Chapter Essence
This chapter examines how Artificial Intelligence — a General-Purpose Technology (GPT) — is reshaping the global economy and charts a pragmatic, India-specific strategy for navigating rapid technological change and persistent uncertainty. The core argument: India must not replicate the unsustainable, capital-intensive, GPU-heavy frontier model approach of advanced economies.
Instead, the chapter makes the case for a bottom-up, frugal AI strategy — multiple sector-specific solutions under a single national vision — grounded in India's comparative advantages: deep human capital, heterogeneous domestic data, and strong institutional coordination. On governance, the chapter advocates sequencing over speed: enable experimentation first, scale proven solutions next, introduce binding obligations only where risks are most pronounced. The proposed data governance framework balances openness to cross-border flows with accountability and domestic value retention. The government's role is that of an enabler and coordinator, not a prescriptive regulator. The central message: India's opportunity lies in deploying AI in a way that is economically grounded and socially responsive.
73%
Data centres in High-Income Countries (June 2025)
88%
Organisations using AI in at least one function (McKinsey 2025)
3%
India's share of global data centres by count
2%
India's share of startups curating AI training data
40
Quarters simulated in the ABM stress-test model
24%
Annual data centre demand growth (conservative baseline)
₹8/kWh
Baseline power tariff in ABM simulation
100 Cr+
Indians with wired/wireless broadband access (Dec 2025)
1.8 Mn
Farmers served by AI agricultural networks across 12 states
58.4%
AI utilisation concentrated in High-Income Countries (Apr 2025)
The global conversation on AI has fundamentally shifted from speculation to adoption. AI is no longer a distant technology — it is being integrated into organisations worldwide, even if in an experimental capacity.
McKinsey Survey (1,993 firms, 2025): 88% of organisations surveyed reported using AI in at least one business function — up from 50% in 2022. Of these, 31% are in the process of scaling across the organisation, while 7% have fully deployed and integrated AI.
AI Adoption Trajectory (McKinsey)
- 2022: 50% of organisations using AI in at least one function
- 2023: 55%
- 2024: 72%
- 2025: 88%
Global AI Utilisation Distribution (April 2025)
- High-Income Countries: 58.4% of all AI usage
- Upper-Middle-Income Countries: 22.5%
- Lower-Middle-Income Countries: 18.7%
The US–China AI Duopoly and the Frontier Divide
While AI usage has grown globally, the capability to design and train large foundational models remains highly concentrated in a handful of firms. These firms exercise significant control over market resources — compute, data, specialised hardware, and deep pools of technical talent — erecting high barriers to entry for others.
Export Restrictions on Advanced Chips: Export controls on the most advanced processors required for frontier model development create a fundamental asymmetry. Most countries may end up participating in AI primarily as users, while a few will shape its trajectory, standards, cultural leanings, and pricing. This is a geostrategic tool akin to the semiconductor supply-chain leverage described in Chris Miller's "Chip Wars".
Key Structural Asymmetries in the Global AI Ecosystem
Capability
Frontier vs Application
- Frontier model development highly capital-, compute-, data- and energy-intensive
- Only a few firms globally can compete at the frontier
- Most countries remain application-layer participants
Labour
Scale vs Inclusion
- AI raises marginal productivity of capital over labour in white-collar sectors
- Rapid AI adoption risks displacing workers faster than the economy reabsorbs them
- Critical tension for labour-abundant India
Governance
Open vs Proprietary Models
- Proprietary models are opaque "black boxes" — users cannot know what changes are made
- Open-weight models offer lower entry barriers, greater adaptability, less vendor lock-in
- India must balance openness with stewardship
Infrastructure
Compute vs Resource Constraints
- AI data centres consume enormous electricity and water (up to 20 lakh litres/day per centre)
- Globally, data centres consume 56,000 crore litres of water annually
- Grid stability risks: 1,500 MW load loss event in Northern Virginia (July 2024)
Geopolitics
Strategic Autonomy vs Integration
- AI has emerged as a geostrategic asset
- Export controls, tech transfer restrictions, weaponisation of AI inputs
- Overdependence on foreign systems carries systemic risks
- Complete self-sufficiency neither feasible nor efficient
Policy
Regulation vs Innovation
- Regulatory compliance imposes fixed costs that scale poorly for smaller firms
- Frontier AI firms in high-income countries can absorb compliance costs; India's fragmented ecosystem cannot
- Minimal regulation creates systemic risks in critical sectors
Data Centre Geography (World Bank, June 2025): 73% of all data centres by count are located in High-Income Countries. China holds 11%, other Upper-Middle-Income Countries 11%, and India only 3%. Meanwhile, the United States (40%), EU (21%), Other HICs (17%), and UK (9%) dominate startups curating AI training data — India holds a mere 2% of this segment.
