Beyond the Cloud: Sovereign AI Insights from India AI Impact Summit 2026

The world of Artificial Intelligence is moving fast. While the "ChatGPT moment" introduced the power of large language models, a more profound shift is underway in the boardrooms of India's largest organisations: Sovereign AI, keeping data, models, and compute under direct, local control rather than relying on external clouds.

At the India AI Impact Summit 2026 in New Delhi, this shift takes center stage in a panel titled “Beyond the Cloud: The Sovereign AI Moment.” The discussion focuses not on model demos, but on what it truly takes to run AI inside public institutions and national infrastructure.

On stage at the India AI Impact Summit, the conversation was guided by some of the leading AI experts: Bernard Nguyen, Director of Engineering at NVIDIA; Ritwik Rath, Executive Director at HPCL; Balasubramanian V, Chief General Manager at NABARD; and Abhinav Aggarwal and Raghav Aggarwal, Co-Founders of Fluid AI

What does Sovereign AI look like in practice, and why is the public sector treating it as core infrastructure? Let’s explore.

What is Sovereign AI?

Sovereign AI means running AI on in-country, on-prem infrastructure, so the organization controls the models, platform, compute, and data.

For Indian organisations, it matters for independence from hyperscalers, predictable costs vs token-based pricing, and compliance with the DPDP Act. It also enables Indic-language models that better serve India’s diverse users.

At the India AI Impact Summit, one key question stood out: Why choose Sovereign AI over the traditional cloud model?

HPCL’s Ritwik Rath explained "being a PSU oil company... we are a national critical information infrastructure setup so it also makes it imperative for us to ensure that the data stays on prem to the extent possible and if not it stays within the geography of the country".

This transition is driven by three primary factors: Security, Compliance, and Cost.

1. On-Premise AI for Public Sector Security

For organizations like HPCL, AI infrastructure decisions go far beyond innovation strategy. They sit at the intersection of national interest and operational continuity. Energy supply chains, refinery operations, and distribution networks form the backbone of India’s economic stability.

In such an environment, sending sensitive operational or logistics data to external hyperscalers is not just a technological choice, it carries geopolitical and strategic implications.

HPCL’s approach, described as “cautious adventurism,” reflects this balance. They are actively experimenting with advanced agentic AI systems to improve efficiency and decision-making. But they are doing so on a secure, on-premise foundation. Innovation is encouraged, but never at the expense of sovereignty or control.

2. DPDP Act 2023 and AI Compliance Requirements

The legal landscape in India is changing. Balasubramanian V, CGM at NABARD, points out that the Digital Personal Data Protection (DPDP) Act of 2023 and the subsequent rules have fundamentally changed how public undertakings must handle data.

Balasubramanian V, put it succinctly:

“The evolving regulatory and statutory requirements place a lot of responsibility on public undertakings. The choice of going for an on-prem implementation becomes a compliance requirement rather than a technological preference.”

For a regulator like NABARD, moving AI on-premise is now a compliance requirement. It allows them to:

• Ensure data localization.
• Maintain a controlled infrastructure.
• Achieve strategic independence, ensuring they are not beholden to a single external provider forever.

3. On-Premise AI vs Cloud: A Cost Perspective

One of the most significant barriers to enterprise AI scaling is the cloud's token-based pricing model. Organizations often face unpredictable costs because every interaction with a chatbot or AI model consumes tokens, and there’s little control over usage spikes.

In contrast, on-premise Sovereign AI offers budgetary predictability. Organizations can calculate the cost of hardware, compute, and orchestration over a five-to-six-year horizon. Once the infrastructure is built, the cost remains the same regardless of the extent of usage.

Open Source and Specialized Models

The shift toward on-premise AI wasn’t possible at scale a few years ago. Today, it is.

Bernard Nguyen, Director of Engineering at NVIDIA, pointed to the explosion of open-source AI models.

“Earlier, the most powerful models were closed and accessible only through APIs”. Now, high-performing open models are released frequently, with weights available on platforms. This changes everything.

Enterprises can now run and fine-tune open models on their own infrastructure. NVIDIA supports this with GPUs and proven training recipes. Sovereign AI is now practical.

