The country that gave the world its engineers is struggling to build its own AI. Here is the uncomfortable truth about why — and what it will take to change it.
India produces nearly a million STEM graduates every year. It is the world’s largest user of ChatGPT on mobile. Its software engineers power Silicon Valley. And yet, in 2026, the country still had to scramble to assemble a handful of foundational AI models for a government summit — while the United States and China have been shipping frontier intelligence for years. Something is structurally wrong, and it is time to name it.
The gap is no longer a matter of aspiration or intent. It is a matter of hard numbers, institutional culture, and two decades of choices that optimised for cost arbitrage instead of original creation. The Indian IT industry, worth nearly $300 billion, finds itself at a crossroads — adapt or be rendered redundant by the very wave of automation it helped deploy for global clients.
The Outsourcing Trap
For thirty years, India’s IT industry operated on one genius insight: talented engineers at a fraction of Western wages. It was a model that built global empires — TCS, Infosys, Wipro — and lifted millions into a new middle class. But it was, at its core, a labour arbitrage play, not an innovation engine.
The problem with labour arbitrage is that it gets competed away. First by cheaper geographies, and now — fatally — by machines. A recent Citrini Research report laid it bare: the entire model was built on one value proposition — Indian developers costing a fraction of their American counterparts. AI coding agents now perform similar work at near-zero marginal cost. Major global clients are re-evaluating outsourcing contracts, some walking away or renegotiating at steep discounts.
TCS saw annual revenue slip year-over-year and its CEO admitted seeing “degrowth”. HCL warned of three to five percent revenue dips due to “AI deflation”. The bill for three decades of playing it safe has arrived.
Compiled from Q4 FY2026 earnings reportsThe tragedy is that none of this was unforeseeable. The signs were there throughout the 2010s — cloud computing eroded on-premise infrastructure work, agile replaced waterfall, no-code tools bit into low-complexity development. Indian IT adapted at the margins, never at the core. The core remained: hire, staff, deliver, bill by the hour.
The Indigenous Model Problem
In February 2026, India hosted its AI Impact Summit in New Delhi and invited China — an acknowledgement that the world’s largest democracy has much to learn from a state that treated AI as national infrastructure from day one. What India could showcase were a handful of foundational models shortlisted just months prior.
The IndiaAI Mission — launched in 2024 with a ₹10,372 crore allocation — did produce real results. Sarvam AI, BharatGen, Gnani, and Socket launched models with genuine strength in Indic language benchmarks. Sarvam AI’s models showed impressive accuracy in document understanding across India’s many languages. These are real wins.
Despite India being the largest user of ChatGPT on mobile globally (13.5% of all users) and the third-largest user of China’s DeepSeek, the adoption of indigenous Indian LLMs remains negligible. Indians enthusiastically consume American and Chinese AI — but rarely reach for something built at home.
But these are startups running on government grants — not the $300 billion IT services industry deploying its full weight. TCS, Infosys, and Wipro have been conspicuously absent from the race to build foundation models. They are positioning themselves as implementers of AI built elsewhere — wrapping GPT-4 and Anthropic models in an AI veneer. It keeps the lights on. It does not build sovereign capability.
Why the Gap Persists: Five Structural Reasons
- A culture of delivery, not discovery. Indian IT firms are masterclasses in project execution. They are not designed for the long, expensive, uncertain loops of foundational research. AI model development requires years of compute spend with no guaranteed billable outcome — anathema to the quarterly-revenue-focused services model.
- Compute poverty. Training large AI models requires GPU clusters at scale India simply hasn’t had. Even after provisioning 38,000+ GPUs under IndiaAI Mission, projections suggest India needs closer to 3.3 million GPUs by 2030. The sovereign compute gap is enormous — and hyperscalers won’t fill it.
- Brain drain to the model builders. India’s best AI researchers work at OpenAI, Google DeepMind, Meta FAIR, and Anthropic. The incentives — compensation, compute access, peer collaboration — are structurally superior in the West. Retaining world-class AI talent at home requires world-class conditions India has yet to create at scale.