The economic implications of AI remain the subject of intense debate. Early evidence tempers some extreme predictions on labour disruption, yet the structural shifts underway demand careful policy attention.
Concept: The Productivity J-Curve
Historically, General-Purpose Technologies do not deliver immediate productivity gains. Adoption requires reorganisation of workflows, new complementary skills, and institutional adaptation — all of which take time. The result is a "J-curve": an initial dip or plateau in measured productivity, followed by eventual gains once the technology is sufficiently embedded. AI is expected to follow this non-linear adoption path.
- Early stage: Labour is complemented as organisations restructure around new systems
- Substitution pressures: Emerge only once the market saturates and productivity gains no longer lead to significant cost reductions
- Key uncertainty: The pace at which this transition occurs, and how rapidly labour markets can adapt
- T.S. Eliot paraphrase (Survey): "The change comes not in a single shock, but in a quiet, steady drift."
Emerging Labour Market Evidence
Yale's Budget Lab: The broader US labour market has NOT experienced discernible disruption due to AI so far — providing near-term reassurance, especially for labour-abundant economies like India.
Brynjolfsson, Chandar & Chen (2025): The difference in job prospects between occupations highly exposed to AI and those with relatively low exposure is minor at this stage.
Caution Warranted: The positive short-run findings do not invite complacency. Productivity gains from augmentation have a ceiling. The labour intensity of output is marginally declining in AI-exposed sectors. Unless labour adapts and new skills are acquired, reductions in labour-intensity of GDP will worsen over time.
Box XIV.1 — Structural Shift in US Professional Services (PBIS Sector)
An Autoregressive Distributed Lag (ARDL) model was estimated using US non-supervisory employment in Professional, Business, and Information Services (PBIS) for March 2016–July 2025, with a post-December 2022 regime dummy (the GenAI inflection point).
| Finding |
PBIS Sector |
Non-PBIS Sector |
Interpretation |
| Post-2022 indicator (short-run) |
+25.74 (significant) |
+1.58 (modest) |
Structural shift in PBIS; not a general economy-wide effect |
| Interaction (output × post-2022) |
-2.572 (significant) |
~0 (insignificant) |
Labour intensity of output has marginally declined in PBIS post-GenAI |
| Long-run AI coefficient |
+9.409 (significant) |
+1.90 (insignificant) |
Sustained structural change concentrated in AI-exposed sectors |
| Long-run interaction |
-0.938 (significant) |
~0 |
Reduced marginal employment responsiveness to output growth in PBIS |
Key Takeaway: Not an abrupt contraction in employment, but a meaningful change in how employment responds to output growth in the post-GenAI period. The PBIS sector is exhibiting a distinctly different adjustment pattern concentrated in sectors with high AI sensitivity.
Agent-Based Model (ABM) — AI Infrastructure Stress Test (Box XIV.2)
Concept: Agent-Based Model (ABM) for AI Compute Expansion
The ABM is a policy stress test (not a forecasting tool) that simulates how data centre and AI compute capacity expansion is shaped by the interaction of financial constraints, infrastructure readiness, and external supply dependencies over a 40-quarter (10-year) horizon.
- Agents: Data centre operators who individually face financing, GPU, and grid constraints
- Feedback loops: Financial conditions, grid availability, and hardware access interact and shape one another
- Goal: Surface structural vulnerabilities, identify policy leverage points, and reveal limits of price-based or subsidy-led approaches when constraints are binding
ABM Simulation Scenarios
| Scenario |
Parameters |
Key Finding |
| Scenario 1: Baseline |
Demand growth 24% p.a.; Power tariff ₹8/kWh; Borrowing cost 9% p.a.; India's GPU share 4% of global demand; Grid capacity added with 12-quarter lag |
Bottlenecks emerge sequentially — finance first, then GPU access dominates. Hardware supply uncertainty, not demand or power, is the central long-run limiting factor. |
| Scenario 2: High Foreign GPU Demand |
Foreign GPU demand grows at ~double the baseline rate |
Higher GPU prices worsen project economics, weigh on profitability, and make banks reluctant to finance. Financial bottleneck persists longer. Easing domestic finance alone is insufficient when hardware supply is externally constrained. |
| Scenario 3: High Domestic Demand + Liberal Financing |
India demand at 32% p.a.; India GPU share 6%; Investor hurdle rate lowered to 10% (vs 15% baseline) |
Finance bottleneck eases (2 fewer quarters of waiting), but GPU availability becomes the dominant binding constraint. Abundant capital and strong demand are insufficient to overcome geopolitical limits in advanced hardware supply chains. |
ABM Conclusion: Across all three scenarios, global GPU supply chain dynamics play a meaningful role in shaping India's AI infrastructure expansion. Conventional policy levers — easier finance, demand subsidies — must be complemented by measures that enhance supply-side resilience in access to advanced compute. India's AI strategy must, in the medium term, rely on diversified and resilient access to global compute, while the National Semiconductor Mission works toward longer-term domestic capability.
