India AI infrastructure: Power and compute bottlenecks 2026

Shift from AI pilots to infrastructure-first scale-up

As companies move from experimenting with AI to deploying it at scale, attention is shifting from applications to the infrastructure that powers them. The core requirements are becoming more physical and capital-intensive: data centres, GPUs and compute capacity, cloud platforms, networking, storage, connectivity and energy. Experts cited in the material argue that the next chapter of India’s AI story will be shaped as much by investments in compute and connectivity as by progress in models and applications. The argument is straightforward: applications create demand and business value, but infrastructure determines how much AI activity a country can support and how quickly it can innovate.

Compute and data centres are increasingly “strategic assets”

One expert, Kharbanda, described compute capacity, GPUs, data centres, cloud platforms, networking and storage as strategic assets. The implication is that access, affordability and reliability of these inputs will decide who can run AI services efficiently and at scale. The material also notes that India will eventually need its own compute capacity and data infrastructure, and possibly indigenous large language models. That framing shifts the conversation away from which chatbot wins mindshare and toward who can provide affordable access to computing power.

Where investments are expected over the next 3 to 5 years

The experts highlighted that over the next three to five years, the biggest AI investments in India are likely to occur in data centres, cloud infrastructure and AI compute. These layers are described as the “foundation for everything else”, and the text flags that significant capital will be required for GPU clusters, networking infrastructure, storage systems and energy capacity. In practical terms, that links the AI investment cycle to construction, power procurement, grid connectivity, and large-scale hardware procurement.

The bottleneck is migrating from chips to memory to energy

A central claim in the material is that constraints in AI hardware have shifted over time. By 2025-2026, binding constraints are described as high-bandwidth memory (HBM) and energy infrastructure. It notes that only three companies worldwide can produce the HBM that AI needs, and supply is sold out through 2026. Separately, it states that 50% of global data center projects face delays due to power grid limitations. The conclusion presented is that the most consequential shift is from silicon to energy, with one analyst report quoted as saying the primary risk to AI expansion is no longer chip availability but the inability of electrical grids to meet exponential power demand.

A caution on fab-led AI sovereignty narratives

The material argues that India’s “$15 billion” buys chips “the AI industry doesn’t need”, at nodes that “don’t matter for AI”, and that they would arrive after the bottleneck has moved. It frames this as a structural issue, not a cyclical one, and warns against equating fabrication sovereignty with AI sovereignty. At the same time, it also stresses a separate takeaway: India has access to advanced AI compute today, but still needs to build domestic AI chip capability for tomorrow. The proposed long-term stack includes chip design, packaging, fabs, AI accelerators, talent and the broader semiconductor ecosystem.

India’s “real” opportunity areas above the chip

Instead of positioning fabrication as the primary AI hardware answer, the text points to domains where India’s software talent can drive advantage. It lists inference optimization software, edge AI deployment, energy-efficient data center architecture, and intelligent model orchestration. The rationale is efficiency: the layers above the chip determine how effectively AI compute is used. This also ties back to the view that the next AI race will be won by countries that control the full stack, from chips and data centers to power, cooling, software and supply chains.

Electricity becomes the hard constraint for AI scale

Multiple passages emphasise that AI is constrained by physical infrastructure, especially electricity. Advanced compute workloads require large, continuous and reliable power. The material projects global datacenter electricity demand to exceed 1,000 TWh by 2027, driven significantly by AI workloads. It proposes that AI infrastructure must be developed as a connected national system, from semiconductors to the grid, and says a grid-as-strategic-infrastructure framework should focus on three levers: generation scale-out, transmission modernisation, and intelligent distribution.

India’s data centre growth implies a major power build-out

The material provides several data points on India’s data centre expansion and its impact on electricity demand. India’s data centre capacity is expected to grow from 1.4 gigawatts last year to 9 GW by 2030, and to consume about 3% of India’s electricity by then, up from less than 1% currently. It also states a projection of nearly 1.5 GW today to about 10.5 GW by 2031, described as a six-fold increase in half a decade. Deloitte India’s estimate cited in the text suggests India will need 40 to 50 terawatt-hours of additional electricity to meet projected demand for AI-driven data centres by 2030.

Key numbers mentioned in the material

Market impact: what becomes investable and what becomes risky

The market-facing message in the text is that AI adoption at scale redirects capital and execution risk toward infrastructure. It identifies data centres, compute, networking and storage as core investment areas, but also highlights a rising risk factor: grid readiness. With half of global data center projects facing delays due to power grid limitations, the availability of power and the ability to connect and cool capacity becomes a gating variable. The material also points to transmission, storage integration, open access and distortions in industrial tariffs as constraints, noting that even the Draft National Electricity Policy 2026 acknowledges tariff-related issues.

Analysis: India’s cost of new power is an advantage, but grids matter

The text states that India has roughly 524 GW of installed capacity, but argues the key issue is the cost of adding incremental power, not legacy capacity. It notes that renewable bids in India, including renewable plus storage, have been clearing around ₹3 to ₹5 per kWh in recent tenders, positioning India as one of the few large economies that can still add power at globally competitive prices. But it also stresses that generation is not the only constraint. Transmission, storage integration, open access and tariff structure will determine whether energy remains a bottleneck or becomes a strategic advantage for AI infrastructure.

Conclusion: the next AI chapter is built on compute, grids and efficiency

The material repeatedly returns to one theme: AI runs on physical infrastructure, and electricity is central. Over the next few years, India’s AI trajectory is presented as depending on whether it can build affordable compute and data centre capacity while strengthening the power and connectivity backbone that supports them. It also argues for a more connected national approach that links semiconductors, cloud and data centres to generation, transmission and distribution planning. The next phase, as described, will be decided less by the most popular AI application and more by who can deliver reliable, affordable compute at scale.