India Hasn't Lost the AI Race: Vishal Sikka Says Build Foundation Models - And Build Them Here
Vishal Sikka, former Infosys CEO and an early advisor to OpenAI, is clear: India should build foundational AI models, not just apps. He argues India has the scale, talent, and data to do both - and that the game is still wide open.
His stance cuts against the "just build on top" thesis. For executives, that means the decision isn't binary. The smarter play is layered: invest in models where India has an edge, and spin up high-value applications on top of them.
Why foundation models still matter
Sikka's core point: today's AI is powerful yet inefficient and limited. He highlights a basic comparison - the human brain runs on roughly 20 watts. In contrast, large models burn massive compute and energy. By his estimate, human intelligence is still 12-18 orders of magnitude more efficient. In other words, there's a long way to go.
He also points to gaps in current systems: weak grounding in the physical world, shallow links between concepts and real activities, and limited common sense. That's an opening for new research, new architectures, and new models - not a closed chapter.
"We haven't lost" - and we shouldn't outsource the core
Sikka pushes back on the idea that India should skip model-building and just focus on apps. He notes even academic courses now teach how teams can build foundation models. His take: it's not preordained who gets to do this. If India steps back, it gives away the base layer of future products and policy leverage.
Where India has an edge
India has unique assets other countries don't: a trove of multilingual data, historical manuscripts, and a digital public infrastructure that is the envy of many markets. With the India Stack, the country can build models and applications tuned to real usage at population scale.
The debate: commoditized models vs strategic depth
Some leaders argue models will commoditize and the value will concentrate in applications. Sikka's counter: even if parts commoditize, owning core capability changes your cost curve, de-risks dependence, and creates leverage in standards, safety, and cross-border policy. For a nation of 1.4B, optionality is strategy.
What executives can do now
- Fund a portfolio of models: 1-2 general models (open and commercial) plus domain models for finance, health, agriculture, and public services.
- Build a compute consortium: secure long-term GPU/accelerator supply, power, and cooling. Incentivize domestic inference capacity to lower total cost of ownership.
- Create high-quality, lawful data pipelines: multilingual corpora, annotated public data, and industry data trusts with privacy and consent baked in.
- Invest in talent and research: joint labs with IITs/IISc, fellowships, and return-to-India programs for AI researchers and systems engineers.
- Stand up India-specific benchmarks: Indic languages, code-switching, factual accuracy, and safety aligned to local contexts.
- Launch anchor deployments: government and large enterprises commit to production pilots with clear procurement standards and APIs.
- Partner with open-source communities: reduce vendor lock-in, speed up iteration, and ensure auditability.
- Set governance early: model cards, data lineage, evaluation protocols, and IP structures that keep strategic assets in-country.
Metrics that matter
- Training cost per billion tokens and inference cost per 1,000 tokens.
- Energy per million tokens and latency at 4k/32k context.
- Accuracy and hallucination rates on Indic-language and India-specific tasks.
- Share of inference served on domestic compute.
- Number of production apps with positive unit economics.
90-day starting plan
- Form a model strategy council across industry, academia, and government.
- Run a data audit: identify top 10 proprietary corpora you can legally use and clean.
- Secure compute: lock multi-year accelerator capacity and power for training and inference.
- Kick off a bilingual pilot (e.g., customer support or field ops) using retrieval-augmented generation on Indian data.
- Spin up internal training for product, legal, and security teams to build with confidence.
The bottom line
Sikka's message is blunt: India is big enough to do both - build foundation models and build the widest set of applications. The cost is high, but the cost of dependence is higher. Treat models as national capability, not a nice-to-have.
If you're building internal capability or upskilling teams for this shift, explore practical learning paths by role at Complete AI Training.
Your membership also unlocks:
