The India AI Mission: Whose Sovereignty, and Whose Subsidy?

The India AI Mission — IndiaAI, as the government brands it — is the largest deliberate bet the Indian state has ever placed on a single technology it does not yet own. Cleared by the Union Cabinet in March 2024 with a reported outlay of around ₹10,372 crore spread over five years, it promises subsidised GPU compute, funding for indigenous foundation models, a national datasets platform, fellowships, skilling and application development. Those figures have already been revised more than once, and I will hedge every number in this piece for that reason. What has not changed is the shape of the wager: public money, spent largely through private providers, to buy a country a seat at a table it currently rents.

I want to take that ambition seriously before I start pulling at it, because the reflexive cynicism about state technology spending usually misses what is actually at stake. Depending entirely on someone else’s cloud and someone else’s models for your most consequential infrastructure is not independence. It is a standing arrangement in which the terms can change without your consent. And models that genuinely work in Indian languages — that handle Malayalam morphology, that do not collapse into nonsense when a farmer speaks Bhojpuri into a phone — are a real public good, not a vanity project. The question I keep returning to is narrower and harder: when the public pays for capability, what does the public end up owning?

What the money actually buys

The mission is organised into pillars, and the compute pillar is the one that has money and headlines. Its design is genuinely unusual. Rather than building state-owned data centres, the government empanels private GPU operators — reported participants include Jio, Tata Communications, Yotta and NxtGen among others — and subsidises the hourly rate that startups, researchers and academics pay to use their capacity. The subsidised price has been reported in the region of ₹65 to ₹70 per GPU-hour, a fraction of what the same silicon costs on the open market.

Deployed capacity has been reported at more than 38,000 GPUs by late 2025 and early 2026, against an original target that was far smaller, with a stated ambition of reaching six figures. Treat these numbers as directional. Empanelled capacity, contracted capacity and actually-available-to-a-student capacity are three different things, and public reporting routinely blurs them.

The foundation model pillar issued a call for proposals that reportedly drew over 500 applications, from which a shortlist of roughly a dozen large language model teams was selected, later expanded with a set of smaller models. The IIT Bombay-led BharatGen consortium has been reported as the largest single recipient, at a figure in the region of ₹900 crore to ₹1,000 crore. Sarvam AI and Gnani.ai are among the other names. The first models were shown publicly around the India AI Impact Summit in New Delhi in February 2026 — the summit that also produced the New Delhi Declaration and a flurry of investment pledges reported in the hundreds of billions of dollars across the AI stack.

Alongside these sit AIKosh, a datasets platform meant to pool public and anonymised data so builders are not scraping the internet from scratch, and a fellowship and skilling programme. Those last two are where the story gets uncomfortable. A parliamentary standing committee reported that the mission spent only around 32 per cent of its allocated funds in FY 2025-26, that fellowship uptake was a small fraction of the target — a reported 150 selections against 5,000 places — and that the FY 2026-27 allocation was subsequently cut by roughly half. Announce, underdeliver, re-announce is a familiar rhythm in Indian technology policy, and the mission has not escaped it.

The case for doing this at all

Set the execution aside for a moment. The strategic logic holds up better than critics allow.

India runs an enormous amount of consequential software on infrastructure it does not control. Payments, identity, welfare delivery, banking, health records — much of it sits on three foreign hyperscalers. Adding a model layer on top, where the reasoning happens and the judgements are made, deepens that dependency at exactly the point where it matters most. A country that cannot train its own systems cannot audit them, cannot set their defaults, and cannot refuse an outside party’s terms without breaking something essential. That is not an abstraction; it is the ordinary risk of building a state’s nervous system on a rented substrate. I have argued this at greater length in writing about sovereign AI, and India’s version of the argument is among the more coherent ones.

The language case is stronger still. Commercial frontier labs optimise for the languages where their revenue is. Indian languages are a rounding error in that calculus — vast in speakers, thin in high-quality digital text, expensive to serve well. Left entirely to the market, hundreds of millions of people get a second-class version of a technology that is rapidly becoming infrastructure. Public money correcting that is exactly what public money is for. The same is true of the datasets work: a shared, well-governed pool of Indian data is a genuine commons, and it is the kind of thing no single private firm has an incentive to build.

The strongest argument for the mission is also the sharpest question about it: if this capability is a public good, why is it being built as private property?

Who owns what we paid for

Here is the structural problem. A compute subsidy does not build public capacity. It buys utilisation for privately owned assets at a publicly guaranteed price.

Consider what the arrangement looks like from the provider’s side. You have purchased a large quantity of expensive, rapidly depreciating hardware. Your commercial risk is idle time. The state then offers to underwrite a stream of demand at a rate it tops up, filling your racks with government-sponsored workloads. Your utilisation rises, your depreciation schedule is de-risked, and at the end of the contract you still own every chip. The public has bought GPU-hours. It has not bought a GPU.

