VEERAN ADVISORY
← All research

№ 01 · Feb 2026 · ALLOCATORS · OBSERVERS

India as a Global Inference Hub

The countries that deliver power fastest, not the best models, will capture the inference economy. India's structural case, with Tamil Nadu as a pilot corridor.

Executive Summary

Artificial intelligence is hitting physical limits. In the United States and Europe the constraint is no longer talent or capital. It is power delivery, grid interconnection, permitting, and land. In mature markets, connecting a data centre to the grid now takes longer than building the data centre. [1][2]

This matters because AI is shifting from training to inference. Training happens once. Inference happens every time a model is used, every query, every prediction, every piece of enterprise automation, and it grows with adoption. [3] As inference becomes the dominant workload, AI value capture stops being a contest of algorithms and becomes a contest of infrastructure.

That contest suits India. The countries that deliver power fastest, not the best models, will capture the inference economy. This memorandum sets out India’s structural case, the trap to avoid, and a state-level pilot in Tamil Nadu. The capital has already voted. Microsoft has committed $17.5 billion to India. Google is building a gigawatt-scale AI hub at Visakhapatnam. OpenAI and Tata have announced 100 megawatts with a path to a gigawatt, and Meta has signed its first Indian AI data centre deal with Reliance. [11] The thesis is not a forecast. It is under construction.

At Three Altitudes

For allocators. The binding variable in data centre returns is shifting from land and headline power price to time-to-power: the elapsed time between approval and reliable grid electricity. Multi-year interconnection queues in the US and Europe can impair returns even when tenant demand is strong. [1][2][5] Capital concentrated on a small number of well-prepared, grid-coordinated corridors will outperform capital spread across dispersed projects. For governments, permitting speed and grid coordination are now industrial policy.

For operators. Inference is recurring and operational, not episodic. Once embedded in enterprise workflows it becomes predictable long-term demand. India’s services firms and Global Capability Centres can bind compute to talent, selling compute plus services rather than bare hosting, which lifts utilisation and keeps more of the value onshore. [9] For data centre operators, utilisation ramp and time-to-power decide the economics more than any other variable.

For observers. The AI story is becoming an infrastructure story. The numbers to watch are interconnection queues and megawatts delivered, not model benchmarks. And the strategic point is this: India does not need to win frontier training to matter. It needs to deliver power, reliably and quickly, at scale.

01 · The Shift

Three structural changes drive the argument.

Inference is becoming the dominant workload. A trained model is used millions or billions of times. The economic value migrates from training events to inference volume, especially in enterprise deployments: copilots, search, customer support, analytics, embedded automation. [3]

Infrastructure is becoming the binding constraint. Inference needs persistent power, cooling, networking, and physical capacity. Unlike training, it is continuous, and it lives or dies on uptime and the cost and reliability of electricity. Construction costs per megawatt have risen across most mature markets. [5][6]

And the grid is the bottleneck. In several advanced economies, interconnection delays now exceed construction timelines. European hubs report multi-year connection queues that are already shaping where capacity can go. [1] US interconnection backlogs have grown for a decade. [2] Power availability, not chip supply or capital, is becoming the decisive constraint on where AI runs.

02 · The Landscape

The United States dominates frontier training and capital, but hyperscale growth is constrained by grid bottlenecks, permitting resistance, and queue delays. [2] It will keep sensitive and ultra-low-latency workloads at home. Portions of non-sensitive, batchable inference will seek offshore capacity wherever time-to-power and total cost are superior.

Europe faces structural grid congestion, with legacy hubs carrying the longest connection lead times. [1] It will prioritise strategic autonomy and sensitive workloads onshore, while bulk inference growth meets friction.

The Gulf states deploy capital fast, helped by centralised decision-making and cheap energy. Their long-term position depends on trusted legal frameworks for cross-border data and durable alignment with the major platforms.

India’s profile is distinct: expanding renewable capacity with improving procurement pathways, a data centre market in rapid build-out, national coordination through the IndiaAI Mission, the world’s largest IT services and GCC base, and a federal structure that lets individual states execute and compete. [4][6][7][9]

Infrastructure investors have announced multi-gigawatt Indian pipelines through 2030 alongside the commitments above. [11] The question is no longer whether capacity comes to India. It is whether India captures durable value from it, which is a different problem.

03 · What India Already Holds

The IndiaAI Mission, approved in March 2024 with an outlay of ₹10,371.92 crore over five years, spans compute, innovation, datasets, applications, skilling, startup financing, and safe AI, with an ambition of a federated national compute network referencing around 38,000 GPUs at scale. [4][8]

Multiple states have issued data centre policies covering land, duties, renewable procurement, and single-window clearance, a ready platform for state-level execution. [6][7] India’s renewable build-out and procurement mechanisms, open access, group captive, long-term PPAs, give large power consumers a route to predictable cost and credible carbon positioning.

The gap is framing. These efforts are built around domestic capability and investment attraction. None of them yet positions inference as an export product. That reframing costs little and changes what the same infrastructure is worth.

04 · Why Inference Exports Fit India

Training dominance demands frontier research clusters, proprietary data, multi-billion-dollar risk appetite, and preferential access to cutting-edge accelerators. India trails on all four, and so does nearly everyone.

Inference exports demand something else entirely: reliable power at predictable prices, predictable policy, fast permitting and coordinated grid planning, enterprise-grade compliance, and operational excellence. These map directly onto India’s strengths, especially once the services ecosystem is attached. The workloads are recurring, tolerant of moderate latency, and once integrated into enterprise processes they are sticky for years.

India does not need to out-research San Francisco. It needs to out-deliver Virginia and Dublin on time-to-power. That is a winnable contest.

05 · The Commodity Trap

Hosting compute alone does not guarantee value capture. The failure mode is becoming a low-margin landlord: foreign-controlled infrastructure, value accruing to hyperscalers and energy suppliers, minimal spillover to the domestic economy.

The mitigation is deliberate pairing. Compliance layers that reduce friction for regulated but non-sovereign workloads. Domestic operator capability, Indian-owned or joint-venture data centre and managed-service providers. Long-term enterprise contracts that justify resilient investment. Integration with GCCs so the export is inference plus services, not bare hosting. And public-sector AI adoption to anchor base demand. [9] Done together, hosting becomes embedded economic infrastructure rather than commodity floor space.

06 · The Tamil Nadu Corridor

India’s federal structure means one state can prove the model. Tamil Nadu is the natural candidate: a leading industrial state with deep manufacturing and logistics density, established administrative capacity, and in Chennai a major data centre and subsea cable landing hub that improves international network economics for export workloads. [10] The state has a dedicated data centre policy and renewable scale to support green power contracting.

The pilot is an Inference Corridor: a small number of highly prepared sites rather than dispersed projects. Pre-zoned, pre-cleared land with defined standards. Coordinated grid and substation planning with published capacity roadmaps. Time-to-power service commitments for qualifying projects, with transparent milestones. A compliance and security stack for regulated but non-sovereign workloads. Integration with local talent and GCCs. And one or two anchor projects secured through bilateral channels to create the reference cluster.

None of this requires national restructuring. It requires one state to compress time-to-power and package infrastructure, compliance, and talent into a single export-ready offer. High-density inference will also force the cooling question early, and corridor design should assume liquid-assisted cooling and renewable-linked sourcing from day one.

07 · Risks

Four are live. Grid expansion lagging load growth, in India as much as anywhere. Overconcentration in a few hyperscalers, which weakens domestic bargaining power. Geopolitical sensitivity in cross-border AI hosting, export controls, localisation rules, security review. And technology shifts, model efficiency gains, on-device inference, new accelerator architectures, that change demand profiles and facility design. An inference strategy needs standing monitoring of platform dynamics, semiconductor supply, and regulation in the major demand centres, with the flexibility that state-level execution allows.

Conclusion

AI value capture is being decided by infrastructure, and the infrastructure contest is being decided by time-to-power. India holds the structural cards: renewable scale, a coordinated national compute mission, the world’s deepest services ecosystem, and states able to move at their own speed. The capital has already started arriving. What converts arrival into durable advantage is deliberate positioning: inference framed as an export, value capture engineered through services integration, and one state proving the corridor model first. Tamil Nadu can be that state. The window is open now, while the congested markets are still stuck in their queues.

References

All research →