India can write its own AI. It cannot yet make the thing that runs it

You type a question in Marathi into Indus, the assistant that Bengaluru's Sarvam AI runs on its 105-billion-parameter model. The answer comes back in Marathi. The model was trained from scratch, in India, on compute paid for partly by the Indian government. Everything about that exchange says made here.

Now ask a different question: what did the thinking happen on?

The government has already answered it, in a single line in a March 2026 press release about its own AI mission: "GPUs are highly advanced equipment and are primarily manufactured in one country." It did not name the country. It did not need to.

Think of it as a five-storey building

Here is the one picture to hold for the rest of this piece. Indian AI is a building with five floors. At the top, where you live, are the models that answer you. Below them, the data centre halls and the public pool of rented computing inside them. Below that, the fab that prints chips. Below that, the plants that package them. And at the ground floor, holding everything up, is the accelerator: the chip that does the arithmetic.

India has built, or is building, every floor of this building. The question is which floors it owns and which it leases. We will walk up from the ground.

Ground floor: the accelerator, which India designs around but does not make

An accelerator is a chip built to do one kind of maths, the multiplication of huge grids of numbers, far faster than an ordinary processor can. A GPU (graphics processing unit) is the most common type. Large language models are trained on thousands of them working at once.

Look at the price list of the government's own GPU pool: every accelerator on it comes from an American company, NVIDIA, AMD, Intel, Amazon or Google. NVIDIA's chips are fabricated by Taiwan's TSMC, which Bloomberg describes as the go-to chipmaker for NVIDIA and as earning most of its revenue from processes of 5 nanometres and smaller.

India's position on this floor is real but specific. Electronics minister Ashwini Vaishnaw says 2-nanometre chips are already designed in India, though not fabricated here. An Indian AI accelerator is still a research programme: the government's own account says projects to design processors, general-purpose GPUs and accelerators under the National Supercomputing Mission, on the open RISC-V architecture, "have been initiated".

First floor: packaging, where India has started

A finished chip is two jobs. A fab prints the circuits onto a silicon wafer. A packaging plant, called an ATMP or OSAT, cuts the wafer into chips, seals each in a protective housing and tests it. The second job is easier, and it is the one India has begun.

Micron's plant at Sanand, Gujarat, went into commercial production on 28 February 2026. Its chief executive told an audience in September that it would process tens of millions of chips this year and hundreds of millions in 2027. But note what it receives: DRAM and NAND memory wafers from Micron's global manufacturing network. The silicon arrives; the finishing happens here.

Tata's packaging plant at Jagiroad, Assam, is not yet in commercial production. Mint reported this month that it is slated to start in December.

Second floor: the fab, where the floor is still being poured

India's first commercial fab is Tata Electronics' plant at Dholera, Gujarat, built with Taiwan's Powerchip (PSMC). The government lists it at ₹91,526 crore, with a capacity of 50,000 wafer starts a month on nodes from 110 nm down to 28 nm. A "node" is the generation of the manufacturing process; the smaller the number, the newer and denser the chip.

Two things have shifted since the plant was announced. In July, Vaishnaw said commercial operations are now expected in mid-2028, against an earlier target of end-2026. And Bloomberg reported, citing unnamed people, that the plant would start mostly on 90 nm. Tata's spokesperson said the plan "was always to start with 55nm and 90nm, followed by 28nm".

Hold that next to the ground floor. The foundry that makes NVIDIA's chips earns most of its money at 5 nm and below. Dholera's ceiling is 28 nm. That is not a failure: 28 nm chips run cars, phones and telecom gear, and the new Semicon 2.0 programme approved in July funds research on a 7-to-3 nm path. It means the fab India is building makes the chips around the computer, not the chip at its centre. This decade, Dholera does not touch the ground floor.

Third floor: the halls and the public pool

Above the factories are the data centre halls, the buildings full of power and cooling that host the chips. What one of them consumes in power and water is its own story (what is a megawatt of data centre).

