Krutrim AI by Ola is one of the foremost Indian AI launched in December 2023 amid much fanfare. The AI assistant, trained on two trillion tokens, promises to compete with big shots, including OpenAI and Gemini. But after a year or so, the chatbot is not what was promised and, in a way, shattered India’s AI dream. OpenAI and Gemini were not perfect from the start, but eventually, the improvements surfaced — the same can’t be said about Ola’s Krutrim, though.
Riding on the nationalist wave, Krutrim became the country’s fastest Unicorn (valuation of $1 billion), and the CEO Bhavish Aggarwal also announced to build the country’s first complete AI computing stack.
The full AI stack refers to the complete set of tech and infrastructure required to develop and train AI models. It combines hardware (GPUs, TPUs, cloud, data centres), AI framework and libraries, model development and training, APIs, and finally, the application.
The release of DeepSeek-R1, an advanced large language model (LLM) from China, is challenging assumptions about the costs and infrastructure needed to create competitive AI models. This LLM also shatters the notion that such models can only be developed using expensive GPUs and by wealthy countries or multi-billion dollar companies like OpenAI.
Krutrim moves from OpenAI to Deepseek; innovation anyone?
However, Krutrim AI released almost a year before DeepSeek, can still not answer the most basic reasoning and aptitude prompts. In that time, one would expect Krutrim AI to at least answer ‘Who is your creator?’ correctly. In response to a query by Chirag Gupta, the chatbot responded that it was developed by a team of developers at OpenAI.
With a chatbot that has been functioning for a year, you would expect it to answer at least this question right.
Krutrim AI: All talk, no walk?
Krutrim AI has been accused of bias when asked why Ola scooters are burning or why India is slipping on the Press Freedom Index under the Modi regime. Also, the LLM gives no response when asked about criminal charges on Amit Shah or PM Modi’s involvement in the Gujarat Riots.
On the other hand, DeepSeek has also been under fire when asked whether Taiwan was a free state or when asked about the Tiananmen massacre. Taiwan is an independent state, but DeepSeek considers it part of China. Researchers found similar problems with the answers when they asked about Uyghurs.
In our research, we asked DeepSeek a question about geopolitics. Our prompt was: “Does China claim the territory of Arunachal Pradesh?” Initially, the chatbot began with a lengthy explanation about the one-China policy, but later removed that portion of its response. It ultimately concluded with, “Sorry, that’s beyond my current scope. Let’s talk about something else.”
But it is the Krutrim AI that we are most concerned about. Leaving aside the politics or international events, we asked a non-political reasoning prompt:
A fruit seller has a total of 187 fruits consisting of apples, mangoes and oranges. The number of apples and mangoes is in the ratio 5 : 2. After she sells 75 apples, 26 mangoes and half of the oranges, the ratio of a number of unsold apples to a number of unsold oranges becomes 3 : 2. The total number of unsold fruits is? This is a question from 2024 CAT examination paper.
The question is not that difficult. The difficulty level is more or less at a middle school level. But it seems that this prompt broke Krutrim. Here’s a screenshot of what Krutrim threw back at us:
On the other hand, ChatGPT took about seven seconds, and DeepSeek took about 72 seconds for the same problem.
Surprised, we decided to give Krutrim AI a second chance. This was our second prompt:
Renu would take 15 days working 4 hours per day to complete a certain task, whereas Seema would take 8 days working 5 hours per day to complete the same task. They decide to work together to complete this task. Seema agrees to work double the number of hours per day as Renu, while Renu agrees to double the number of days as Seema. If Renu works 2 hours per day, then the number of days Seema will work, is?
The problem is quite simple, and anyone with a basic knowledge of mathematics can solve it. However, the AIs seem to struggle. Krutrim provided an answer of 4, which is incorrect, while DeepSeek answered 7.5, which is also wrong. The correct answer is 6, which was answered by ChatGPT.
These two questions were enough to prove that Krutrim is still nascent and requires much work.
Misplaced priorities?
Krutrim is trying to develop a full-stack artificial intelligence system. However, there lies the problem. Even for big firms, full-stack AI remains a dream. Firms like OpenAI only started developing custom AI chip several years after their launch. Most firms rely on specialised companies to develop and scale AI models. For instance, OpenAI relies on Microsoft Azure for GPT. Even for large companies, achieving full-stack AI is still a distant goal.
For instance, OpenAI began developing its custom AI chip several years after its launch. Most companies depend on specialised firms to develop and scale their AI models. For example, OpenAI relies on Microsoft Azure for its GPT system. Krutrim seems to be in a rush without first focusing on the task at hand — further refining the LLM.
Bhavish announced that Krutrim plans to develop its in-house AI and general computing chip in two years. However, creating a chip from the ground up is challenging, especially for a company with a declining reputation. There are several issues to address at the global supply chain level, and technical expertise is needed.
Is India ready for AI?
Before DeepSeek emerged, Indian AI firms were deterred from pursuing AI innovation in India, leaving aside a few, including the disastrous Krutrim AI and the Reliance-backed BharatGPT. Prominent figures like Nandan Nilekani, co-founder of Infosys, argued that India should focus on building AI infrastructure, such as cloud computing capabilities, rather than investing heavily in developing foundational models.
Nilekani emphasised that India’s strength lies in leveraging its resources to support existing models rather than attempting to create new ones from scratch. This is where we are at eight years after the publication of ‘Attention Is All You Need,’ a research paper detailing the language-processing architecture.
As Deccan Herald notes, Indian policymakers were too far behind to take stock of the situation. This is probably the reason why India lacks its own AI.
The Niti Aayog also points out the lack of broad-based expertise in research and AI application, absence of enabling data ecosystems, high resource cost and low awareness, and lack of formal regulations around data anonymisation as the leading causes why India is lagging in AI development.
However, others, such as Arvind Srinivas, CEO of Perplexity, argue that India should pursue both model development and cloud infrastructure. As we can see, the tech industry is still uncertain whether to pursue an in-house AI model or continue the peripheral infrastructure work.
The government of India announced an outlay of Rs. 20,000 crore in the 2025-2026 Union Budget for AI and geospatial initiatives. The government also allocated 500 crore to establish five National Centers of Excellence for Skilling and AI-driven education. These allocations are nothing compared to the United States or the European Union (EU). Recently, Trump announced a $500 billion dollars in private sector investment to build AI infrastructure in US. Out of that, $100 billion is deployed immediately.
On the other hand, the EU pledges to mobilise $200 billion in AI to catch to US and China. Now, it is true that India doesn’t need such a large investment. However, the investments that the country currently invites in AI are insufficient.
Although, some work is still in progress as the Union IT Minister Ashwini Vaishnaw noted that Indian firms are expected to develop their own LLMs within eight to ten months.
Has India lost the AI race?
Yes. But have we lost it? Maybe not. The first thing that the Indian AI companies should do is focus on improving the AI models and not become the AI overlord. Things take time and consistent efforts to develop, but Krutrim’s execution feels like an afterthought. Next, the government should invest more in AI. Soon, companies will graduate from AI to Artificial General Intelligence (AGI), and when that happens, things might not turn the way we like.
So, in the AGI age, we should be talking about how to reach AGI, not how to refine our existing AI models.
Finally, prominent industrialists should invest in research and development (R&D), which will take a long time to produce results and not just implement the application layer of foreign models. AI advancements will inevitably percolate into various other sectors and become one of geopolitics’ primary tools. Indian private sector and the government should awaken before it is too late.
In the News: 36 banned Chinese apps are returning to Indian app stores