India's access to cutting-edge compute infrastructure is limited relative to the scale of its AI ambitions. The infrastructure constraints are not merely technical — they involve power, water, finance, and global supply chains simultaneously.
India's Data Centre Gap: India holds only 3% of global data centres by count (as of June 2025, World Bank), compared to 73% in High-Income Countries and 11% in China. This severely limits India's capacity for large-scale AI model training.
Training Data Startup Gap: India accounts for only 2% of global startups focused on curating AI training data — compared to 40% in the US and 21% in the EU. This represents a massively underutilised comparative advantage given India's data scale and diversity.
Infrastructure Constraints Facing India
Power
Energy Constraints
- AI data centres consume massive electricity — US EIA revised growth forecast from 2% to 5% p.a. by 2026 due to AI demand
- India's already strained power grid faces additional risk — Virginia 2024 event showed 1,500 MW load loss caused major frequency deviation
- Grid capacity added with a 12-quarter lag, creating mismatch with fast-growing demand
Water
Water Stress
- AI data centres consume up to 20 lakh litres of water per day
- Global data centres consume 56,000 crore litres of water annually
- Scaling AI data centres in India risks extraordinary stress on already strained groundwater and freshwater reserves
Finance
Capital Constraints
- IBM CEO questioned profitability of AI data centre boom
- Some firms projected to burn $270 bn+ by 2030 while pursuing compute infrastructure
- Tech groups have shifted $120 bn of AI data centre debt off balance sheets
- IMF flagged risk of financial contagion from debt-fuelled expansion
Hardware
GPU Supply Chain
- NVIDIA flagged rising cost of inputs (Q3 2026 earnings)
- OpenAI signed deals to purchase 40% of global high-bandwidth memory supply from just three manufacturers
- India's 4% share of global GPU demand leaves it vulnerable to external supply conditions
- Export restrictions on advanced chips constrain India's frontier AI ambitions
IndiaAI Mission — Key Elements
INDIAai Mission: The national umbrella initiative for coordinating India's AI ecosystem. Key functions include:
- Providing shared GPU access for startups and research institutions
- Pooling existing data centre capacity to create shared cloud compute infrastructure
- Establishing common platforms where open-source AI efforts can be coordinated and audited
- Systematically identifying successful bottom-up applications and encouraging their scaling
- Hosting a government-curated code repository to facilitate rapid experimentation
The National Semiconductor Mission (Chapter 8) is expected to evolve toward building domestic capabilities that can progressively meet a larger share of India's advanced compute demand over the long term, reducing vulnerability to global hardware supply chain disruptions.
Concept: Frugal AI / Jugaad AI — India's Bottom-Up Approach
The global AI ecosystem has diverged along two development paths. The Western path is top-down: frontier models, massive private capital, enormous compute expenditure, intellectual property concentrated in hyperscale firms. India's path, by strategic necessity, is bottom-up: distributed innovation across firms and sectors, strong state coordination, application-specific AI rather than frontier model supremacy.
- Frugal AI prioritises small, task-specific models tailored to defined uses and sectoral needs
- These models are more computationally efficient, easier to fine-tune, and capable of running on locally available hardware (smartphones, personal computers)
- They allow innovation from a broader set of actors: startups, research institutions, public agencies, domain-specific firms
- Enables secure deployment in sensitive sectors (healthcare, defence, public administration) via decentralised compute
- Late-mover advantage: India can avoid costly path dependencies and unsustainable design choices — energy-intensive architectures, opaque practices, ballooning financial commitments — that early adopters are now locked into
Box XIV.3 — Local Ingenuity: Frugal AI in Action Across India
Healthcare
Medical Diagnostics
- Southern India: Non-invasive AI-enabled thermal imaging for early breast cancer screening in low-resource settings — reduces dependence on expensive diagnostic infrastructure
- Eastern India: Portable, low-cost AI-assisted oral cancer screening devices in primary healthcare centres and outreach camps
Urban Management
Cities & Environment
- Bengaluru: AI-based water management systems monitoring consumption and detecting leakages in real time
- Himalayan region: Indigenous sensor networks with machine learning models providing real-time landslide alerts across vulnerable slopes
Agriculture
Farm & Market
- AI-enabled agricultural networks improved market access, price discovery, and logistical efficiency for 1.8 million farmers across 12 states
Education
Learning Analytics
- AI analytics monitoring classroom learning outcomes in Pimpri-Chinchwad — 18 classrooms, 3 schools
- Produced gains in student engagement, teacher accountability, and supervisory capacity
Language & Voice
Bhashini & AI4Bharat
- Bhashini (MeitY): Language and voice-first AI systems extending digital services to populations excluded from text-heavy platforms
- AI4Bharat (IIT Madras): Native language interaction, functioning on low-cost devices — frugal AI at scale
Open Source
Open-Source Advantage
- India is one of the world's largest and fastest-growing communities of open-source developers (GitHub Octoverse 2025)
- Open models have been consistently closing the performance gap vs closed proprietary models (EpochAI, 2024)
- India ranked #2 globally in AI literacy (after the US) as of 2024 (Stanford AI Index 2025)
Idea: The AI-OS Initiative
The chapter proposes an "AI-OS" initiative — a sovereign-backed platform analogous to how UPI and Aadhaar turned digital payments and identity into public goods. The sovereign acts as a monetary shareholder in the effort, turning AI into a public infrastructure layer.