Inside those halls, the government runs something unusual: a GPU pool. Rather than buy chips, the IndiaAI Mission empanels private operators and pays part of the bill for startups, researchers and agencies. In March it counted 38,231 GPUs from 14 providers, at an average of about ₹65 per GPU per hour. By July it had 15 providers and said access was available at under one US dollar an hour. The economics of the pool are in our data brief (the state as compute wholesaler).

One line in the pool's public ledger is Sarvam's: 4,096 NVIDIA H100 chips at Yotta's data centre, a bill of about ₹246.7 crore, of which the state covered about ₹98.7 crore. An Indian company, in an Indian operator's hall, on an Indian subsidy, running American-designed silicon.

Top floor: the models, which are ours

And at the top, the part you touched. Sarvam open-sourced its 30B and 105B models on 6 March 2026, saying training "was conducted entirely in India on compute provided under the IndiaAI mission", with NVIDIA contributing software libraries. A large share of the training budget went to the ten most-spoken Indian languages. BharatGen, the IIT Bombay-led consortium backed by ₹1,058 crore from the IndiaAI Mission, released Param-2, a 17-billion-parameter model in 22 scheduled languages.

These are not wrappers on someone else's model. The weights, the data choices and the engineering are Indian. On the top floor, India owns the flat.

Who holds the lease on the ground floor

So who decides whether more accelerators arrive? Mostly, the US Commerce Department, through export controls: rules on which chips can be shipped where, and with whose permission.

In January 2025 the outgoing Biden administration published the "AI Diffusion" rule, which sorted the world into three tiers. India sat in the middle tier, facing a cap of roughly 50,000 advanced AI chips through 2027. That rule never took effect. On 13 May 2025, two days before compliance was due, Commerce rescinded it and promised a replacement "in the future". If you have read that India is capped at 50,000 chips, you have read about a rule that is dead.

The replacement has not arrived. In March 2026 Reuters reported a draft under which even installations of under 1,000 chips could need a licence, and countries seeking 200,000 or more might have to invest in American AI data centres or give security guarantees. Eight days later Commerce withdrew a planned rule. In February, India signed the US-led Pax Silica declaration on secure chip and AI supply chains. Business Today, reporting on the separate US-India interim trade agreement that month, noted that India faces no export restrictions on advanced AI chips today. One analyst it quoted put the limits of that reassurance precisely:

What the terms look like when Washington does write them is visible in the Gulf. In July, Commerce let the UAE government and approved companies receive AI chips licence-free, noting the UAE had reaffirmed its commitment to make "matching investments" in US AI infrastructure. This month Nazak Nikakhtar, a former acting head of Commerce's export-control bureau, said licences to the UAE and Saudi Arabia were being held while Washington waits for their investment promises to move.

So the honest answer, as of October 2026: there is no fixed cap on India today, and NVIDIA and AMD chips are sold here legally. But access rests on policy, not on a treaty, and policy is drafted, withdrawn and redrafted in Washington on its own clock. India does not hold the lease on the ground floor. It has a good relationship with the landlord.

The government has already priced this in

None of this is hidden from New Delhi. The Economic Survey 2025-26 states it plainly: "India's access to cutting-edge compute infrastructure is limited." In a modelling exercise, the Survey's simulated data centre operators pile up in a "waiting for GPUs" state even when money and demand are plentiful. Its answer is what it calls a frugal AI pathway: small, task-specific models that run on phones and laptops, which need fewer of the chips India cannot make.

That is a sensible bargain with a constraint. It is also an admission of where the constraint sits. The dependency did not disappear when India learnt to train its own models. It moved down the building, from software to silicon.

What we don't know

We do not know what the next US chip rule will say, or when it will come. We do not know when, or whether, Dholera's 28 nm line will run at volume. And we have found no public count of how many GPUs in Indian data centres belong to Indian owners, as against foreign cloud companies.

So the next time an Indian model answers you in your own language, ask the question behind the answer: whose permission did the chip under it need, and who could take that permission back?