- Ad-hoc innovation, not systematic R&D. Most Indian AI initiatives remain stuck at proof-of-concept stage, unable to scale due to legacy IT integration issues, poor data standards, and unclear governance. NASSCOM reports AI-focused R&D investment rose 20% in FY25 — encouraging, but still a fraction of what frontier labs spend in a single quarter.
- Policy incoherence. India’s AI policy features competing visions across Niti Aayog, MeitY, the Economic Survey 2026, and the AI Impact Summit. A comprehensive industrial strategy spanning promotion, regulation, and accountability remains a work in progress.
The Global Comparison That Should Sting
| Dimension | USA | China | India |
|---|---|---|---|
| Foundation Models | GPT-4o, Claude, Gemini, Llama | DeepSeek, Qwen, Ernie | Sarvam, BharatGen (emerging) |
| National AI Strategy | Unified + private sector | AI+ Initiative, sector-wide | Fragmented across ministries |
| Sovereign GPU Compute | Massive investment | State-directed ecosystem | 38K GPUs (need 3.3M by 2030) |
| AI R&D Investment | High & growing | Mandated & growing | Up 20% FY25 — from low base |
| AI Startup Ecosystem | Dominant globally | State-backed, scaling fast | $1.25B funding (2025) — 2× growth |
China’s approach is especially instructive — and uncomfortable. Beijing systematically integrated AI into healthcare, manufacturing, transportation, and governance through a top-down AI+ Initiative. India, by contrast, was “hoping” to have foundational model applicants ready for its summit. The ambition gap is not one of talent — it is structural and strategic.
The Green Shoots Worth Watching
It would be inaccurate to paint only failure. The IndiaAI Mission’s compute portal treats GPU access as a public good — available to startups, researchers, and academics. The Union Budget 2026–27 introduced long-term tax holidays for data centre investments. The AI Impact Summit secured $250 billion in AI infrastructure commitments, signalling that global capital sees India as a viable AI destination.
Within large IT firms, action is shifting. TCS has established a dedicated AI and services transformation unit. Infosys is developing over 100 generative AI agents for client workflows and building vernacular AI models in partnership with IIT Kharagpur and IISc Bangalore. Thirty AI solutions spanning agriculture, climate, and healthcare — from AI-powered soil testing to early TB detection — are in active development under IndiaAI’s programme.
The biggest threat to Infosys is not Wipro — it is a 22-year-old founder in San Francisco with GPT-4 and an internet connection. Indian IT must accept it cannot scale with headcount anymore.
Industry Commentary, 2025What the Next Five Years Must Look Like
- Fund indigenous models as national infrastructure, not just R&D bets. The government must treat LLM development the way it treats highways — as sovereign critical infrastructure requiring long-term committed spend.
- Create world-class conditions to retain AI talent. Compute access, competitive compensation at national labs, and research environments that can compete with the pull of Palo Alto and London.
- Force large IT firms to invest in IP creation. Tax incentives and government procurement mandates should favour companies building proprietary AI — not just reselling foreign APIs.
- Unify the policy architecture. A single high-powered AI authority — not competing committees across Niti Aayog, MeitY, and the PMO — must drive coherent industrial strategy with hard accountability metrics.
- Build for Bharat, not just for global clients. India’s 22 official languages, 1.4 billion citizens, and unique sectoral needs are not weaknesses — they are the most differentiated AI training environment on Earth. That is a competitive advantage no one else has.
The Bottom Line
India is not unintelligent about AI — it is structurally mis-incentivised. Three decades of outsourcing success created institutions optimised for execution at scale, not for the long-horizon, compute-intensive research that produces frontier AI models. Changing that requires rewiring how India’s technology industry thinks about value creation.
The good news: India has done this before. The country that seemed forever destined to be a low-cost back office eventually produced global CEOs, world-class pharmaceutical generics, and a Moon mission that cost $75 million. The capacity for transformation exists. But AI waits for no one — the compounding advantages of the US and China grow with every training run. India’s clock is ticking.
The question is not whether India can build world-class AI. It is whether India will choose to — urgently enough, coherently enough, and boldly enough to matter on the global stage.