- Structured datasets: Expand availability of anonymised, machine-readable datasets in priority sectors (health, agriculture, finance, education, governance)
- Pooled compute: Pool existing data centre capacity to create shared cloud compute infrastructure accessible to startups and researchers
- Code repository: Government-hosted, community-curated platform under IndiaAI Mission — a secure, transparent space where developers, researchers, and enterprises share code and build on each other's work (akin to GitHub for Indian AI)
- Open-weight models: Prioritise innovation on open-source and open-weight platforms so shared innovation enables India to achieve more with less
- Private sector participation: Large domestic firms absorb risk and scale successful applications — transforming India from "IT back office" to "AI front office"
In the AI era, data is a core factor of production. India's scale and diversity of domestically generated data constitute an important comparative advantage — but this asset remains largely underutilised.
India's Data Moat — An Untapped Strategic Asset
India's unique data advantage stems from the heterogeneity and scale of its domestic data sources across multiple critical sectors:
- UPI: World's largest real-time digital payments system — vast transactional behavioural data
- Aadhaar: 1.4 billion+ biometric identity records — unique population-scale dataset
- ONDC: Open Network for Digital Commerce — diverse commerce and preference data
- Health Records: Population-scale health and disease burden data across a highly diverse demographic
- Agriculture: Crop, weather, soil, and market data across varied agro-climatic zones
- 100 crore+ broadband users: Massive, growing base of internet-connected users generating behavioural data
- Training data scarcity: Current stock of AI training data is expected to run out; models collapse when trained on synthetic data — making India's human-generated data increasingly precious
Proposed Data Governance Framework — Three Core Objectives
1. Openness
Preserve openness to cross-border data flows, recognising their importance for innovation, investment, and global integration. No rigid data localisation mandates.
2. Regulatory Oversight
Ensure regulatory oversight and enforceability over large-scale processing and use of Indian personal data, irrespective of where the processing occurs.
3. Domestic Value Retention
Promote domestic value retention so that Indian data contributes meaningfully to the development of India's own AI capabilities and research ecosystem.
Framework Principles (Box XIV.6)
| Principle |
What It Means |
| Accountable Portability over Rigid Localisation |
Data may move across borders, but entities processing Indian data at scale must ensure auditability and traceability |
| Risk-Based Data Categorisation |
Data classified by sensitivity and economic significance — large-scale behavioural, transactional, personal, and inferred datasets receive differentiated treatment due to strategic AI training value |
| Graduated Obligations |
Regulatory requirements scale with risk and size — high-impact uses require enhanced transparency; startups and research institutions have eased compliance |
| Mirrored Data for Oversight |
Eligible entities maintain contemporaneous mirrored copies of relevant datasets within India — no mandates for domestic processing |
| Incentive-Compatible Value Retention |
Firms extracting significant commercial value from Indian data contribute to domestic AI ecosystem through flexible menu-based mechanisms (local model training, financial contributions, compute sharing, R&D investment) |
| Transparency-Centred AI Regulation |
Focus on dataset provenance, standardised model documentation, impact assessments for high-risk uses, and post-deployment monitoring |
| Positive Incentives over Prescriptive Mandates |
Participation in certified domestic compute/data environments is voluntary but rewarded with reduced audit burdens and faster clearances |
| Access as the State's Primary Lever |
Compliance linked to eligibility for government datasets, AI missions, regulatory sandboxes, and public procurement — shaping incentives without expanding statutory controls |
DPDP Act 2023 as Foundation: The framework builds on the Digital Personal Data Protection Act, 2023, expanding data categorisation to incorporate distinctions based on sensitivity and economic use. Subordinate legislation and rules allow regulatory adaptation without frequent statutory amendments.
