Image Source: Unsplash
India’s artificial intelligence (AI) ambitions are substantial and increasingly visible. The country has emerged as a major hub for AI applications, digital public infrastructure, and large-scale deployment of AI-enabled services. AI adoption across healthcare, agriculture, logistics, education, finance, and governance is accelerating, supported by India’s vast digital population, rich multilingual data, and deep engineering talent.
Yet beneath this optimism lies a structural weakness: India remains heavily dependent on foreign computational infrastructure. It has talent, data, and growing demand for AI—but lacks compute capacity commensurate with its ambitions. This gap may prove decisive.
The global discourse on AI has largely been dominated by models—oscillating between the promise of large language models, the risks of autonomous systems, and the race towards artificial general intelligence. Governments debate regulation, companies unveil ever-larger models, and investors pour capital into AI applications. But beneath this visible layer lies a deeper strategic reality: AI power is increasingly shaped not merely by who designs the best algorithms, but by who controls the infrastructure that makes those algorithms possible.
In the emerging political economy of AI, compute—advanced chips, GPU clusters, hyperscale data centres, high-speed interconnects, cooling systems, and reliable energy—is becoming the primary bottleneck for frontier AI development. Data and algorithms remain indispensable, but unlike data, compute is scarce, capital-intensive, and geographically concentrated. While model architectures can often be replicated, training and deploying frontier AI systems require access to enormous computational resources, industrial supply chains, and power infrastructure.
This marks a profound shift in digital power. The internet era was shaped by control over data and platforms. The AI era is increasingly defined by control over compute.
From Data Sovereignty to Compute Sovereignty
For much of the last decade, digital policy centred on data sovereignty—where data is stored, how it moves, and who profits from it. Today, data alone is insufficient. A country may possess vast data and engineering talent yet remain dependent without computational infrastructure, making compute sovereignty a defining strategic question of the AI age.
Compute sovereignty is a nation’s ability to access, govern, and scale critical computational resources without excessive external dependence. It ranges from full sovereignty (domestic control over critical compute infrastructure) to strategic autonomy (secure access through trusted alliances and diversified supply chains), and finally dependency (reliance on external providers), with most countries falling into the latter two categories.
Compute sovereignty is a nation’s ability to access, govern, and scale critical computational resources without excessive external dependence.
Compute extends far beyond GPUs. It depends on advanced chips, AI accelerators, cloud infrastructure, high-bandwidth memory, advanced packaging, ultra-fast interconnects, and thermal systems—critical layers concentrated in a handful of firms and countries. Such concentration creates strategic vulnerability, much like in the energy, manufacturing, and telecommunications sectors.
This risk is already visible. A small number of firms control frontier AI compute, turning it into a geopolitical chokepoint akin to rare earths or advanced lithography. United States (US) chip export controls on China reflect a simple logic: constrain compute, and one constrains frontier AI capability.
The scale is staggering. Training frontier models requires tens of thousands of GPUs, with clusters of more than 100,000 GPUs becoming the new benchmark. OpenAI’s GPT-4 likely used tens of thousands of NVIDIA A100-class GPUs, while xAI targets superclusters of up to one million GPUs. At US$25,000–40,000 per H100 GPU, a 100,000-GPU cluster costs billions before accounting for power, cooling, and networking.
Inference adds another challenge: deployment costs increasingly match or exceed training costs. For developing countries, foreign cloud infrastructure offers scale but creates long-term vulnerabilities through pricing power, embedded tech stacks, and limited bargaining power. Open-weight models lower entry barriers but do not eliminate compute dependence.
Left unchecked, compute dependency could deepen global inequality, dividing the world between compute-rich powers that build foundational AI and compute-poor economies limited to downstream adaptation.
The core question is not whether countries will rely on global infrastructure—they will—but whether they can retain strategic autonomy. Left unchecked, compute dependency could deepen global inequality, dividing the world between compute-rich powers that build foundational AI and compute-poor economies limited to downstream adaptation. For India, the challenge is not necessarily training the largest models but securing sovereign inference capacity to scale AI domestically and avoid becoming merely a renter in the AI economy.
The New Geography of AI Power
AI compute is highly concentrated. The United States remains dominant, backed by semiconductor leadership, hyperscale cloud infrastructure, capital, and research institutions. NVIDIA controls roughly 80–90 percent of the advanced AI accelerator market, while Amazon Web Services, Microsoft Azure, and Google Cloud dominate global cloud infrastructure. AI dependency now extends beyond chips to cloud stacks, storage, networking, and AI tooling.
