"A stack is sovereign when it is a glass box, not just when it sits inside your borders." That is how Prof. Ganesh Ramakrishnan of BharatGen frames SVAIAM, a new reference architecture and implementation of a sovereign AI stack for India, built by BharatGen, IBM and NxtGen.
The work grew out of discussions in the AI Alliance India community, starting at our India technical meet in November 2025. It gained grounding in February 2026, when letters of intent were signed with the Government of Andhra Pradesh at the IndiaAI Summit, and the partners then worked with the state government on the use case and its trusted data sources. We are glad to share what came of it.
Download the SVAIAM reference architecture (PDF) →
The gap: sovereignty one layer at a time
Most sovereign AI efforts solve one piece of the problem: a sovereign model, sovereign infrastructure, or a data residency policy. Each matters. None of them, on its own, gives a government or enterprise a stack it can run, govern and scale. SVAIAM covers every layer, from the hardware to the application a citizen actually uses.
Five tests of a sovereign stack
The paper sets out five characteristics a stack needs to be sovereign:
- Data residency and control. Data, metadata and derived knowledge stay under the jurisdiction's control.
- Operational sovereignty. The platform can be deployed, run, monitored and secured with no required outside operator.
- AI choice. Models from any provider can be selected, combined, replaced and fine-tuned, with no lock-in.
- An open ecosystem. Open standards and modular components that local builders can extend.
- Security compatibility. The stack runs in nationally controlled environments, air-gapped if needed.
The authors' point is that these should not come at the expense of quality, cost, developer speed, scale or governance.
The stack: who builds each layer
- Sovereign infrastructure: NxtGen's multi-accelerator sovereign IaaS, spanning AMD, NVIDIA and Intel hardware
- Data and AI operations: Red Hat OpenShift, llm-d and OpenRAG
- Data: government and academic data sources
- Models: BharatGen's sovereign models (Shrutam for speech to text, Sooktam for text to speech, Param2 for text) alongside open models such as Google's Gemma 4
- Common governance: IBM watsonx.governance
- AI developer tools: open-source tools including Docling, Langflow, OpenRAG, CUGA and SDG_Hub
- Control plane: IBM Sovereign Core, running across the stack
Each component plugs in through open interfaces, so any model or tool can be swapped as better options arrive.
The proof: a voice agent for farmers
To show the architecture working end to end, the team built a reference application with the Government of Andhra Pradesh: an agricultural advisory agent for farmers. A farmer asks a question by voice on WhatsApp in Telugu, Hindi or English. Shrutam transcribes it. An agent plans the response and calls governed government data sources, including e-NAM market prices, Soil Health Card data, AgriWatch forecasts and Vyavasaya Panchagam farming guidance. The answer is grounded in those sources, every call is traced, and Sooktam speaks the reply back in the farmer's own language. All of it runs on NxtGen's sovereign cloud.
Sovereign does not have to mean slower or costlier
The paper reports results from a benchmark run jointly by IBM, NxtGen and Red Hat on a 20-GPU cluster spanning three hardware vendors:
- 91% higher throughput with llm-d routing, compared with a standard Kubernetes baseline
- 5.4× faster time to first token in the same test
- Answer accuracy improved from 0.47 to 0.88 as the team iterated against a fixed use-case specific benchmark in watsonx.governance
As the authors note, gains vary by model and workload.
In their words
"A stack is sovereign when it is a glass box, not just when it sits inside your borders. In SVAIAM, Shrutam (STT) hears the farmer, Sooktam (TTS) speaks back in their own language, Param (text LLM) is a strong sovereign option for reasoning, and Patram (VLM) extends that reach to the documents and images that matter to them. We are grateful to IBM and NxtGen for building this with us."
Prof. Ganesh Ramakrishnan, Principal Investigator and Board Member, BharatGen
"NxtGen powers SVAIAM’s sovereign infrastructure with a multi-tenant, heterogeneous AI platform built on Red Hat OpenShift and OpenStack — enabling scalable, cost-optimized deployments with full operational control, while integrating open-source innovation across inference, orchestration, and governance."
Abhisyant Anasapurapu, VP and Head of Products, NxtGen
"IBM is committed to bring sovereign AI from a concept to an implementable full-stack architecture, leveraging India's sovereign AI models and AI infrastructure with a portable software stack built on open-source and open standards, while showing how to achieve high quality, performance and developer agility, while optimizing cost. SVAIAM is a realization of this design."
Amith Singhee, Director of IBM Research India and Chief Technology Officer of IBM India/South Asia
Build on it
SVAIAM is a reference, not a finished product. It is meant to be adapted to other sectors, languages and jurisdictions, and the authors invite the community to extend it. If you are working on sovereign AI, in India or anywhere else, we would like to hear from you.
Read the full reference architecture (PDF) →
Get updates like this from the AI Alliance community: subscribe to our newsletter.