Data Scarcity Warning: Research suggests the current stock of human-generated training data for LLMs is nearing its limits. Crucially, AI models "collapse" when trained recursively on synthetic data (Shumailov et al., Nature 2024). This makes India's unique, diverse, human-generated data an increasingly scarce and strategically valuable resource.
The labour market implications of AI are neither uniform nor predetermined — they are critically shaped by how skills, tasks, and institutions evolve together (Liu et al., 2025 — two centuries of technological change evidence). AI may follow a different trajectory from earlier automation waves: potentially increasing demand for experience-intensive roles rather than simply rewarding formal education.
Denmark Evidence (Renault 2025): Most Danish workers benefit from the adoption of AI. US PBIS Sector Evidence (Survey Box XIV.1): Not abrupt displacement, but a structural shift — the labour intensity of output has marginally declined in AI-exposed sectors, suggesting non-linear trajectories ahead.
Box XIV.4 — Where Human Value Lies in an AI-Driven Economy
AI as a Powerful Ship — Not an Independent Navigator
"AI is best understood as a powerful ship rather than an independent navigator: it can move faster and farther than any human, but without a knowledgeable captain who understands the vessel and the waters they are navigating, it is as likely to drift aimlessly as it is to arrive anywhere useful."
- Domain depth: AI can access vast troves of knowledge but lacks context and salience — humans must supply deep subject-matter understanding to frame right questions and evaluate outputs critically
- Continuous reading: Effective AI use requires frequent engagement with high-quality reading materials — AI amplifies the returns to prior knowledge
- System architects: Cognitive workers must decompose complex problems, sequence inquiries, impose constraints, define evaluation criteria — structured thinking over prompt engineering
- Cognitive atrophy risk: MIT and Microsoft studies independently find that dependence on AI for creative work and writing contributes to cognitive atrophy and deterioration of critical thinking
Human Capital Recommendations
Talent Pipeline
Diaspora & Industry-Academia
- Attract diaspora talent with hands-on experience building large models — have them train others (tacit "underground knowledge")
- Practitioner fellowships, flexible teaching roles for industry experts, structured apprenticeship models
- Reference: EU industry-academia collaborations, China's Young Thousand Talents Program
Earn-and-Learn
Early Industry Integration
- High-school, vocational, and early tertiary pathways integrated into credit-bearing industry fellowships from Class 11
- Students earn both academic credits and paid work experience through apprenticeships
- NEP 2020's MEME provisions, Academic Bank of Credits, National Credit Framework as policy foundation
- Viksit Bharat Shiksha Adhishthan Bill 2025 as first step in freeing universities to adapt curricula
Foundational Skills
Primary Education Reform
- As AI automates routine cognitive tasks, long-term talent utilisation depends more on foundational skills
- Prioritise: literacy, numeracy, reasoning, problem-solving, communication, socio-emotional skills, curiosity, self-regulation
- India Skills Report 2025 (Wheebox): Employers place highest emphasis on engagement, adaptability, and problem-solving
New Job Sectors
Understaffed High-Skill Jobs
- Nursing & Geriatric care: Already understaffed; India's dependency ratio will double in the next decade
- Other high-skill, long-apprenticeship sectors: culinary sciences, advanced metalwork, experiential hospitality, surgeons, physiotherapists, advanced electricians, early childhood educators
- Physical, human-centric, hands-on jobs hold immense potential for meaningful employment
Cognitive Risk of GenAI in Education: Niall Ferguson ("AI's Great Brain Robbery") and studies by MIT and Microsoft independently warn that widespread student use of GenAI as a substitute for creative and critical thinking contributes to cognitive atrophy. Combined with social-media-induced anxiety and depression, this may permanently dent employment prospects for a generation.
India's regulatory philosophy must be calibrated to its economic realities — neither the prescriptive omnibus approach of the EU nor the minimalist voluntary principles of the US, but a sequenced, risk-weighted, incentive-based framework.
