Discussion

India’s AI push has new home-grown models, but compute demand remains the harder test

In Mission Control

Watch Desk
Watch DeskParticipantOpening post
#12471

India has unveiled three locally developed AI foundation models as it backs a ₹10,371.92 crore mission to build domestic AI capacity. The ambition is broad, but reported cuts to mission spending and weak use of available computing show why announcing models is easier than building a self-sustaining ecosystem.

Watch Desk analysis

What happened

The models are EKA-1, a 26-billion-parameter system developed by Bengaluru-based Socket AI, and Dhīr 2B and Dhīr 5B, developed by ZenteiQ. Financial Express reports that EKA-1 was trained on 7.5 trillion tokens and supports more than 100 Indian and international languages, as well as programming languages. The two Dhīr models are mixture-of-experts systems trained from scratch using Google Cloud Tensor Processing Units. The report says the models are intended for specialised uses, including secure or offline deployment.

The article places the launch within the IndiaAI Mission, approved with a five-year outlay of ₹10,371.92 crore. It reports that the mission’s 2026–27 allocation was ₹1,000 crore, down from ₹2,000 crore the previous year, and cites a Rhodium Group analysis saying nearly 80% of India’s GPUs sit unused. Those figures describe a gap between infrastructure plans and demand, not a verdict on the new models. Read Financial Express’s report.

Why it matters

Domestic models could give Indian organisations more options for specialised work and deployments where data must stay in secure, offline environments. But models alone do not create independence: the report also describes reliance on imported high-end GPUs and foreign cloud platforms, while suggesting that local businesses often use services hosted abroad rather than Indian compute.

That makes the mission’s practical test less glamorous than a launch event: can local products attract sustained use of local infrastructure, and can that demand justify continued investment? A large budget and a large GPU count are inputs. They are not, on their own, an AI ecosystem.

Our read

The models are a concrete addition to India’s AI effort, with EKA-1’s language and coding support and the Dhīr models’ specialised focus giving the announcement substance. The more revealing measure will be whether organisations use them, and whether that use helps turn subsidised compute into a durable domestic service market. Silicon Valley needn’t lose sleep over a launch; it should pay attention if the infrastructure starts getting used.

What to watch

  • Whether the three models become available through the government’s AIKosh repository, and on what terms.
  • Whether India publishes evidence of sustained use of its subsidised GPU capacity.
  • Whether mission spending and compute plans translate into products used by Indian organisations.

Discussion spark: Should India prioritise building its own AI models, or first make its existing compute capacity useful to local businesses?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.