With recent advancements in large language models and tools like ChatGPT, Indians have been pushing for the need to develop a homegrown AI model so that the country is not left behind. There has been rising pressure on the government to fund the development of LLMs that would help reduce India’s dependency on the US and China.
In April, the Centre picked Bengaluru-based startup Sarvam AI to build an indigenous LLM. The company will receive support from the government, such as GPU compute resources worth Rs 200 crore. However, funding the development of these LLM models through taxpayers’ money may not be in our best interests. These models have exponentially high maintenance cost, which keeps increasing with every new version, while offering diminishing gains in performance. Building a futuristic technology also requires a more fundamental research culture that we do not have in the first place. Not addressing these concerns first and leapfrogging to build an LLM would give us a superficial victory at best.
Rising costs, diminishing returns
If you are a regular user of ChatGPT, you must have felt that the system has not improved drastically with each update over the last 12 months. At the same time, the cost to build these models has kept increasing since 2021, with estimates that the training of ChatGPT-4o cost about $100 million. GPT 4.5 must have cost significantly higher, being the largest language model trained by OpenAI. CEO Sam Altman described this model as a “thoughtful person”, but one “who won’t crush benchmarks”.
Going by previous trends, this model must have required a few thousand crores and yet hardly offered any gains. OpenAI announced in April that they will be discontinuing this model for developers, which comes down to its unreasonably high serving cost. Most industry players, including OpenAI and Google, have not obtained their recent performance gains by training a bigger or different language model. The Indian government should be taking notes and instead invest in alternative approaches that market-leading companies may not have focused on due to intense competition and investor pressure.
All that glitters is not gold
While the Ghibli-style image trend has helped AI companies get some traction lately, they have been burning enormous cash with no return on investment in sight. Take OpenAI as an example. Since 2023, the company has raised a new funding round every 12 months. In 2023, they raised over $10 billion followed by $10.6 billion in 2024, and now $40 billion in 2025. That is over $60 billion of investor money in three years. Meta is planning to allocate a good chunk of its projected $65 billion in 2025 capital expenditures to AI, and so is Google. Does the Indian government have the vision and budget to potentially invest thousands of crores on an Indian LLM, or are we just giving token money to build a system that would never really be usable?
Also read: AI helps us get over the limits of our cognitive ability. We must embrace it
The final hurdle
It is not the capital that is holding back India in this race. Look at Krutrim, an LLM released by Ola. Despite all the funding, they are yet to produce a barely usable system compared to those developed global competitors. The fundamental requirements to come up with groundbreaking technology are exceptional talent, research calibre, visionary scientific leadership, minimal administrative hurdles, and then and only then, capital to make sure that money problems of running such an effort can be solved.
India still needs to retain its top talent—the ones who go abroad and lead AI efforts at OpenAI, Google, Meta—reduce administrative hurdles for businesses, and foster a high-quality research culture in academia, which is the foundation of such innovation. Leave it to the private markets and investors to figure out their way from there on.
Open-source AI is alive and accessible
While fears that China or the US may be left with sole access to AI sound alarming, we need to remember that most technologies around the world have been built on the back of open-sourcing, which is releasing software code for free public use. This has continued with LLMs, where players like Microsoft, Mistral AI, Meta, DeepSeek, Alibaba, and others have and continue to release their LLMs for free. These models can be downloaded easily, do not have a backdoor to any foreign governments, and are within 5-10 per cent of the performance range of commercial models. If Indian companies need to research AI, they can easily start with one of these open-source models and build their own using them. There is no need to fund the creation of a new language model and reinvent the wheel.
Also read: AI regulation gets trickier with Grok. India needs adaptive, not reactionary policies
No clear path
Let’s assume that the government goes ahead and keeps funding the development of an LLM. Then who takes care of the recurring talent and maintenance costs, which are often 20-30 per cent of a company’s budget worth thousands of crores? Why should a user prefer this model over OpenAI, Google, xAI or DeepSeek? Will this Indian model perform equally well in programming, science, and mathematics as the rest of the models? If a private company like Sarvam AI has been tasked with developing the model, why not just have private investors fund the project? The government should be easing administrative hurdles, not betting on private markets.
It’s not that we shouldn’t be pursuing AI or funding its growth. There are many ways in which state funding could be of use. The digitisation of instructional datasets of local languages would help global and local AI companies make their LLMs understand Indian languages better. Curated knowledge bases and documents that can help debias state-of-the-art AI models on Indian history, narratives, and religions can educate global and local audiences on Indian issues. There can be various applications of LLMs in governance.
For now, we need to be aware that there is no clear path for a return on investment from LLMs. And so far there are no signs that we are anywhere close to even general intelligence. Let us have the private markets play out this bet. There is no dearth of areas where taxpayers’ money can be better spent.
The author is the head of product and machine learning at Narravance, New York. He tweets @pratik_ratadiya. Views are personal.
(Edited by Aamaan Alam Khan)
I’ll be commenting more as a strategist than a technical expert, so my opinions may stem from limited technical depth in the field.
Context:
I really appreciated your take on the steps the Centre is taking to build their own LLMs. You suggested that instead of building from scratch, the government should leverage existing open-source models. You also cited Ola as an example of a company building its own LLMs. Trust me, their vision is far broader than most people realize. Ola is no longer just a cab-sharing company. it’s evolving into a Data and AI company. Let me explain why.
Let’s draw a comparison using western analogies. We often think Ola is analogous to Uber. But a more accurate analogy would be between Ola and Tesla. Why? Around 2021, Ola entered the EV market and is now working in both the EV and AI sectors seeking to build a data network effect similar to what Tesla has achieved.
But here’s the twist: unlike Tesla, Ola already operates a massive ride-hailing network, giving it access to real-time motion data at scale. This puts them in a unique position to build and optimize electric vehicle infrastructure while training AI models using fleet-level data leading to personalization, optimization, and predictive capabilities.
India’s population creates an “unfair advantage”, more users, more two-wheelers, and hence more running data to train models on. This explains why they entered two wheeler and not Four Wheeler market. If data is the new oil, why not also build the right engine (model architecture) that aligns with that oil?
If Uber doesn’t adapt to such a data-first model, it risks becoming obsolete. Ola, in contrast, is creating a flywheel: cab-sharing → EV adoption → more data → India-specific model training.
Now to your main point, why not just fine-tune open-source models instead of building new ones?
That’s where architecture comes in.
Fine-tuning open-source models like Mistral or LLaMA gives you some control but not full autonomy. You’re still bound to architectures and training pipelines built for Western contexts. For example:
Most open-source tokenizers are heavily biased toward Latin scripts.
India needs tokenizers that are optimized for Indic languages and its dialects.
Many “open” models (like LLaMA) are not licensed for commercial use, limiting redistribution, monetization, and modification rights.
If we aim to build a long-term AI ecosystem, especially one that supports India’s linguistic and cultural diversity, we need to control the foundation model’s architecture, not just fine-tune someone else’s and remain with scalability limitations.
Full control could be:
Monetizing APIs on your own terms.
Licensing the model to governments, schools, and startups.
Creating a sovereign tech stack for AI.
If a private Indian sector company like Ola recognizes this and is investing in building such infrastructure, shouldn’t the Indian government consider doing the same?
Is this really your concern ? We waste billions of dollars in freebies but this is the problem ? Seriously what a stupid article.