General vs. Specialized Agents

Large foundation models excel at deep reasoning and general intelligence. They are powerful, versatile, and capable of handling complex, multi-domain tasks. But that power comes with trade-offs, they are computationally heavy, slower in certain workflows, and more expensive to operate at scale.

In contrast, smaller, specialized models can be trained to be best-in-class for highly specific tasks.

Specialized models can do one job well without a big general model, e.g. tenders, medical reports, or finance docs that are often faster and cheaper.

A sovereign AI strategy enables both layers:

• Large models for strategic reasoning and orchestration
• Smaller, specialized agents embedded within workflows

As enterprises move from experimentation to scaled deployment, this modular, specialized agent architecture may prove to be the most efficient and sustainable path forward.

From Isolated Use Cases to a Platform Approach

Many AI initiatives fail because they start as disconnected pilots, without a unifying structure.

Raghav Aggarwal emphasized that scale demands a different mindset.

“If AI has to work across 20,000 employees, it cannot be a series of point use cases. It has to be platformized.”

NABARD adopted this platform approach by building a single orchestration layer to host multiple models and agents.

• Reusability: The same data pipelines and infrastructure can support multiple departments. What works for one workflow can be extended to another.
• Centralized Governance: A unified risk, compliance, and privacy framework applies across the organization, reducing duplication and regulatory gaps.

Rather than reinventing AI for every function, the institution builds once and scales intelligently.

Real-World Impact: Case Studies in Efficiency

1. HPCL’s Medical Report Parsing (The "4 Hours to 18 Minutes" Story)

One of the most striking examples of efficiency comes from HPCL's automation of periodic medical examinations.

• The Problem: HPCL, with around 5,000 employees undergoing annual medical exams, faced a huge productivity drain. Manually entering and parsing reports took an average of 4 hours per employee, totaling roughly 20,000 person-days annually.
• The AI Solution: A specialized GenAI agent was deployed on HPCL’s platform to parse reports automatically.
• The Impact: Processing time dropped from 4 hours to just 18 minutes, with humans verifying content per DPDP guidelines. Because of this impact, HPCL has gone under a complete transformation and have launched 24+ new GenAI use cases across departments.

2. NABARD’s Unstructured Data Insights

For a regulatory body like NABARD, the challenge was not just speed, but visibility into their own archives.

• The Problem: 60–70% of NABARD’s data was unstructured, making traditional searches and analytics ineffective.
• The AI Solution: A GenAI PoC was used to analyze unstructured files efficiently.
• The Impact: The PoC demonstrated actionable insights from unstructured data. Because of this success, NABARD has planned the launch of 20+ new GenAI use casesacross departments

3. Real-Time Voice AI for Logistics Safety

Another impactful example comes from using Voice AI to bridge the digital divide in logistics

• The Problem:
  In logistics operations, IoT sensors can detect safety events like harsh braking. However, language barriers and literacy constraints often limit effective real-time communication with drivers in the field.
• The AI Solution: A Voice AI system was deployed to trigger real-time calls to drivers in their local language when such events occur, enabling immediate, context-aware safety communication.

These examples highlight the real-world impact of AI in improving efficiency and decision-making. By adopting Fluid AI, HPCL and NABARD are now able to scale their AI initiatives, securely and efficiently, and launch multiple new GenAI use cases. Fluid AI is helping organizations turn past successes into a foundation for a broader, future-ready AI strategy.

Looking Ahead

Sovereign AI isn’t about ditching the cloud, it’s about control. With on-prem deployment, organisations keep data secure, meet DPDP requirements, and plan costs predictably, without relying on global hyperscalers.

As the founders of Fluid AI concluded, this shift requires both a mindset reset and immediate execution. Abhinav Aggarwal emphasized:

“We're so used to seeing a problem we solved before and take the most complicated route or approach to solving it... [we must] leverage agentic AI in super interesting ways and really reframe the way we think about these problems.”

Raghav Aggarwal reinforced the urgency:

“I think the key thing is agentic and sovereign is like water it's a utility... that's where organizations need to go to the gym today... you need to build that muscle, you need to build a partner network, you need to build a governance... and you need to get it all together lined up today rather than wait for it to happen a year or two down the line.”

Sovereign AI is not optional infrastructure, it is foundational. And the time to build is now.