Bengaluru-based CoRover is best known for launching BharatGPT, positioned as India’s indigenous generative AI platform. At the India AI Impact Summit, the company announced a large-scale deployment of its multilingual platform powered by Nvidia Nemotron Speech, along with the launch of the BharatGPT DeskAI Appliance built on Nvidia Grace Blackwell architecture. The platform enables secure, sovereign AI at population scale, supporting 0.8 million+ concurrent users with natural voice and text interactions. In an interview, Ankush Sabharwal, founder & CEO, CoRover.ai, discussed with Sudhir Chowdhary the importance of sovereign AI and the role of BharatGPT. Excerpts:
Is BharatGPT really India’s own AI – or just ChatGPT with an Indian wrapper?
BharatGPT is not a wrapper around ChatGPT. We are not aspiring to compete in the B2C chatbot race. We are a B2B, enterprise-focused AI company. Our goal is to provide ChatGPT-like capabilities to enterprises, but trained on legitimate, proprietary data that we either own or are authorised to use. We do not scrape publicly available content and repackage it. Even if something is legally accessible, that does not make it ethically usable. Sovereign AI must be built on trust.
Today, BharatGPT powers large-scale multilingual deployments across sectors like financial services, travel, utilities, citizen services and other transaction-heavy environments. Our AI platform, built in collaboration with Nvidia using Nemotron Speech models and accelerated by Grace Blackwell architecture, supports over 800,000 concurrent users at population scale, delivering near-zero latency voice, video and text interactions.
We operate a three-layer architecture: Classic AI/NLP systems, secure graph-based RAG models including smaller purpose-built language models, and optional integration with external LLMs like ChatGPT or Gemini if enterprises choose. Our innovation lies in orchestration, sovereignty, and secure enterprise deployment, not hype. With the launch of BharatGPT DeskAI Appliance and BharatGPT Mini, we are also enabling a fully offline, fully private AI assistant that runs directly on Nvidia AI infrastructure on a desktop, no internet required, full data control retained by the organisation.
How is BharatGPT being used in healthcare? Do Indian hospitals trust AI enough to let it interact directly with patients?
Healthcare is still an evolving segment for us. We are not deeply embedded in hospital systems yet, but we work with organisations like Population Foundation India and global NGOs. Our broader approach in healthcare and other sectors is to deploy secure, domain-specific AI agents that operate within enterprise governance frameworks. With graph-based secure RAG and task-driven AI agents, the goal is not just answering questions but enabling end-to-end workflows while maintaining enterprise-grade security and data control. Trust is earned through deployment discipline, not marketing claims. That is why sovereign AI and controlled infrastructure matter deeply in sensitive domains like healthcare.
Why should India build its own AI when US tech giants already dominate the field?
We should not build AI out of FOMO or to prove a point. The question is purpose. If our objective is to solve a problem efficiently, we must use whatever works – cheaper, faster, better. Sovereignty is not isolationism; it is resilience. True sovereign AI means that if global systems are disrupted, your critical systems should still function.
That requires control over deployment architecture, data governance and operational continuity. It does not mean reinventing everything from scratch – but it does mean ensuring your AI systems reflect your languages, priorities and enterprise needs. Our partnership with Nvidia demonstrates that sovereignty can coexist with global collaboration. Built with Nvidia Speech models and accelerated by Blackwell architecture, BharatGPT shows that multilingual AI can be deployed securely and efficiently at scale in India. Trust follows sovereignty.
Can AI really work in India’s many languages and accents?
Yes, but only within defined use cases and domains.We have over 1.3 billion user interactions across our AI applications since 2016. That scale has given us deep experience in multilingual and dialect-rich environments. However, we do not believe in one universal model that works perfectly across all languages, domains and geographies. Healthcare in Delhi differs from healthcare in rural India. Banking regulations in India differ from the Middle East or the US. So language capability must be domain-specific. With Nvidia Nemotron Speech models and libraries, we have significantly improved speech speed, stability and latency – which is critical for voice-led service journeys at population scale. Multilingual AI works when it is grounded in real enterprise workflows, not just trained on generic data.
Can India realistically compete in LLMs, or should we focus on AI platforms and apps instead?
At CoRover, we focus on AI solutions and full-stack enterprise AI platforms rather than building massive general-purpose LLMs. Training a large model using open-source algorithms and GPUs is not the differentiator. The real value lies in orchestration, governance, deployment, security and domain integration. That is why we invest in graph-based secure RAG, domain-specific small language models and task-driven AI agents capable of completing end-to-end workflows. That said, India should experiment. Innovation may come from new model architectures, distributed AI approaches or more efficient compute utilisation. Reinventing how models are built is more meaningful than replicating what already exists.
Will AI chatbots replace call centres in India? IT services firms are already feeling the heat from AI disruption…
Yes – and in some cases, they already have.We have deployments where not a single human agent is required in the call centre for certain workflows. But this is not purely a displacement story. Humans are meant to be creative and solve complex problems, not repeat scripted playbooks. AI can automate repetitive interactions while enabling knowledge workers to move up the value chain. Over time, IT services firms must transition from effort-based billing models to outcome-based models. This AI wave is an opportunity for reinvention, not decline.
What policy support does Indian AI startups need most right now – compute, data or govt procurement?
The government has already provided significant support in terms of computer and data initiatives. Data alone is not a differentiator. If everyone has the same data, differentiation disappears. The real need now is adoption. There must be a demand pull from enterprises and public institutions articulating real problem statements. Startups should partner with domain experts, solve specific problems, and build credible solutions. Government procurement should accelerate utilisation and deployment, not just fund experimentation. We are now at the stage where AI solutions must be implemented at scale.