India’s AI Challenge Is Bigger Than Just GPUs
| AUTHOR | Gopal Tomar |
| DATE | September 21, 2026 |
| CATEGORIES | Artificial Intelligence AI Policy Semiconductors Digital Infrastructure Technology Policy |
Recently, there has been a huge discussion all around the world on whether India needs AI, then AI needs GPUs, and therefore India needs to buy/build more GPUs. And the reason seems very much straightforward: countries without them will be dependent on others.
But the important question to ideate upon is: if India had every GPU it needed tomorrow, then would it have achieved AI sovereignty, which seems to be the main goal of all this?
It is obvious that modern AI needs a lot of computational power, but there are other factors too which are part of the system that will lead to the goal being achieved. So, India’s real challenge is not just to have more chips. It is to build an AI stack that can turn computing power into useful, affordable, and sovereign capability. A bit of a deep dive into the AI system shows that it has several layers, like electricity, data centres, networking, computing hardware, data, models, model adaptation, and then the application and, at the end, delivery to users.
And this same understanding is broadly reflected by the India AI Mission, and its architecture is beyond compute so as to include our own foundation models, datasets, applications, talent, startup financing, and reliable, trusted AI with safety. Following up on this, as of 2026, official figures put the India AI Compute portal at more than 38,000 GPUs alongside 1,050 TPUs, giving cloud-based compute access to researchers, startups, MSMEs, academia, and other eligible users. So, it’s a sequential pipeline in which a GPU needs a data centre to operate, electricity to power it, networks to connect it, models and data to run on it, and then users and, most importantly, applications capable of delivering value to citizens and businesses.
Now the main question is: which part of this AI stack should India own, which should it control, which should it secure, and where can it rely on others?
The least visible layer of the AI stack is electricity, which might become one of the biggest constraints. Data centres need a lot of electricity, and obviously, AI is increasing that demand. The International Energy Agency estimates that global data-centre electricity consumption could roughly double from 405 TWh in 2025 to around 950 TWh by 2030, mainly because data centres are growing even faster.
So, India’s AI strategy cannot separate semiconductor policy from energy and infrastructure policy. A thousand additional GPUs are useful only if the base is there to run them, which is physical infrastructure. For reducing strategic dependency, a semiconductor push is required. We are building capabilities in chip design, manufacturing, packaging, and testing. However, advanced AI chips still depend on complex global supply chains and technologies that require high specialization.
India’s main focus should be on developing indigenous models, mainly because language, culture, and domain-specific needs create strategic value. And that is what is being supported by the IndiaAI Mission by using Indian datasets and custom sectors such as healthcare, education, agriculture, climate, etc. But sovereignty does not require India to reproduce every frontier model developed in the US and China. So, instead of training a foundational model from scratch, open and accessible models can be fine-tuned for use cases that matter to India’s users.
India has a high stake in the high demand for AI for several use cases, and this gap is an opportunity for a lot of emerging startups as well. So, India’s strength may not be in building the world’s largest AI models but in using them at scale. And the aligned advantage which India has is the technology workforce, a growing AI ecosystem, and platforms like UPI and Aadhaar, which show that India can build and deliver digital services to millions of people.
For every layer which I discussed earlier, India should therefore ask four questions: Can we control it? Can we access it reliably? Can we substitute it if geopolitical conditions change? Can we deploy it at national scale? And the answers will differ for each layer.
In conclusion, India’s AI challenge is therefore not a choice between building chips and leaving other layers untouched. It should be: which parts of the AI stack must India control, which can it access through trusted partners, and which are the niche areas that the rest of the world cannot copy.