Veeran Advisory · Rohith Devanathan · February 2026

Research Memo

India as a Global Inference Hub

A Strategic Memorandum

Executive Summary

Artificial intelligence is hitting physical limits. In the US and Europe, the constraint isn't talent or capital — it's power delivery, grid interconnection, permitting timelines, and land availability. In mature markets, getting a data centre connected to the grid now takes longer than building the facility itself. [1][2]

This matters because AI is shifting from training to inference. Training happens once — episodic, concentrated, capital-intensive. Inference happens continuously: every query, every prediction, every piece of enterprise automation. As models move from labs into production, inference becomes the dominant workload. Multiple industry assessments confirm this trajectory. [3]

The question this memorandum addresses: can India capture long-term AI value by positioning itself as a global inference hub?

We examine India's current positioning, global competitive dynamics, infrastructure bottlenecks in advanced economies, and the potential for state-level execution. Tamil Nadu is explored as a possible pilot.

Scope boundary: This focuses on export-oriented inference workloads that tolerate moderate latency and aren't sovereignty-sensitive (batch analytics, enterprise back-office, content workflows). It excludes ultra-low-latency applications like high-frequency trading or classified defence systems.

Glossary of Terms

Artificial Intelligence (AI) Software systems capable of performing tasks that typically require human intelligence: classification, prediction, reasoning, generation.

Training Building an AI model by exposing it to massive datasets and adjusting internal parameters. Capital-intensive, episodic, dominated by a handful of frontier organisations.

Inference The operational use of a trained model. Every prediction, classification, summary, anomaly detection, or recommendation is inference. Recurring and scales with adoption.

Compute Processing capacity for AI workloads, typically GPUs or specialised accelerators.

Latency Response time. Some applications need ultra-low latency. Many enterprise workloads tolerate moderate delay.

Throughput Total computational work processed over time.

Time-to-Power Elapsed time between project approval and delivery of reliable grid electricity at required capacity, including interconnection and grid reinforcement. [1]

Sovereign Compute Infrastructure hosted within a jurisdiction and subject to its regulatory control — data, security, export regimes. [4]

Structural Shift in AI Economics

The dominant narrative in AI focuses on frontier training: larger models, more parameters, proprietary datasets. But as AI moves from research to production, the constraints shift from intellectual to infrastructural.

Three structural changes are now visible:

1.1 Inference Becomes the Dominant Workload

Once a model is trained, it's used repeatedly. A single foundation model serves millions or billions of queries. The economic value shifts from training events to inference volume — especially in enterprise deployments like copilots, search, customer support, analytics, and embedded automation. [3]

1.2 Infrastructure Becomes Binding

Inference needs persistent access to power, cooling, networking, and physical data centre capacity. Unlike training (centralised and periodic), inference is continuous and depends on uptime, availability, and the cost and reliability of electricity. Data centres are capital-intensive infrastructure assets, and construction costs per MW have risen across many markets in recent years. [5][6]

1.3 Grid and Permitting Constraints Are Rising

In several advanced economies, grid interconnection delays now exceed data centre construction timelines. In European hubs, grid connection can take multiple years and is shaping where capacity can be deployed. [1] In the US, interconnection queue backlogs have grown substantially over the past decade, with large volumes of capacity awaiting connection. Power availability is becoming the decisive constraint. [2]

This shift means AI competitiveness increasingly depends on the ability to host reliable inference infrastructure at scale — not just frontier research capabilities.

Global Competitive Landscape

2.1 United States

The US dominates frontier model development and capital concentration. But hyperscale infrastructure growth is constrained by grid bottlenecks, local permitting resistance, and interconnection queue delays. [2] The US will anchor sensitive and ultra-low-latency workloads, but portions of non-sensitive, batchable, or moderate-latency inference demand will increasingly seek offshore capacity where time-to-power and total cost of ownership are superior.

2.2 European Union

Europe faces structural grid congestion and lengthy interconnection timelines in several markets. Legacy data centre hubs have long grid connection lead times, influencing future capacity location. [1] Europe will prioritise strategic autonomy and sensitive workloads onshore, but bulk inference growth faces structural friction — particularly in power-constrained metropolitan areas.

2.3 Middle East (UAE, Saudi Arabia)

Gulf states have rapidly deployed capital toward hyperscale data centre and AI infrastructure, aided by centralised decision-making and strong state capacity. Execution speed and energy cost profiles are competitive advantages. Long-term inference export positioning depends on integration with global enterprise ecosystems, trusted legal frameworks for cross-border data handling, and durable alignment with major software platforms and demand centres.

2.4 India

India presents a distinct structural profile:

Expanding renewable energy capacity and improving procurement pathways for large consumers
Growing data centre investment, with credible market research indicating operational capacity around 1–1.5 GW by the mid-2020s [6][7]
National coordination around AI infrastructure and compute provisioning through the IndiaAI Mission [4][8]
A large IT services and Global Capability Centre (GCC) base capable of integrating infrastructure with talent and operational delivery [9]
A federal structure enabling state-level experimentation and differentiated execution models

India is already building foundational capacity. The question is whether this capacity can be strategically positioned to capture recurring inference demand — both domestic and export-oriented.

India's Current Position and Momentum

India has initiated several relevant national efforts that form a credible foundation.

3.1 IndiaAI Mission

In March 2024, the Union Cabinet approved the IndiaAI Mission with a total outlay of ₹10,371.92 crore over five years. The programme includes pillars covering compute capacity, innovation, datasets, applications, skilling, startup financing, and safe AI. [4][8]

Government communications indicate an ambition to build a federated national AI compute network, with scale targets referencing around 38,000 GPUs as the mission scales. [4] This positions India with a nationally coordinated compute backbone supporting both training and inference.

3.2 Data Centre Policy Momentum

India's data centre market continues to expand, driven by cloud, digital services, and increasingly AI. Market research from established infrastructure analysts describes sustained growth in capacity and pipeline activity. [6][7] Multiple states have issued data centre policies offering incentives and facilitation mechanisms — land, stamp duty, electricity duty, renewable procurement pathways, single-window clearances — creating a platform for state-level execution experimentation.

3.3 Renewables and Power Procurement Pathways

India's renewable energy buildout and evolving procurement mechanisms (open access, group captive, PPAs) offer a pathway to improve power cost predictability and carbon positioning for large, power-intensive assets like data centres, subject to state regulatory specifics and grid constraints.

These initiatives show significant momentum. But they're primarily framed around domestic capability building and infrastructure attraction — not explicitly positioning inference as an export vector. A strategic reframing could complement rather than replace existing efforts.

Why Inference Exports Represent a Distinct Opportunity

Training dominance requires

Frontier research ecosystems and dense clusters of elite AI researchers
Proprietary datasets and extensive data acquisition capabilities
Extremely large capital pools and risk tolerance for multi-billion-dollar training runs
Preferential access to cutting-edge accelerators and deep semiconductor supply chain integration

Inference exports require

Reliable power at competitive and predictable prices
Predictable policy (tax, regulatory, data protection regimes)
Fast permitting and coordinated grid planning to minimise time-to-power
Enterprise-grade compliance and security frameworks
High uptime and operational excellence in data centre operations

These requirements align better with India's structural strengths — particularly when combined with its services ecosystem and large GCC base.

Inference workloads are recurring and operational, not episodic. They grow with AI adoption across enterprise and public-sector processes, are often compatible with moderate latency (especially back-office and batchable workloads), and become embedded into workflows, creating predictable long-term demand once integrated.

This suggests inference offers a scalable pathway for value capture where India can compete on reliability, speed of execution, and integration with talent and services — rather than on frontier training leadership alone.

Economic Considerations

Data centre economics are influenced by construction cost per megawatt of IT load, power cost and predictability over the contract horizon, time-to-power, utilisation ramp speed, and contract structure (pre-leasing, take-or-pay features, power pass-throughs).

Industry benchmarks indicate that construction cost per MW in mature markets has risen in recent years, with cost ranges varying by geography and supply chain conditions. [5][6] High utilisation drives operating leverage; mature operators typically target strong EBITDA margins on stabilised assets. [5] Payback and IRR expectations depend heavily on utilisation and time-to-power; multi-year grid delays can materially impair returns even when tenant demand is strong. [1][2]

If India can compress grid coordination timelines and standardise permitting at the state level, it achieves a structural advantage over congested markets — even when headline power prices are similar. The objective isn't to compete purely on cost. It's to compete on reliability, speed of execution, and integration with services and talent.

Avoiding the Commodity Trap

Hosting compute alone doesn't guarantee durable value capture. Risks include becoming a low-margin location where value accrues primarily to foreign hyperscalers and energy suppliers, hosting foreign-controlled infrastructure without building domestic operating capability, and capturing limited spillover benefits if workloads remain disconnected from domestic enterprises, GCCs, and public-sector transformation.

To mitigate this, inference positioning should be paired with:

Compliance alignment layers that reduce friction for regulated industries
Domestic operator capability building, including Indian-owned or JV data centre and managed service providers
Long-term enterprise contracts that justify resilient infrastructure investment
Integration with GCCs, enabling "inference-plus-services" exports rather than stand-alone hosting
Public-sector AI adoption to anchor base demand and create reference deployments aligned with national priorities [9]

This transforms hosting from commodity capacity into embedded economic infrastructure with stronger domestic spillovers.

State-Level Execution: The Tamil Nadu Lens

India's federal structure enables state-level experimentation. Tamil Nadu presents characteristics worth examining as a pilot environment for an "Inference Corridor".

7.1 Structural Advantages

Tamil Nadu is a leading industrial state with strong manufacturing and logistics density, major ports, and established administrative capacity. Chennai is a significant data centre and connectivity hub, including subsea cable landing relevance, which improves international network economics for export-oriented workloads. [10]

Tamil Nadu has published a dedicated data centre policy framework, providing a base for incentives and facilitation mechanisms, signalling intent to attract data centre investment. [10] Renewable generation scale and procurement pathways in the state can support renewable-linked power contracting where feasible under regulatory conditions.

7.2 A Tamil Nadu "Inference Corridor" Pilot

A state-level "Inference Corridor" operationalises the inference export thesis through a small number of highly prepared sites rather than dispersed projects. Key elements:

Pre-zoned, pre-cleared sites with defined environmental and building standards
Coordinated grid and substation planning, including redundancy design and published capacity roadmaps
Time-to-power service commitments for qualifying corridor projects (with transparent conditions and milestones)
A compliance and security stack to reduce friction for regulated but non-sovereign workloads
Integration with local talent and GCC ecosystems to enable compute + services exports
Anchor projects and bilateral positioning via consulates and central missions to secure one or two credible reference clusters

This pilot doesn't require national restructuring. It demonstrates how a state can compress time-to-power and package infrastructure, compliance, and talent into an export-ready offering.

7.3 Cooling and Energy Considerations

High-density inference clusters elevate cooling and energy constraints. Corridor design should explicitly anticipate modern cooling architectures (including liquid-assisted cooling) and align energy sourcing with renewable procurement where feasible.

Risks and Constraints

Grid expansion lagging relative to load growth in electrifying economies
Overconcentration in a small number of hyperscalers, limiting domestic bargaining power
Geopolitical sensitivity in cross-border AI hosting (export controls, localisation requirements, security concerns)
Technological shifts (model efficiency, on-device inference, new accelerator architectures) altering demand profiles and facility design assumptions

Any inference export positioning requires ongoing monitoring of platform dynamics, semiconductor supply chains, and regulatory developments in major demand centres, coupled with flexible state-level execution.

Implications for Capital Allocation

An inference-oriented strategy has direct implications for how different capital providers and operators coordinate.

Infrastructure capital

Data centres and digital infrastructure

Infrastructure investors underwrite long-term, high-utilisation campuses whose economics are highly sensitive to utilisation ramp and time-to-power. In markets with multi-year grid connection lead times, even strong tenant demand can be delayed by interconnection bottlenecks, impairing return profiles. [1][2][5]

Energy developers and power contracting

Inference corridors create opportunities to co-develop renewable capacity alongside data centre campuses, improving bankability through long-term offtake agreements and stabilising lifetime power cost. This requires alignment between corridor planning, open access rules, and grid reinforcement sequencing.

GCC operators and services firms

India's services ecosystem is positioned to bind compute to talent. Co-locating AI operations, integration, and managed services with inference infrastructure creates 'compute + services' export platforms rather than stand-alone hosting footprints, improving utilisation and strengthening domestic value capture. [9]

Hyperscalers and cloud platforms

Platform operators decide which workloads can be located offshore for inference and under what compliance conditions. Their decisions are sensitive to time-to-power, policy predictability, power sourcing options, and integration with global regions — more than marginal differences in land cost.

For India to position itself as a global inference hub, these actors cannot operate in isolation. The binding constraint is whether capital allocation converges on a small number of well-prepared corridors with coordinated grid planning, rather than being spread thinly across uncoordinated projects.

Conclusion

AI value capture is increasingly shaped by infrastructure constraints rather than algorithmic novelty. Countries able to reliably deliver power, grid capacity, land, cooling, and compliant data centre infrastructure at scale are well positioned to host inference-heavy workloads. [1][2]

India has already initiated foundational capacity building in compute, renewables, and data centre infrastructure, including a nationally coordinated IndiaAI Mission with a five-year outlay of ₹10,371.92 crore and an ambition to build a large federated compute network. [4] A strategic evaluation of inference exports — particularly through state-level pilots like a Tamil Nadu inference corridor — offers a realistic pathway for long-term positioning without requiring dominance in frontier model training.

This memorandum offers a framework for examination, not a prescriptive roadmap. The objective is to contribute to structured dialogue around AI infrastructure and capital allocation in the Indian context, and to inform discussions with policymakers, state governments, and industry leaders.

References

[1] Ember. Grids for data centres in Europe. June 2025. https://ember-energy.org/app/uploads/2025/06/Grids-for-data-centres-in-Europe.pdf
[2] U.S. Federal Energy Regulatory Commission (FERC). Explainer on the Interconnection Final Rule (Order 2023). 2024. https://www.ferc.gov/explainer-interconnection-final-rule
[3] McKinsey & Company. The next big shifts in AI workloads and hyperscaler strategies. December 2025. https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-next-big-shifts-in-ai-workloads-and-hyperscaler-strategies
[4] Government of India, Press Information Bureau. IndiaAI Mission / Compute Capacity releases. 2024–2025. https://pib.gov.in
[5] Houlihan Lokey. Real Estate Market Update: Data Centers. Summer 2024. https://cdn.hl.com/pdf/2024/real-estate-market-update-data-centers-summer-2024.pdf
[6] JLL. Global Data Center Outlook. 2026. https://www.jll.com
[7] CBRE India. India's Data Centre Market in a New Era. 2025. https://www.cbre.co.in/insights/reports/india-s-data-centre-market-in-a-new-era
[8] IndiaAI (official portal). IndiaAI Compute Capacity. https://indiaai.gov.in
[9] NITI Aayog. AI for Viksit Bharat. 2025. https://niti.gov.in
[10] Data Center Dynamics. Tamil Nadu Data Centre Policy coverage. https://www.datacenterdynamics.com