Global Regulatory Approaches (Comparative)
- EU: Omnibus "EU Artificial Intelligence Act" — comprehensive mandatory regulation by risk tier
- China: Separate legislations governing separate AI applications
- United States: Guiding principles — voluntary and non-binding
- India's proposed: Risk-based, proportionate, incentive-compatible, sequenced framework via MeitY governance guidelines
Proposed AI Institutional Architecture
- MeitY AI Governance Group + Technical Committee: Light, incentive-based, risk-weighted governance foundation
- AI Economic Council (Box XIV.5): Coordinates technology deployment with education/skilling infrastructure — calibrates pace of AI adoption, assesses labour market impacts
- AI Safety Institute: Analyses emerging risks, conducts safety evaluations, red-teaming, scenario-based testing, transparency reporting
AI Economic Council — 5 Core Governance Principles (Box XIV.5)
| # |
Principle |
Meaning |
| 1 |
Human Primacy & Economic Purpose |
AI adoption must be subordinate to human welfare and economic inclusion; every major deployment must demonstrate a credible pathway to net social and economic benefit |
| 2 |
Labour-Market Sensitivity by Design |
AI policy must internalise India's labour structure (high informality, skill heterogeneity, limited safety nets) — labour impact assessments ex ante, with mitigation plans baked in |
| 3 |
Sequencing over Speed |
AI adoption phased in line with institutional readiness and skill pipelines — classify AI uses into 'deploy now', 'pilot', or 'defer' based on readiness |
| 4 |
Co-evolution of Technology & Human Capital |
Skill policy must stand equal to technology policy — AI push must proceed with parallel educational reform, vocational adaptation, and reskilling pathways |
| 5 |
Public Interest Safeguards & Ethical Non-Negotiables |
Strict lines around: surveillance misuse, worker monitoring, algorithmic discrimination, opaque decision making — no conception of safe AI is credible without protecting individual rights |
AI Safety Risks — Key Categories
Biosecurity
AI + Synthetic Biology
- Open-source CRISPR kits now accessible to hobbyist researchers and DIY scientists
- Combined with AI models generating genomic sequences and guiding gene-editing, the threat landscape changes dramatically
- A motivated individual with computing access could engineer pathogens — AI lowers the barrier to entry significantly
Behavioural Risk
Social Sycophancy
- Widely deployed models exhibit "social sycophancy" — over-affirming users' actions at rates significantly higher than human benchmarks (Cheng et al., 2025)
- This persists even in interpersonal harm or unethical behaviour contexts
- Increases user trust and reliance while reducing willingness to engage in corrective actions — a perverse incentive structure
Corporate Opacity
Big-Tech Safety Washing
- AI Lab Watch demonstrates big-tech firms obfuscate how evaluations are conducted, hide reasoning, and provide dubious interpretations
- Claims of implemented safeguards are unsubstantiated with public evidence
- Significant information gap between AI developers and end-users
Prohibited Applications
Non-Negotiable Restrictions
- Predictive policing
- Facial recognition misuse
- Exploiting psychological vulnerabilities
- Inferring emotions without consent
- Classifying individuals by behavioural or personality traits
International Cooperation on AI Safety: India's proposed Safety Institute should partner with established counterparts — UK's AI Security Institute and NIST's AI Risk Management Framework (US) — for joint evaluations of high-risk models, shared computing access, and globally interoperable AI safety standards. India does not need to do this in isolation.
"Regulate and Deploy Simultaneously": A central insight of the chapter. Unlike early AI adopters who scaled in a regulatory vacuum and are now locked into unsustainable commitments, India can sequence regulation alongside deployment — designing more resource-efficient systems aligned with public objectives from the outset. Being a late mover, done right, is an advantage.
The chapter proposes a phased, sequenced roadmap for India's AI strategy — building coordination first, capacity next, and binding policy leverage last, allowing institutions and markets to co-evolve.
Three-Phase Roadmap
| Phase |
Focus |
Key Actions |
| Phase 1 — Coordination & Experimentation |
Operationalise existing institutions; align incentives; enable bottom-up innovation |
- Expand IndiaAI Mission shared infrastructure
- Launch government-hosted community code repository
- Evolve DPDP framework via subordinate legislation for functional data categorisation
- Scale earn-and-learn pathways and curricular flexibility
- Focus on application/sector-specific, small, open-weight models
|
| Phase 2 — Selective Scaling & Formalisation |
Medium-term; evidence from Phase 1 guides scaling decisions |
- Expand shared and certified domestic computing infrastructure
- Voluntary participation by large firms linked to regulatory facilitation
- Formalise risk-based, proportionate AI regulation
- Codify graduated obligations for AI firms by scale and sector
- Embed oversight within existing sectoral regulators (not single omnibus AI law)
- AI Safety Institute moves from analysis to structured scenario testing and red-teaming
|
| Phase 3 — Resilience & Sustained Adaptation |
Long-term; strategic autonomy and labour market resilience |
- Strategic partnerships and diplomacy for advanced computing hardware access
- Reduce vulnerability to external shocks in AI supply chains
- Sustained adaptation of labour markets and education systems
- Primary education prioritises foundational cognitive and socio-emotional skills
- National Semiconductor Mission builds domestic advanced chip capabilities
|
India's Comparative Advantages in the AI Era
Human Capital Depth:
- Ranks among top global contributors to AI research output (Georgetown University)
- Deep pool of AI technical talent globally (Paulson Institute Global AI Talent Tracker 2.0)
- Ranked #2 globally in AI literacy after the US (Stanford AI Index 2025)
- World's largest and fastest-growing open-source developer community (GitHub Octoverse 2025)
Strategic Positioning:
- Late-mover advantage: Learn from early adopters' costly path dependencies
- Diverse domestic data across health, agriculture, finance, education, public administration
- Strong institutional coordination capacity (demonstrated by UPI, Aadhaar, ONDC)
- Application-led innovation rather than frontier model supremacy
India's Strategic Choice Framed (Para 14.96): "AI does not confront India with a single policy question, but a series of choices that must be made under conditions of heightened uncertainty and resource constraints. The central challenge for India is in what it builds domestically, what it sources globally, what it regulates early, and what it deliberately allows to evolve. Passive consumption is the riskiest position of all."
