India, June 27 -- There is a long-standing debate playing out in public. Despite having a massive pool of engineering talent, India did not build global software product companies. That is why companies like Microsoft, Google and Meta are based in Silicon Valley and not Bengaluru. This undercurrent of anxiety is now taking a different shape: Why haven't Indians built the world's leading artificial intelligence models? "Do you ask British Airways to make aircraft?" asks Harish Mehta, co-founder of NASSCOM, founder of Onward Technologies and author of 'The Maverick Effect.' "So, why do you ask services companies to make products? The DNA of a services company and a product company are entirely different, right from the CEO down to the junior-most engineer." The structure of Indian IT companies was designed to organize, scale and manage large groups of people. When they did try to build software tools, they operated within their boundaries. Crucially, because these firms served as partners to global product giants, building rival products would have meant competing with their own ecosystem. Mehta recalls an early vision to pursue both products and services. "But we failed to build products early on because we didn't know how to write modular code," he says. "We burnt huge monies." Developing a proprietary tech product demands massive upfront capital, years of speculative research, and carries a high failure rate. Not just that, India's early ecosystem was difficult for product builders due to widespread software piracy and bureaucratic regulations. "In Silicon Valley, a company can pivot 19 times - as Instagram did. But an early Indian startup faces an immediate dead end," says Mehta. That is why India's IT pioneers focused on a different kind of innovation: IT services. This outsourcing model was a significant economic leap, but because it is a process innovation, it remains largely invisible to critics. India became a tech power because it focused on what was affordable and practical. This pragmatism built a stable social safety net and established the modern Indian middle class. Today, however, public choices suggest a shift away from that original logic. Public policy is pushing the country toward the high-stakes world of physical hardware manufacturing. Consider the India Semiconductor Mission (ISM) 2.0, where cumulative approved outlays have reached Rs.1.64 lakh crore. While policymakers view this as a path to self-reliance, physical factories remain deeply connected to global networks. To manufacture a single silicon wafer, a factory requires specialty chemicals from Japan, design software from California, advanced machinery from the Netherlands, and gases from Eastern Europe. If a single geopolitical link breaks, a highly expensive physical asset can sit idle. A similar shift is occurring in corporate strategy. At a recent Annual General Meeting, Tata Consultancy Services leadership discussed plans to deploy 500,000 AI agents within three years to handle routine workloads, alongside a visible cooling in mass campus recruitment. Markets initially assumed that replacing entry-level human workers with software agents would automatically lower costs and raise profit margins. But there is a practical mathematical side to this transition. Running enterprise-grade AI agents requires significant spending on cloud compute power and usage fees paid to platform providers. Industry estimates suggest an average cost of roughly Rs.4.5 lakh ($5,000) per agent annually. For a company deploying 500,000 agents, that translates into an ongoing technology infrastructure bill of roughly $2.5 billion paid directly to overseas providers. When asked how they plan to balance this infrastructure bill against profit margins, or why they aren't using their cash reserves to build independent models from scratch, TCS declined to comment. The reality is that running a fully automated enterprise is far more complicated and expensive than the trend suggests. Mehta points out that while venture capitalists like Vinod Khosla predicted a rapid decline in tech work when ChatGPT arrived in 2022, widespread labour collapses have not occurred. "Globally, no one knows exactly how this will play out," he said. "But existing work is not being easily replaced either." To bolster his case, he makes the point that there are still 100 million lines of legacy COBOL code running the world's infrastructure; "Even if firms can maintain $2 per line of code, it is a massive opportunity." Combined with emerging fields like robotics or space mining, he argues there is no empirical proof of a sudden human labour collapse. "All we can say is that we have not done any mass layoffs." The services engine is not disappearing; it is changing its tools. True economic maturity means recognizing that managing a reliable, flexible network is a profound strength. As we invest in new technologies, we should not abandon the culture of calculated risk management that secured our middle class in the first place....