As Indian enterprises move from AI pilots to production, the question is no longer just “which model?” but “which stack?” With geopolitical shifts, data sovereignty concerns, and the rise of domestic AI initiatives like BharatGen, CIOs must decide how to balance global frontier models with India-built alternatives without compromising on performance, cost, or control. In this interview with CIONow, Aditya Maheshwari, a BharatGen consortium member, explains where sovereign AI fits in the enterprise stack, how CIOs can adopt it pragmatically, and what it will take to scale safely while building long-term capability.
Can sovereign AI turn into a business imperative, and not just a policy conversation?
Post current geo political events, most of the Big Tech are dancing to their own government’s tune, which means even AI is becoming like Digital Public Infrastructure – a public good.
Its applications are spreading widely and affecting everyone, not just businesses but ordinary citizens also, through technologies like OCR, image recognition, and generative AI. So the sovereign stack, from compute to application layer, including models, APIs, and data, needs to be sovereign if a country wants to protect its space.
India, for example, wants to expand its leadership in digital public infrastructure globally through UPI and other services. Over the last 10–15 years we have extended digital governance to citizens, which has also helped create a digital startup ecosystem. We have seen many billion-dollar startups and unicorns. This is not merely a policy decision; it’s the future of the economy. Investment and valuations in this sector are already huge and will only grow.
What will make CIOs trust India-built AI models enough to integrate them into enterprise workflows?
There are two aspects: intent and capability. Intention covers trust and commitment; capability covers performance and business viability. Compared with frontier models, BharatGen is still evolving, so trust must be built from both sides. Many enterprises, across tiers, are already using BharatGen models, so adoption is happening from both ends.
We are also building domestic capability. Sovereign ventures, like BharatGen, pride themselves on transparency – their data, models and teams are open source and auditable, something that is tricky for businesses, but important as they plan for the future. Concerns linger around future price increases, discontinuance of services or products of foreign players, issues amplified by geopolitical movements and valuations changing at lightning speed.
We need to build an Indian AI infrastructure ecosystem, not just BharatGen but foundational and applied AI companies together, similar to how the software ecosystem matured. Indian IT has earned global trust through coordinated investment, skills, and local institutions. To make the Indian AI story global, stakeholders must invest in trust. Rather than asking “why trust,” the conversation should be about forming partnerships and jointly developing the ecosystem.
Where does BharatGen fit versus global giants, which excel at coding, financial math, and complex reasoning?
Different model sizes serve different tasks. Smaller models (e.g., 17B parameters) won’t directly compete with 500B+ models on some tasks, but they can be ideal for many enterprise use cases. The opportunity is to identify tasks where smaller, local models suffice such as internal automation, document querying, education tools, defence-specific workflows, and other non-public tasks. Success depends on mapping specific tasks to the right model size and deployment approach.
When organizations scale out from testing AI, what’s the biggest challenge they’re going to face deploying India-centric models such as BharatGen?
There are several key challenges:
- Compute: You will usually need on-prem resources or access to rented cloud instances with these open source models. Compute availability is the top constraint.
- Evaluation and benchmarking: Creating reasonable, domain specific benchmarks are an unsolved problem. Public benchmarks are helpful but it is easy to game these or they will accidentally spill over into your training data. The reality in your system might differ from performance on benchmark tasks.
- Real world fidelity: Sometimes state-of-the-art models even fail at simple tasks relevant to their domain such as pulling data out of OCRs even if their benchmark score is excellent.
Many Indian firms currently use global foundation models. How should enterprises balance global models with India-specific models like BharatGen? Should this balance be determined by cost, control, or relevance?
All three. In business, there are immediate, mid-range and long-term priorities. If the customer asks for a certain model: be it from Anthropic or OpenAI or Grok enterprises will make sure it is delivered and that decision will be purely on the basis of cost and availability. But if companies focus only on short-term gain and don’t invest in domestic AI capabilities, they risk long-term dependency on foreign, closed-source models. That risk may not be evident now but grows over the medium to long term.
BharatGen and other domestic players aim to build and support the wider ecosystem, not just sell models. There’s scope for many partnerships, and the Ministry of Electronics & IT and the Department of Science & technology are focused on ecosystem development rather than backing a single startup.
How should Indian CIOs approach Bharat Gen adoption alongside existing foreign models, and plan for a gradual shift in dependency?
The practical approach is hybrid and incremental. Start small: form a compact in-house team, partner with Indian enterprises that are building modules (not just BharatGen), and gradually develop expertise. With that internal capability you can make informed decisions about token budgeting, on-prem versus cloud, required GPU capacity, and future outlook. Until you build a research and development core, you won’t fully understand operational trade-offs. CIOs who grasp token economics and infrastructure nuances will avoid costly information asymmetries in the AI age.
Many enterprises lack data pipelines to scale local LLMs. Is that a major blocker?
Absolutely. Preparing AI training data from text, images, and websites, transforming it from rectangular/tabular forms to instruction/QA/summarization formats, is a major challenge. We are working on data-pipeline solutions within BharatGen to help enterprises generate usable training data, but this remains a widespread gap.
Which enterprises or use cases are ready to adopt BharatGen today, and which should wait?
Ready now:
- Internal automation tasks (document search, querying scanned PDFs, knowledge-base assistants).
- Sectors or functions with lower-stakes, high-volume tasks—certain banking document workflows, education content tools, some defence and public-sector applications.
- Teams willing to pilot parts of their pipeline with a local ecosystem partner.
Should be cautious / may wait:
- High-stakes, public-facing products that demand cutting-edge broad-reasoning or superior open-domain inference compared to frontier models.
- Workloads that currently require 500B+ parameter level performance for generalized reasoning or highly complex logic.
Enterprises should evaluate task-level requirements rather than sector-level labels. If the task aligns with the capabilities of smaller/open models, adopt now; otherwise phase in adoption while monitoring model progress.
Our models are openly available on AIKosh and Hugging Face, with 135,000 downloads so far. On enterprise engagements, we are focused on co-building with industry leaders across 4 key segments: Governance, BFSI, Healthcare and Education.
