In August, the IndiaAI Mission's shared computing pool crossed 45,000 GPUs, up from a standing start of 10,000 barely two years ago. Along the way, the government has sanctioned more than 93 lakh subsidised GPU-hours across 237 projects, and identified 20 indigenous foundation models for state backing, including Sarvam AI's 30 billion and 105 billion parameter models and BharatGen's multilingual work out of IIT Bombay.
Those are good numbers. But they are not, on their own, the interesting part of this story. The interesting part is how much of India's actual AI potential still sits ahead of these numbers rather than inside them.
What's already working
Start with what the mission has genuinely got right. It has made compute a public utility rather than a private luxury. H100 access through empanelled providers comes well below commercial hyperscaler rates, which means a research lab in Pune or a two-person startup in Kochi can now run experiments that used to require a Bay Area seed round. That is a real change in who gets to build AI in this country, not just a change in who gets to talk about it.
The choice of what to build with that compute has also been sound. Rather than chasing a single giant general model, the mission backed a spread of mid-sized, India-specific systems: Gnani.AI's speech-to-speech engine, Avataar's video generation model, and language models built for the reality that India runs on dozens of languages that most global AI labs have never seriously trained for. That specificity is not a consolation prize. It is arguably India's clearest opening, because no lab in San Francisco or Beijing is going to out-build India on Indian languages.
The potential still on the table
Here is where it gets more interesting. The IndiaAI Mission has been running in parallel with a much larger, slower-moving project: Semicon India 2.0, a ₹1.27 lakh crore push to build actual chip manufacturing on Indian soil. Twelve fabrication and packaging projects are already approved across six states. The Tata Electronics fab in Dholera, India's first serious wafer fabrication facility, is targeting first silicon by the end of this year.
Put those two programmes side by side and a bigger picture appears. Right now, India rents its compute, buying GPUs designed and fabricated elsewhere. But a country that can design its own chips, package them domestically, and train its own models on its own languages is playing a different, longer game than a country simply renting cloud time. That combination, sovereign compute plus sovereign silicon plus a billion-plus native speakers of languages the rest of the world has largely ignored, is not something any other AI power can simply copy. It has to be built here, or not at all.
What's genuinely missing
None of this happens automatically, and it is worth being honest about the gaps.
India's own semiconductor push is still years away from producing the advanced logic chips that actually power AI training, the kind TSMC and Samsung make at the leading edge. Domestic fabs will handle packaging, memory, and mid-range logic long before they touch a GPU-class chip. That gap will not close this decade without sustained investment and, more importantly, sustained patience from policymakers who will be tempted to declare victory early.
There is also a translation gap between prototype and product. The mission has produced 62 prototypes and dozens of hackathon-born solutions, but turning a promising demo into something a hospital or a farmer collective actually uses every day is a different kind of work, and it needs a different kind of funding than a compute subsidy.
And the compute build-out itself needs private capital riding alongside public money, not trailing behind it. Government-subsidised GPUs are a floor, not a ceiling. The countries and companies pulling ahead globally are the ones where public infrastructure attracts private billions on top of it, not the ones where the state carries the whole load alone.
How to actually get there
None of these gaps are exotic. They are the ordinary, fixable kind.
The clearest move is to treat IndiaAI compute and Semicon India as one continuous pipeline rather than two separate ministries' projects, so that GPU subsidies and chip fabrication funding are planned against the same five-year horizon instead of announced in separate press cycles. A second is building open, high-quality datasets in Indian languages and making them freely available the way AI Kosh has started to do, because the biggest constraint on Indic-language models right now is not compute, it is clean training data. A third is creating a real bridge from the AI Centres of Excellence to actual deployment budgets in state governments and public hospitals, so prototypes stop dying quietly after the demo. And a fourth is making it simple and fast for private capital, domestic and diaspora alike, to co-invest in Indian compute and chip capacity, rather than routing every serious AI bet through a subsidy application.
India did not have a credible national AI compute story two years ago. It has one now, imperfect and still small by global standards, but real. The next two years will decide whether that story stays a well-funded pilot or turns into the infrastructure a fifth of the world's population actually needs. The pieces for the better outcome are already on the table. What is left is mostly the work of putting them together.