This is not a scandal; it is a design choice, and it has real advantages. The state avoids operating data centres it has no competence to run, capacity comes online faster than any public procurement could manage, and the hardware refresh risk stays with the party best placed to bear it. But it means the durable asset accumulates on private balance sheets while the public holds a consumption receipt. When the subsidy ends, the compute does not become ours. It becomes theirs to price.

The model funding raises the same question in a different register. If a consortium receives several hundred crore of public money to train a foundation model, what obligations follow? Are the weights released openly? Under what licence? Are the training datasets documented and reusable? Can a competing Indian startup build on the output, or does the public grant simply establish an incumbent with a permanent head start? These are answerable questions, and the answers should be conditions written into the grant rather than goodwill hoped for afterwards. The DeepSeek and open-source AI episode demonstrated how much strategic value open weights can create for a national ecosystem — and how quickly a closed release forecloses it.

This is the mechanism I have described elsewhere as how technology gets captured: not through conspiracy, but through the ordinary drift by which socialised costs and privatised returns become the default arrangement, one reasonable-sounding contract at a time. Public subsidy without public conditions is simply the fastest way to fund a concentrated market.

Sovereignty, on imported silicon

There is a further limit that the rhetoric tends to skip past. Almost every accelerator in those 38,000-plus GPUs was designed in the United States and fabricated in Taiwan or South Korea. India’s semiconductor programme is moving — the first fabrication and packaging projects under the India Semiconductor Mission were reported to be entering production in early 2026, and a second phase was announced in the February 2026 budget — but nothing in that pipeline produces frontier AI training chips. India also signed on to the US-led Pax Silica arrangement at the summit, which is a sensible hedge on supply chains and simultaneously an admission of where the leverage sits.

So the honest description is this: India is building a national AI capability on hardware it must import, using export-control regimes it does not set, from vendors whose allocation decisions are made elsewhere. That is still meaningfully better than having nothing. It is not sovereignty. Calling it sovereignty makes it harder to see the parts that remain hostage.

And the physical footprint arrives with its own politics. Data centres need land, water and enormous quantities of power, and the communities that supply them are increasingly unwilling to be treated as an externality — as the data centre protests in India have made plain. A compute strategy that does not account for who pays that bill is not a complete strategy.

Compute is the visible half of AI capability. Data quality, research depth and state competence are the invisible half — and they are where the money is not going.

The foundations nobody photographs

GPU counts are legible. A minister can announce them; a headline can carry them. That legibility is itself a distortion, because it pulls money toward the thing that photographs well and away from the things that determine whether any of it works.

Indian-language data at genuine quality and scale — curated, rights-cleared, dialect-aware, annotated by people paid properly — is slow, unglamorous, labour-intensive work. So is building research capacity: funded PhD positions, laboratories that can retain people against private salaries, institutions where someone can spend six years on a hard problem. So is public-sector capability: civil servants and technical staff inside government who can specify a contract, evaluate a model, and tell a vendor no. Underspending on fellowships while compute capacity races ahead is precisely the wrong ratio, and the reported budget cut compounds it rather than correcting it.

A country can rent compute. It cannot rent a research culture, and it certainly cannot rent the institutional competence to govern what it has bought.

What conditions should attach

I am not arguing the mission should be abandoned. I am arguing that publicly funded gains should carry public conditions, and that these should be specific enough to be enforced.

• Equity or access for subsidy. If the state underwrites utilisation of private compute, it should hold something durable in return — a standing reserved allocation at cost, a public stake, or price ceilings that survive the subsidy period.
• Open weights as the default for publicly funded models. Not a preference stated in a press release. A licence condition in the grant, with the burden of justification on anyone seeking an exception.
• Documented, reusable data. Training corpora built with public money should be catalogued and made available to other Indian builders, with provenance and consent recorded.
• Published allocation data. Who received subsidised GPU-hours, at what price, for what. Without this, no one outside the process can tell whether the mission widened access or concentrated it.
• Real spend on the unglamorous pillars. Ring-fenced funding for data, fellowships and institutional research capacity that cannot be quietly reallocated to hardware when a summit approaches.

None of this is exotic. It is the ordinary discipline applied to any public investment in an industry with a tendency toward concentration — and AI’s tendency toward concentration is not a tendency, it is a structural fact.

The question underneath all of it is the same one I would ask of any state funding a frontier technology: who ends up owning the thing the public paid to create? India has answered the first half of that question with real money and real ambition, and I would rather it made this bet than stood aside. But the second half is still open, and it will close quietly — in contract terms, licence clauses and allocation rules that nobody reports on — long before anyone notices it has been decided. That is the moment worth paying attention to. Not the summit. The paperwork.