The Sovereignty Trap

In late 2023, India began looking for a home-grown artificial intelligence champion to counter the rise of global giants such as OpenAI and Google. Sarvam AI, founded by researchers Vivek Raghavan and Pratyush Kumar, emerged as a promising candidate. Backed by $41 million from investors including Lightspeed, Peak XV and Khosla Ventures, the startup was selected by the government in April 2025 to develop a sovereign foundational model. It received 4,096 Nvidia chips, with the government bearing roughly Rs. 99 crore of its Rs. 247 crore computing bill.

The young company thus acquired the mantle of national technological ambition and substantial public support before establishing its capabilities against global competitors.

Borrowed Brains

The gap between ambition and achievement became apparent as Sarvam began releasing its models. In December 2023, it launched OpenHathi, essentially Meta’s Llama 2 model adapted for Hindi. In May 2025, it introduced Sarvam M, built on France’s Mistral Small. The model struggled to attract developers, recording just 334 downloads.

In February 2026, Sarvam attempted a more ambitious leap with Sarvam 30B and Sarvam 105B, trained from scratch. The latter adopted an architecture comparable to China's DeepSeek V3, but independent testing by Artificial Analysis gave it a score of just 18, against nearly 48 for leading global models. Its consumer chatbot, Indus, meanwhile, faced computing constraints and had attracted only 100,000 downloads by October.

India’s emergence as a global technology services hub was built on the success of companies such as TCS, Infosys and Wipro. That achievement fostered a belief that Indian programming talent could reproduce almost any technological success at a fraction of the cost. The same confidence was carried into artificial intelligence, where the economics are radically different.

In June 2023, OpenAI chief Sam Altman was asked in Delhi whether a small Indian team with $10 million could build a foundational model. His dismissive response provoked considerable indignation. Yet, stripped of its tone, it pointed to a fundamental economic reality: developing frontier AI requires resources on a scale that ordinary software development does not.

Training foundational models demands enormous computing capacity, reliable electricity, specialised chips and highly skilled researchers. Frontier models can require clusters of tens of thousands of advanced graphics processing units. Sarvam’s 4,096 chips, substantial by Indian standards, remain modest beside the computing infrastructure available to leading American and Chinese laboratories.

The financial asymmetry is equally stark. In 2024, China and the United States each spent around $1 trillion on research and development when measured at purchasing-power parity. India’s expenditure was approximately $30 billion at market exchange rates. The comparison underlines the enormous resource gap confronting Indian companies attempting to compete at the technological frontier.

Expecting a relatively young startup to match corporations with vast research budgets and computing infrastructure is to underestimate the economics of the AI revolution.

Costly Ambition

Part of the problem is the stigma attached to the word ‘wrapper’ in Silicon Valley. The term is often used dismissively for companies that build user-friendly applications around an artificial intelligence model developed elsewhere. Indian founders, anxious to establish technological independence, have consequently felt compelled to claim that they are building foundational models from scratch.

But there is little reason for India to regard the application layer as an inferior technological pursuit. Indeed, it may offer a more sustainable route to commercial success. The Transformer architecture underpinning modern language models originated in a Google research paper published openly in 2017. Its subsequent development demonstrates how technological breakthroughs can generate value far beyond the institutions that first produce them.

Foundational models also depreciate rapidly as successive generations arrive. A model that commands attention today can become obsolete within months. Indian companies can exploit this cycle by adapting open-source models and building specialised applications around them. This is the approach that helped create India's formidable software services industry, projected to reach $315 billion in revenue by March 2026.

The real opportunity lies in solving problems that global businesses cannot address simply by plugging a public AI model into their existing systems.

Large corporations want AI to predict supply-chain disruptions, automate customer support and detect financial fraud. But deploying public models directly within sensitive corporate networks raises concerns about data security, confidentiality and regulatory compliance. Companies that can build secure software layers connecting powerful AI models to legacy systems, while meeting industry-specific requirements, can create considerable value.

HCLTech’s $150 million investment in Sarvam illustrates this commercial direction. The partnership is intended to take its voice agents and language technologies to banks, insurance companies and other enterprises. Indian developers can also build voice-based applications for the country's linguistic diversity, serving citizens across the 22 constitutionally recognised languages and integrating with established digital infrastructure such as the Unified Payments Interface.

India need not win the foundational-model race to become a significant AI power. It can build the systems that make artificial intelligence useful, accessible and commercially viable across industries and populations.

The Sarvam experience offers a lesson in the distinction between technological ambition and technological strategy. National pride cannot substitute for computing power, research expenditure or sustained scientific investment. Nor should a promising startup be burdened with expectations that only the world's wealthiest technology ecosystems can realistically meet.

India should continue investing in indigenous research and foundational capabilities, but it must distinguish strategic autonomy from the pursuit of technological prestige. Building practical applications, enterprise workflows, secure integrations and multilingual tools is not an admission of defeat. It is a commercially viable strategy in its own right.

Rather than spend billions trying to win a hardware-intensive contest against better-funded rivals, India can leverage its engineering talent to make AI work for businesses and citizens. The route to technological sovereignty may lie less in building every component of the machine than in mastering how to use it.

(The writer is a strategy and transformation leader. Views personal)