Late Mover as Opportunity (Para 14.97–14.99): Early AI adopters scaled under a regulatory vacuum and cheap capital — they are now locked into energy-intensive architectures, mounting financial commitments, and unclear revenue pathways. Some governments are even discussing backstops against potential fallout. India has the benefit of hindsight. It can design AI systems that are more resource-efficient and aligned with public objectives from the outset — sequencing regulation alongside deployment. Late adoption need not imply lagging ambition.
Practice MCQs — UPSC Prelims Style
Q1. In the context of the Economic Survey 2025-26, Artificial Intelligence (AI) is classified as a "General-Purpose Technology (GPT)". Which of the following characteristics best defines a General-Purpose Technology?
- (a) A technology developed exclusively by governments for public sector use
- (b) A technology with pervasive applications across multiple sectors that enables and requires complementary innovations to deliver economic gains over time
- (c) A technology that delivers immediate and uniform productivity gains across all sectors upon adoption
- (d) A technology whose development requires frontier-scale compute and capital, limiting its use to advanced economies
Correct Answer: (b) The chapter treats AI as a General-Purpose Technology — one that is pervasive across sectors and enables a wide range of complementary innovations. Like earlier GPTs (electricity, the internet), AI does not deliver instant uniform productivity gains; it requires organisational adaptation and complementary investments, leading to the characteristic Productivity J-Curve. GPTs are not government-exclusive and are not limited to advanced economies.
Q2. The concept of the "Productivity J-Curve" in the context of AI adoption refers to which of the following phenomena?
- (a) An immediate surge in productivity followed by a gradual decline as AI systems become obsolete
- (b) A situation where only the top 10% of firms benefit from AI while others experience productivity losses
- (c) An initial dip or plateau in measured productivity gains following AI adoption, before eventual gains materialise as the technology becomes embedded and complementary skills develop
- (d) A pattern where AI productivity gains are limited to J-sector (Joint-sector) firms only
Correct Answer: (c) The Productivity J-Curve describes the non-linear adoption path of General-Purpose Technologies. Early adoption requires reorganisation of workflows and development of complementary skills, which initially suppresses measured productivity. The Survey cites the change in how PBIS sector employment responds to output as evidence of this structural shift — the change comes "not in a single shock, but in a quiet, steady drift" (paraphrasing T.S. Eliot).
Q3. According to the Economic Survey 2025-26 data (World Bank, June 2025), what is India's share of global data centres by count?
- (a) 11%
- (b) 7%
- (c) 5%
- (d) 3%
Correct Answer: (d) As per the World Bank data cited in Chart XIV.2 of the Economic Survey 2025-26, India holds only 3% of all global data centres by count as of June 2025. High-Income Countries hold 73%, China 11%, and other Upper-Middle-Income Countries 11%. This data centre gap is a key constraint for India's AI infrastructure ambitions.
Q4. The IndiaAI Mission, as discussed in the Economic Survey 2025-26, is primarily intended to serve which of the following functions?
- (a) Develop frontier AI models to compete with US and Chinese systems at the global level
- (b) Coordinate a bottom-up AI strategy by providing shared infrastructure, pooled compute access, common open-source platforms, and governance frameworks to scale diverse sector-specific AI solutions
- (c) Regulate all AI deployments in India through a mandatory licensing regime
- (d) Exclusively manage India's AI exports to other developing countries under technology transfer agreements
Correct Answer: (b) The IndiaAI Mission is positioned as the umbrella initiative for coordinating India's bottom-up AI strategy. It provides shared GPU access for startups, pools data centre capacity into shared cloud infrastructure, establishes open-source coordination platforms, and hosts a government-curated code repository. It is explicitly NOT about chasing frontier models but about democratising access to AI infrastructure for diverse sector-specific applications. The chapter proposes an "AI-OS" initiative under this umbrella — turning AI into a public good like UPI and Aadhaar.
Q5. The Digital Personal Data Protection (DPDP) Act, 2023, is referenced in the Economic Survey 2025-26 chapter on AI. The proposed data governance framework for the AI era, building on the DPDP Act, is best described as which of the following?
- (a) Mandatory data localisation requiring all processing of Indian data to occur within India's borders
- (b) Complete liberalisation of cross-border data flows with no regulatory oversight requirements
- (c) Accountable portability with risk-based categorisation, graduated obligations, and mirrored data copies within India — ensuring regulatory oversight without mandating domestic processing
- (d) A uniform compliance regime applying identical obligations to all firms regardless of size, sector, or data sensitivity
Correct Answer: (c) The proposed framework consciously avoids rigid data localisation mandates. Instead, it establishes "accountable portability" — data may move across borders, but entities processing Indian data at scale must ensure auditability, traceability, and a mirrored copy within India for regulatory oversight. Obligations are risk-weighted and graduated — scaling with size, scope, and sensitivity of data use. Startups and research institutions face lighter compliance. The DPDP Act 2023 provides the foundation, with expansion via subordinate legislation.
Q6. The "Frugal AI" or bottom-up approach to AI development advocated by the Economic Survey 2025-26 for India is characterised by which of the following features?
- (a) Massive centralised data centres, frontier model development, and heavy private capital investment
- (b) Complete reliance on open-source models developed abroad, with no indigenous innovation component
- (c) Small, task-specific, application-oriented models that run on locally available hardware, enabling distributed innovation from startups, research institutions, and domain firms without requiring expensive compute infrastructure
- (d) A model focused exclusively on defence and strategic sectors, with AI kept away from commercial applications
Correct Answer: (c) Frugal AI / Jugaad AI prioritises small, computationally efficient, application-specific models tailored to sector-specific needs. These models can run on smartphones or personal computers — eliminating the need for energy-intensive large data centres. Examples from the chapter include AI-powered thermal imaging for breast cancer screening in southern India, landslide alert systems in the Himalayas, and agricultural networks serving 1.8 million farmers. Bhashini and AI4Bharat (IIT Madras) are flagship examples of this approach. India is one of the world's largest open-source developer communities, making this approach strategically well-suited.
Q7. The Agent-Based Model (ABM) used in the Economic Survey 2025-26 to assess India's AI compute expansion (Box XIV.2) simulated 40 quarters (10 years) of capacity expansion under three scenarios. Which of the following conclusions was common across all three scenarios?
- (a) Financial constraints are the permanently binding bottleneck throughout the simulation horizon
- (b) Grid/power availability is the dominant constraint, indicating that India must invest primarily in energy infrastructure
- (c) Global GPU supply chain dynamics play a meaningful role in shaping India's AI infrastructure expansion, even when domestic demand is strong and financial conditions are favourable
- (d) India's AI compute expansion faces no significant barriers if the government provides sufficient fiscal support
Correct Answer: (c) The ABM stress-test is explicit that across all three scenarios — baseline, high foreign GPU demand, and high domestic demand with liberal financing — developments in global GPU supply chains consistently play a meaningful role in limiting India's AI infrastructure expansion. In the baseline, GPU access becomes the dominant long-run bottleneck. In Scenario 2, elevated foreign GPU demand keeps financial and hardware constraints binding simultaneously. In Scenario 3, even abundant capital and high demand cannot overcome hardware supply limits. The policy implication: conventional levers (easier finance, demand subsidies) must be complemented by supply-side resilience measures for advanced compute access.
Q8. According to the Economic Survey 2025-26, which of the following best captures the chapter's recommended approach to AI regulation in India?
- (a) Adopt the EU's Artificial Intelligence Act model — a comprehensive omnibus law with strict mandatory requirements across all AI applications
- (b) Follow the US approach of voluntary, non-binding guiding principles with no statutory AI regulation
- (c) Impose a moratorium on all AI deployment until comprehensive safety regulations are in place
- (d) A risk-weighted, sequenced approach: enable experimentation first, scale proven solutions next, introduce binding obligations only where risks are most pronounced — simultaneously deploying and regulating AI rather than waiting for full regulatory certainty
Correct Answer: (d) The chapter explicitly advocates a middle path — neither the EU's prescriptive omnibus approach nor the US's minimalist voluntary principles. India should "regulate and deploy simultaneously" rather than waiting for perfect regulatory frameworks before adopting AI. The sequencing principle — experimentation first, scaling next, binding obligations last — allows institutions and markets to co-evolve. MeitY's AI Governance Guidelines and the proposed AI Economic Council and AI Safety Institute form the institutional architecture. Oversight should be embedded within existing sectoral regulators rather than through a single omnibus AI law, with obligations graduated by risk and scale of AI application.
