Sovereign AI vs Open AI: The Real Choice for India

“Sovereign AI” is the most misused term in 2026. It is used to mean everything from “India should build its own foundation model” to “India should ban OpenAI and Anthropic.” Both are wrong. The real choice is more nuanced: which layers of the stack are sovereign and which are open? This essay proposes a framework.

The four layers of the AI stack

Every AI system has four layers:

1. Compute. GPUs, TPUs, custom accelerators, networking, data centre capacity. The physical infrastructure.
2. Models. Foundation models (GPT-5, Claude 4, Gemini 2), fine-tuned models, custom models. The intellectual property.
3. Runtime. The agent runtime, the inference layer, the safety layer, the orchestration layer. The software that turns a model into a system.
4. Data. The training data, the fine-tuning data, the user data, the context data. The information.

Each layer has a different sovereignty profile. The mistake is to treat “sovereign” as a binary — either the whole stack is sovereign or it isn’t. The real answer: each layer has a different requirement.

The framework

For each layer, ask three questions:

1. Is the layer commoditised? If yes, the sovereignty question is less important. If no, sovereignty is critical.
2. Is the layer a strategic dependency? If yes, the country should have at least one sovereign option. If no, the country can rely on the global market.
3. What is the cost of being wrong? If a foreign dependency turns hostile, what is the cost? If the cost is existential, sovereignty is required.

Layer 1: Compute

• Commoditised? No. Compute is concentrated in TSMC (fabrication), ASML (lithography), NVIDIA (GPUs). The supply chain is fragile.
• Strategic dependency? Yes. Compute is a hard dependency for every AI capability. Without compute, no models, no runtime, no data.
• Cost of being wrong? Existential. If NVIDIA stops shipping to India, India’s AI capability collapses overnight.

Sovereignty recommendation: high. India should pursue sovereign compute through the India Semiconductor Mission, the IndiaAI compute cluster, and partnerships with TSMC. The goal: at least 50% of India’s AI compute should be sovereign by 2030.

Layer 2: Models

• Commoditised? Partially. The frontier model market is an oligopoly (OpenAI, Anthropic, Google, Meta, Mistral). But open-weights models (Llama, Mistral, DeepSeek, Qwen) are narrowing the gap.
• Strategic dependency? Yes, but less than compute. Multiple model providers create redundancy.
• Cost of being wrong? High but not existential. If OpenAI stops serving India, Anthropic and Meta can fill the gap.

Sovereignty recommendation: medium. India should pursue sovereign foundation models for Indic languages and for strategic use cases (defence, public sector). The Indic language models (BharatGPT, Airavat, etc.) are a start. The goal: at least one sovereign model that is competitive on Indic-language benchmarks by 2027.

Layer 3: Runtime

• Commoditised? No. The runtime layer is where India has a real chance to lead. The agent runtime, the safety layer, the observability layer, the orchestration layer — these are the layers where India can build differentiated, sovereign capability.
• Strategic dependency? Yes. The runtime is what turns a model into a production system. Without a sovereign runtime, India is dependent on US and Chinese tooling.
• Cost of being wrong? Medium. The runtime is replaceable; the question is how much it costs to replace.

Sovereignty recommendation: very high. India should be a global leader in the agent runtime, safety, observability, and orchestration layer. The Substrate Pattern, Defence in Depth, and Tiered Governance Model are the framework. The investment should be in Indian companies (Neul Labs is one) that build these layers.

Layer 4: Data

• Commoditised? No. Data is the long-term strategic asset. The country that has the most data — and the best data governance — wins the long game.
• Strategic dependency? Yes. Data is the input to every AI capability. Without data, the models are useless.
• Cost of being wrong? Existential in the long term. If India’s data is owned by foreign companies, the country has lost the most important strategic asset of the 21st century.

Sovereignty recommendation: very high. India’s data should be sovereign. The Digital Personal Data Protection Act (DPDPA, 2023) is a start. The IndiaAI dataset initiative, the National Data Sharing Platform, and the India Stack are the policy infrastructure. The goal: at least 80% of India’s strategic data should be under Indian jurisdiction and Indian governance by 2027.

The K-shaped outcome

By 2030, I expect a K-shaped AI world:

• Top stack: US hyperscalers + US-allied jurisdictions (UK, Japan, Korea, Australia, EU) on US-origin compute, US-origin models, US-origin standards.
• Mid stack: EU + India + Brazil + Indonesia, running mixed-origin stacks with strong local audit and enforce layers.
• Long tail: A long tail of jurisdictions that depend on whichever of the above will sell to them.

India is in the mid-stack today. The bet is to push India to the top of the mid-stack by 2030. The four bets above are the path.

The pragmatic answer

“Sovereign AI vs Open AI” is the wrong question. The right question: which layers are sovereign and which are open? The framework above is the answer:

• Compute: high sovereignty. India must own its own compute.
• Models: medium sovereignty. One sovereign model for Indic languages and strategic use cases.
• Runtime: very high sovereignty. India can and should lead the runtime layer globally.
• Data: very high sovereignty. Indian data under Indian governance, always.

The choice is not binary. The choice is layered. Get the layers right, and India wins the AI century.

— Dipankar Sarkar Founder, Neul Labs · IIT Delhi alumnus · Takshashila Institution alumnus

This essay is a follow-up to “AI Sovereignty Is the New Stack Question” (2026-06). The frameworks I cite — Substrate Pattern, Defence in Depth, Tiered Governance Model — are documented at dipankar.name/frameworks/.