72 crore shares, equivalent to 2.43 percent of Ola Electric’s total shares and 8.09 percent of his personal stake.

Krutrim has launched data centre operations in Bengaluru and Hyderabad to support its AI ambitions. It plans to increase its data centre capacity, aiming to reach 1 GW by 2028 through quarterly expansions. 

Regarding the claim of building AI models from scratch, Ola has positioned itself as developing foundational models independently. Despite rumours suggesting a reliance on fine-tuned models like Meta’s Llama-2 or Google’s Gemini, the company maintains that its models are original creations.

Developing its own models will enable Krutrim to tailor them to its specific needs and requirements, ensuring better performance and adaptability. And apart from helping differentiate itself from competitors, it also allows the company to maintain control over its data, reducing reliance on third-party models and minimising potential data security risks.

“Our goal isn’t to burn billions of dollars like OpenAI or Anthropic. Instead, we focus on efficient training. For example, our 12 billion-parameter Krutrim 2 model performs as well as Llama 70 billion, GPT-4, or Gemini, which are 10 times larger, on Indian use cases,” claims Khatri. While Krutrim may be slightly behind in English-language tasks, it’s on par when it comes to Indian languages, he adds.

The company developed and utilised its own cloud infrastructure to train Krutrim 2. As a result, it demonstrated India’s ability to build and deploy its own cloud infrastructure for AI research, reducing dependence on foreign technologies.

Aggarwal has criticised Google and Microsoft for not subsidising cloud costs for Indian developers. He says Krutrim would be the most affordable platform, challenging global hyperscalers. Experts argue that designing custom AI chips, training models, building infrastructure, and delivering a developer platform is a complex task that requires vast expertise, resources and capital expenditure.

The Road Ahead

Krutrim is eyeing to tap both the B2B and B2C markets. For the latter, it is integrating AI into apps like maps, electric vehicles and ride-hailing. For B2B, it offers cloud services and APIs, used by many companies for speech, voice and other models. “Our fully vertical integration lets us serve both markets, much like Google—with Ola as our biggest customer,” says Khatri.

Ola Electric, though, has been facing a bumpy ride that includes raids on its stores in Madhya Pradesh and Maharashtra, and discrepancies in its sales figures. The company is struggling with mounting losses, falling sales, and dwindling stock prices, amid concerns over its deliveries and after-sales services.

The biggest roadblock Krutrim faces right now is training efficiently, especially considering India’s unique context. It refers to the process of optimising the training of Krutrim’s AI models to achieve better performance, accuracy and scalability while minimising resources.  

With most Indian data stored on pricey public clouds like Amazon Web Services and Google Cloud Platform, affordability is a major hurdle. However, Krutrim achieved a breakthrough by training Krutrim 2 on its own cloud, showcasing India’s potential for indigenous cloud infrastructure. It assumes significance because it reduces dependence on foreign cloud services and paves the way for more affordable and accessible cloud computing solutions in India. 

Rajinder Balaraman, MD of Z47 (formerly Matrix Partners India) and an early investor in the Ola Group, emphasises that India must adopt a comprehensive approach to AI innovation. “You can’t just focus on one aspect of the problem. We need to innovate across the entire stack,” he says.

India faces unique hurdles in its AI journey, including limited access to advanced chips. This uncertainty highlights the need for a full-stack AI company in India. To innovate, it also needs to reduce the cost of AI. While DeepSeek has made progress in model design and algorithms, India requires a more comprehensive AI and chip ecosystem. “The emergence of a domestic AI and chip ecosystem is vital for India’s growth. This is why we need to play across the entire stack,” adds Balaraman.

Krutrim building an AI stack, from chip to cloud, reflects a new wave of AI inference at scale, says Winston Ma, an adjunct professor at New York University’s School of Law. This new computing is necessary to unleash the agentic wave of applications and new levels of intelligence. “Krutrim will need huge amounts of capital to build up Nvidia GPU clusters, and it must have expertise in the interconnectivity of the full stack, because networking is the next bottleneck after the data centres themselves,” adds Ma.

As Krutrim continues to advance, the question remains: Will it be able to deliver India’s own AI model—one that can rival the global leaders?