Indian IT's AI conundrum: What model to use—ready-to-build or build-from-scratch

Infosys and Tech Mahindra are building their own small AI models to help clients save costs.
Infosys and Tech Mahindra are building their own small AI models to help clients save costs.

Summary

  • Infosys, Tech Mahindra are building their own small language models to service its clients, while TCS, Wipro and HCLTech want to build on foundational AI tools.

India’s largest IT services companies are divided when it comes to selling their AI solutions.

Infosys Ltd and Tech Mahindra Ltd are building their own small AI models to help clients save costs, while Tata Consultancy Services Ltd, Wipro Ltd and HCL Technologies Ltd want to build on foundational AI tools that are already present in the market.

The size of an AI model depends on the data that is fed into them. Small AI models are trained on smaller data sets, whereas larger models, better known as large language models (LLMs), are trained on greater amounts of data. As small language models (SLMs) are trained on lesser data, they can perform lesser but more specialized functions compared with larger AI tools.

Most companies rely on internal data sets that are readily available to them, helping them build smaller models. India’s second-largest software services provider Infosys, which ended the previous fiscal with $18.6 billion in revenue, has been building small language models to service its clients.

Also read | In charts: How India’s IT giants fared in their weakest quarter

“In Generative AI, we have built four small language models for banking, IT operations, cyber security, and broadly for enterprises," said chief executive Salil Parekh in the company’s post-earnings press conference on 16 January.

The Bengaluru-based company’s management first mentioned this development in the company’s post-earnings interaction with media following its second quarter results in October last year.

“The reason for the small language model, we believe, we have some very good data sets within Infosys. And we are taking some, let us call it, clean data sets from outside the industry and so on. These then help train the small language model," said Parekh in response to a question during the conference call with analysts on 17 October 2024.

Infosys chairman Nandan Nilekani backed Parekh's comments less than a month later.

“Small language models trained on very specific data are actually quite effective . . . everybody will build models, but I think they don’t have to build these gigantic ones," said Nilekani in an interview with Financial Times published on 27 November 2024.

Also read | Q3 throws focus on deal cycle, tenure for India's biggest IT service providers

Nilekani added he was “not so sure" that companies would want to bear the high costs, the potential “black box" of data, and copyright liabilities associated with large language models behind popular applications such as OpenAI’s ChatGPT.

The small-language-model bandwagon

Infosys was joined by smaller peer Tech Mahindra in the third quarter to hop on to the SLM bandwagon.

“We have since moved from these LLMs to creating small language models and tiny language models," said Mohit Joshi, chief executive of Tech Mahindra, in the company’s post-earnings call with analysts on 17 January. “Customers really find the relevance of these models in small use cases. They (small language models) allow customers to solve fairly specific problems without using up too much of compute or carbon."

Tech Mahindra, India’s fifth largest software services provider, ended the previous fiscal with $6.3 billion in full-year revenue, a third of what Infosys reported.

Also read | Indian IT’s top 5 might end this fiscal year a little better than the last

Citing an example, Joshi added that the small language model could figure out a search functionality on an executive's desktop without relying on an external agent like ChatGPT.

Tata Consultancy Services, the country’s largest software services company, has built its own Gen AI model called WisdomNext, which it launched in June last year. Still, the company has not called this a large or small model.

One of its executives on condition of anonymity stated that large language models pose risks to data security.

“Biggest problem in using large AI models is that it poses data security issues. If the company is feeding sensitive data into an LLM, there are high chances of it being leaked by a third party," said a senior TCS executive on condition of anonymity.

“This is why customers prefer smaller in-house AI models because these are reasonably cheaper than larger models and are built inhouse so the data is protected," added the executive.

When small beats large

According to Abhishek Kumar, analyst at JM Financial, customers want small language models for specific problems. “SLMs are trained on a few million to a few billion parameters and might be customised for use cases," he said, adding that LLMs are trained on hundreds of billions of parameters. “Even the testing for these smaller models and price of computing is cheaper as compared with a large AI model."

Also read | Mint Primer: What if ChatGPT’s AI search engine clicks with users?

Sriram Raghavan, vice-president of IBM Research AI, had earlier explained, “They (small language models) are important because hardware requirements for AI are one of the biggest cost points for clients, so small models and fit-for-purpose models allow you to then extract return of investments more effectively."

Raghavan, in an interaction with Mint last year, added that small language models could help clients save their costs by up to 50 times for a specific use case. “My point is that we're not talking 5%, 10% in savings, we are talking a real order of magnitude difference."

For now, none of the country’s largest software service providers calls out revenue from Gen AI, which gained steam following the launch of ChatGPT in November 2022. Gen AI is known for its human-like capabilities, especially content production in audio, visual, and written form.

In contrast, Accenture Plc, which is the world’s largest software services company, reported $900 million in Gen AI revenue last fiscal. Its revenue from the new technology made up 1.4% of its total revenue of $64.9 billion.

And read | India’s generative AI startups look beyond building ChatGPT-like models

Catch all the Corporate news and Updates on Live Mint. Download The Mint News App to get Daily Market Updates & Live Business News.
more

topics

MINT SPECIALS