India’s Sovereign AI Models Run on Imported Silicon
3 min readIndia now has sovereign AI models it can point to. Sarvam AI, BharatGen, Gnani and Socket have all shipped, trained end to end on Indian soil and covering Indian languages. The awkward part sits one layer down: every GPU those models trained on was imported.
What the mission has built
The IndiaAI Mission launched in March 2024 with roughly $1.25 billion committed across seven pillars, including compute, foundation models, datasets and safety. Against 506 applications, the government selected 20 indigenous foundation model proposals for support, split between 12 large multimodal models and 8 small language models, with the intellectual property staying with the applicants rather than the state. Sarvam has released models at 30 billion and 105 billion parameters. A nine-institution academic consortium is working across all 22 scheduled languages.
The compute side has moved quickly too. Around 34,000 GPUs are deployed across Indian data centers and rented to registered startups, academics and government agencies at roughly 65 rupees per GPU-hour, a deliberately subsidised rate. Another 20,000 are being added to reach 54,000, with a stated target of 100,000 public GPUs by December 2026.
The layer nobody owns
None of that silicon is Indian. As ThePrint put it, India has built a sovereign AI model while almost everything underneath is rented. The weights, the training data and the language coverage are domestic. The accelerators, the fabrication and the supply chain that delivers them are not, which means a subsidised GPU-hour is only as sovereign as the export policy of whoever made the chip.
That is a different position from the rest of Asia. Japan is funding Noetra to build a national model on Nvidia hardware but is also backing domestic packaging and semiconductor work. China is pushing models onto Huawei Ascend silicon specifically to remove the dependency. India has so far optimised for the model and application layers, where the returns arrive fastest, and deferred the capital-intensive part.
Why it matters
The strategy is defensible. Chips take a decade and tens of billions; models and language coverage take a year or two and deliver visible results, which is why Anthropic opened a Bengaluru office with India now among Claude’s largest markets. But sovereignty measured at the model layer is a claim about who owns the weights, not about who can keep the machines running if supply tightens.
Watch for whether the December GPU target gets paired with anything on domestic fabrication or packaging. If the compute base keeps growing while remaining entirely imported, India ends up with the largest rented AI infrastructure in the region rather than a sovereign one.