India is buying compute faster than it is building power
From 38,000 GPUs toward 100,000, plus a $15 billion gigawatt-scale campus. The bottleneck is shifting from chips to electricity and land.
India’s AI compute position has changed faster than almost any other part of its technology policy. The IndiaAI Mission, cleared in March 2024 with a five-year corpus of ₹10,371.92 crore, originally targeted 10,000 GPUs. It has deployed roughly 38,000, offered to startups and researchers at subsidised hourly rates.
On 17 February 2026, Union Minister Ashwini Vaishnaw announced a further 20,000 units, with the government targeting 100,000 GPUs by the end of the year.
Running alongside the public programme is a private one an order of magnitude larger: Google’s roughly $15 billion AI hub at Visakhapatnam, which broke ground on 28 April 2026 and includes India’s first gigawatt-scale data centre campus and a new international subsea gateway.
Why subsidised GPUs matter
The public compute portal — which absorbs close to 45 per cent of the mission’s spending — addresses a specific market failure. Frontier-scale training is available to firms that can pay hyperscaler rates. Indian research groups and early-stage startups largely cannot, and the alternative is not cheaper compute but no compute.
Subsidising access at a few tens of rupees per GPU-hour is a straightforward way to let people build things. Whether they build things worth the subsidy is a question the programme cannot answer in advance, and shouldn’t be expected to.
The constraint nobody solved by buying chips
A gigawatt of data centre capacity requires a gigawatt of generation, transmission to deliver it, and water or an alternative to cool it. India’s compute ambition is now large enough that these are the binding constraints rather than GPU supply.
This is why the Visakhapatnam campus is inseparable from Andhra Pradesh’s energy programme — the state’s renewable targets, and the NTPC green hydrogen hub at nearby Pudimadaka with its 7 GW substation. The clean-energy commitment attached to the Google project is credible only because that build-out is happening in parallel.
Where it does not happen in parallel, data centres will be commissioned into grids that cannot supply them, and the shortfall will be met by whatever generation is actually available.
The subsea gateway
The least-discussed element may be the most structurally significant. India’s international connectivity has landed principally at Mumbai and Chennai. A third major gateway on the east coast changes the country’s network topology and reduces a concentration risk that has existed for two decades.
What this does not settle
Compute is an input. India now has considerably more of it, at a subsidised price, with much more coming. What it does not yet have is a domestic frontier model developer at the scale the compute would support, or a clear answer on where the training data for Indian-language models comes from.
Buying capacity is the part of an AI strategy that can be executed by procurement. The rest cannot.