At an AI server launch in New Delhi on a muggy afternoon in late March, Sushil Pal, a mid-level bureaucrat, made a statement that most people in the room had likely been considering for months. AI servers were not covered by India’s flagship hardware subsidy, which is being hailed as the foundation of its AI aspirations. Laptops were covered. tablets. all-in-one computers. The hardware that was important in 2019. Not the type that uses a frontier model in 2026.
MeitY had discreetly announced a policy review of the ₹17,000 crore PLI 2.0 scheme by the following morning. No press conference was held. No shiny slide show. Just the gradual realization by trade journals that India’s AI hardware push’s design had outdated the technology it was attempting to emulate.
To put it simply, price is the trigger. Due primarily to the doubling of HBM memory, Nvidia’s Blackwell GPUs have increased by 15 to 23 percent this year. The memory shortage is now anticipated to persist well into 2027. As part of its independent AI mission, India had pledged to deploy about 58,000 GPUs. The invoice no longer uses the same math as the cabinet note. Yotta, the Hiranandani group’s data center division, alone ordered nearly $1 billion worth of Nvidia chips. Meanwhile, Jensen Huang’s company signed agreements with Reliance and Tata. The rate at which the nation is purchasing computers is nearly faster than the policy that underpins it.
Observing this, it seemed as though India was torn between two narratives it was telling itself. One was the self-reliance pitch, Atma Nirbhar Bharat, in which Vaishnaw pledged to develop an indigenous GPU in four years. The other was the global pitch, which included a $200 billion infrastructure goal, zero taxes for foreign cloud providers until 2047, and tens of billions of dollars from Google and Microsoft. To increase domestic capacity, the first story requires subsidies. The second narrative continues to make those subsidies more affordable to disregard.
The government-built AI supercomputer in Pune, AIRAWAT, has a symbolic significance that is difficult to ignore. 656 GPUs in total. Clusters larger than fifteen times that size are used by Meta and Microsoft to train their models. India is not a generation behind. It is an order of magnitude behind.
Recalibrating where India can realistically hold ground in the supply chain appears to be the main goal of the rethink. One such location is chip design; although they primarily work for Intel, AMD, and Nvidia, about a fifth of the world’s chip designers reside in Bengaluru, Hyderabad, and Noida. Another is system integration. Given India’s land, power, and Chandra-level talent pool, data centers make up a third. The most ambitious objective—and perhaps the wrong one to pursue—remains the manufacturing of the GPU itself, which every minister frequently brings up in speeches.
The implication is more difficult to understand for global supply chains. India can become a useful middle node, where Nvidia silicon is assembled, racked, cooled, and rented out to half of South and Southeast Asia, if it shifts its incentives toward AI servers and design IP rather than legacy chip fabs. In the context of Modi’s speeches, that is not leadership. However, it’s also not nothing. It appears that investors are more interested in the plumbing than the model. Ten months later and a few billion dollars wiser, India may be adopting that perspective.
Whether the updated PLI will be bold or just more organized is still up in the air. In the upcoming weeks, stakeholder consultations are planned. Until then, the nation continues to build data centers faster than the grid can keep up, purchase GPUs that it did not budget for, and wait to see what level of sovereignty it can truly afford.