AI competition increasingly resembles an industrial arms race. China is investing heavily in indigenous compute, with firms such as Huawei building domestic AI accelerators and cloud ecosystems, recognising that AI sovereignty depends on control over silicon, compute, cloud, and energy.
Beyond these two powers, a widening compute divide is emerging. Europe retains strong research capabilities but lags in chips and cloud infrastructure, while middle powers such as Japan, South Korea, Singapore, the United Arab Emirates (UAE), and Saudi Arabia are treating compute as strategic infrastructure. This divide separates countries that generate AI power from those that merely rent it.
The semiconductor supply chain exposes further fragility. Critical chokepoints persist in chip design, fabrication, lithography, packaging, and memory. TSMC produces around 90 percent of the most advanced chips, while ASML dominates advanced lithography, and high-bandwidth memory is concentrated among firms such as SK hynix. Any disruption would reverberate globally.
Although algorithmic advances—seen in Llama from Meta, models from Mistral AI, and architectures from DeepSeek—improve efficiency through quantisation, distillation, and sparse architectures, they do not diminish the importance of compute. They only increase their productivity. As AI scales across billions of users and devices, compute demand will continue rising. The vital question is whether countries can secure enough compute to remain meaningful players in the AI economy.
India’s AI Ambition and Infrastructure Deficit
India’s AI ambition faces a core infrastructure challenge. Its AI market is poised for strong growth, and the success of digital public infrastructure—anchored by National Payments Corporation of India’s Unified Payments Interface, digital identity, and public digital rails—has positioned India as a potential global AI use-case leader. Yet this optimism is incomplete.
India remains dependent on imported advanced chips and foreign cloud infrastructure. Despite the India Semiconductor Mission and state incentives, domestic semiconductor manufacturing remains nascent, with no approved project matching advanced-node fabrication in Taiwan, South Korea, or the United States. The IndiaAI Mission, with over INR 10,000 crore in funding, aims to address this through a national AI compute facility, but it remains modest relative to frontier global clusters.
The challenge extends beyond chips to energy. AI-scale data centres require massive amounts of electricity, cooling, land, water, and grid reliability, with advanced facilities consuming hundreds of megawatts and next-generation campuses exceeding one gigawatt. As AI scales, compute increasingly becomes an energy—and often a water—question.
The strategic challenge is therefore not just building data centres, but building sustainable AI energy systems through renewables, storage, transmission upgrades, and grid modernisation.
For India, where power demand is already rising due to industrialisation, urbanisation, and digital expansion, AI infrastructure will intensify pressure on the grid. The strategic challenge is therefore not just building data centres, but building sustainable AI energy systems through renewables, storage, transmission upgrades, and grid modernisation. Without overcoming these infrastructure bottlenecks, India risks becoming a major AI market but not a major AI power. A large market attracts AI applications; tactical capability requires infrastructure ownership—or trusted access.
A Compute Strategy for India
India’s AI strategy requires a conceptual shift: beyond talent, startups, regulation, and use cases, AI policy must treat compute as strategic national infrastructure.
India should pursue five priorities. First, build sovereign and federated AI compute clusters through public-private partnerships for startups, universities, and public institutions. Second, establish a national compute governance framework or sovereign AI compute grid to allocate subsidised compute for academia, startups, defence, and public missions. Third, focus semiconductor policy on realistic strengths—advanced packaging, chip design, compound semiconductors, AI hardware, and electronics manufacturing. Fourth, develop AI energy zones with reliable power, renewable integration, cooling, and transmission capacity for hyperscale infrastructure. Fifth, build a public inference layer, positioning India as a leader in efficient inference, multilingual AI, edge AI, and small language models. The more achievable route to AI leadership may lie not in training the largest models, but in mastering deployment at scale.
The more achievable route to AI leadership may lie not in training the largest models, but in mastering deployment at scale.
India must also deepen trusted partnerships with the United States, Japan, and Europe to strengthen semiconductor resilience, cloud capacity, and AI research—while ensuring such partnerships expand domestic capability rather than create new dependence.
Ultimately, compute must be treated as foundational infrastructure. Just as digital public infrastructure transformed payments and identity, sovereign AI compute could become the next core digital utility. The AI race is no longer just about who writes the smartest code, but who controls the infrastructure—chips, clouds, data centres, and power systems—that enables intelligence at scale. For India, the imperative is clear: if it seeks to shape, not merely consume, AI, compute sovereignty must move to the centre of national strategy. The next digital divide may be defined not by connectivity, but by compute.
Amal Chandra is an Indian author, public policy analyst, and political commentator.
The views expressed above belong to the author(s). ORF research and analyses now available on Telegram! Click here to access our curated content — blogs, longforms and interviews.